That dashboard has a blind side. Four unconnected sources landed the same warning on it in the third week of August 2026—no shared data, no shared authors, no shared funding^1,2,3,4. Their warning: your people’s judgment is wearing down as a byproduct of AI use, and nothing on the dashboard would show it. That means your review this week measures the wrong layer.

That wrong layer is a question I left open two weeks ago, in Everybody Named the Problem. Nobody Teaches the Solution. How do you measure the judgment rebuild inside your own P&L? This Call walks straight at it. What converged, what it says about the metrics you report, and where measurement actually stops.

Four Unconnected Sources Converged in One Week: AI Use Wears Down the Judgment That Aims It

Every week brings executives a fresh AI warning. Most trace back to the same recycled report. These four are different, and you can check the difference yourself—four unconnected directions, four different methods:

1. The Harvard experiment (August 19). Harvard Business School researchers ran a randomized field experiment—228 evaluators making 3,002 screening decisions on real innovation submissions, with data gathered in 2024⁵. Harvard Business Review carried the finding to executives on August 19, 2026, under a blunt headline: AI is undermining leaders’ judgment¹. The next section explains the full result, including the half the headline omits.

2. The Stanford payroll panel (revised August 2026). Stanford economists Brynjolfsson, Chandar and Chen analyzed millions of records from Automatic Data Processing (ADP), which runs payroll for 26 million US workers. Employment of workers aged 22–25 in the most AI-exposed occupations sits 19% below where it would be if it had kept pace². A year earlier the gap was 15%—a 4-point widening. Underneath: entry-level work built on codified knowledge is slowing, while tacit-knowledge work grows for experienced workers—patterns the authors call descriptive, not causal, scoped to their ADP sample². The same panel anchored You Priced the License. You Never Priced the Trust. two weeks ago; the continuing slow down of junior hiring is this Call’s highlighted evidence from the same study.

3. The practitioner warning (August 19). Futurist Ross Dawson, an executive advisor last on this page in Your AI ROI Is Hiding in the Judgment Layer, said it plainly on his own Humans + AI podcast: “by default, the way we are building and using AI, and the way we’re bringing AI to organizations, is taking away from human judgment and capabilities.” The human learning loop, he argues, “will basically erode unless we augment it”³. Dawson advises on the remedy he names—nevertheless, his expertise lends his voice credibility.

4. The decision-habit warning (August 17). Philip Topham, who advises founder-led private companies with $20M to $250M in revenue, published the adjacent failure on August 17, 2026: operators running AI-era decisions with habits fitted to a company that no longer exists. “AI did not create this problem. It changed the track underneath it”⁴. His warning is about misfit, not erosion—and it counts toward one claim only: executive decision habits are the binding constraint.

Zero shared data, zero shared authors, zero shared funding: the false-convergence check passes. And none of the four measures whether any organization’s judgment is improving or eroding—not the experiment, not the payroll panel, not either practitioner. The four name the variable. No metric in the class your AI program ships with carries a judgment line that scores it.

So if you are extracting ROI from AI already deployed, the erosion mechanism runs inside tools you already bought—assistants that ship agreeable by default, attached to your decision loops—while the adoption dashboard reads green. Later sections price what judgment is worth when a firm designs for it.

And if you are implementing now, the sharpest lesson from these data is that outcome can be determined by a design decision made at build time. A design review today costs a meeting. Discovering judgment erosion later costs you the rework of habits your teams have already formed around AI use.

The Same AI Made Decisions Better as a Black Box—and Worse When It Explained Itself

That design decision is the experiment’s own finding, and the half the headline omits comes first: plain black-box recommendations—a score with no story attached—improved decision quality in the 2024 Harvard experiment⁵. AI in the loop was not the problem.

The problem arrived with the explanation. With a persuasively written narrative attached, the same model pushed false negatives—good submissions the evaluators incorrectly rejected—up 14.9 points, against 8.4 for the black box⁵. That 6.5-point difference means the narrative nearly doubled the black box’s error—which is why the override step your people run is the asset in play. The 72 experienced evaluators performed a little better. Their false negatives rose 12.9–14.5 points⁵.

That override suppression is the mechanism, and the authors name it: narratives “suppress productive overrides by substituting persuasive text for independent verification”⁵. The evaluators stopped checking because the narrative sounded checked. Two boundaries keep that verification claim honest: this is a working paper—not yet through journal review—and it tested one innovation-screening task⁵. And the two eroding capacities you may have seen quoted are the HBR authors’ framing, not the study’s¹.

Most importantly, the persuasion is not exotic—it ships in the box. Anthropic researchers, in work published in 2024 that examined their own models alongside competitors’, found that sycophancy—agreeing convincingly instead of answering correctly—is a general behavior of flagship AI assistants, a documented by-product of training on human approval⁶. The same by-product bit OpenAI in April 2025: the company shipped an update to 500 million weekly users and rolled it back within days. Its post-mortem opens plainly—”It aimed to please the user”—and concedes the prior reward signals had been “holding sycophancy in check”^7,8. I unpacked how this happens in It’s Not Thinking. It’s Predicting.

That baseline is why the Harvard team’s design clause matters: effective collaboration “requires designs that preserve rather than supplant independent human judgment”⁵. So for your deployments the test is simple. Wherever your AI explains itself persuasively inside a decision loop, your people need a verification step where the narrative cannot talk humans out of their own review justifications.

Developers Said AI Made Them Faster—the Stopwatch Showed 19% Slower

You might think that asking your people is the cheapest measurement plan; but one lab checked it against a clock. The nonprofit AI-evaluation lab METR (Model Evaluation and Threat Research) randomized real tasks—AI allowed or not—across 16 veteran open-source developers working their own mature codebases⁹. The 2025 run covered 246 real tasks.

With early-2025 AI tools the developers took 19% longer—while believing AI had sped them up by 20%. They had forecast 24% efficiency before starting⁹. Their own read was wrong in sign, not just size—and that kills the cheapest measurement plan. Treat the finding as directional—one study, built on self-reported time inside a randomized design. Even so, it is the best evidence on record that perception fails exactly where you need it to hold.

That perception gap outlived the study. METR has since labeled the 19% slowdown historical; its February 2026 update believes developers are now likely faster with AI. That means the tools improved while people’s own estimates stayed wrong—the gap between felt speed and measured speed is the finding that survived every revision¹⁰. What the lab could not survive was adoption itself, and by early 2026 the drift was structural. Then between 30% and 50% of its developers—up to half the panel—were declining tasks they would have to complete without AI, and METR wrote that such self-reported speedups “can be quite unreliable”¹⁰. METR lost its control group.

That loss is coming for your organization too. Once your people will not work without AI, participation-based measurement stops working—and perception-based measurement goes with it. So repeatable measures will have to rest on external signals rather than self-report. Naming those signals for each decision loop is the real work. It is exactly the work your adoption metrics most likely skip.

Make Adoption the Target and Your People Will Give You Adoption—Not Better Decisions

That skipped work runs into the oldest law of measurement. In 1975 the economist Charles Goodhart wrote it down: “any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes”¹¹. For an operator the law reads simply: manage to a measure, and the measure stops telling you the truth. Your people will deliver the measure. Make adoption the goal and you will get adoption: logins, prompts, token counts, whether or not a single decision improves. So a green adoption dashboard proves pressure, not return—and that leaves your decision quality unmeasured.

This is the line I drew in June 2026, in The Sophistication Gap: adoption is a headcount; sophistication is a capability. A headcount can tell you who has access; it cannot tell you whether outcomes improved. Behind that sophistication line sits an analysis by the University of Texas at Austin with KPMG, the global professional-services firm¹². They examined 1.4 million real prompts—and found roughly 5% of users working at the level that changes how work gets done. That means your usage report and your outcome report are different documents.

So most teams fall back on a survey, and organizational-learning research has already run that experiment for three decades. In 2004, researchers validated the field’s flagship instrument—the Dimensions of the Learning Organization Questionnaire, by Yang, Watkins and Marsick, N = 836—with half the sample held out to check the result: a seven-dimension questionnaire in which even the performance outcomes are respondents’ perceptions, a limit its own authors state. It’s also worth noting that the validators were also the framework’s creators¹³. The important take away is that surveys anchored on self-perception–exactly the mechanism the METR results employed–are structurally deficient. You need metrics that score decision quality, which is the true indicator of increased or decreased quality in your AI implementation’s outcome layer.

Scoring Judgment Itself Is a Limit Research Hasn’t Fully Explored—Certainly Not With AI as a Factor

So what is judgment worth in money? Researchers have already priced its presence—twice, in opposite directions, at the same kind of firm. In November 2020, Management Science published a field experiment at an automobile spare-parts retailer: merchants overrode the finished decisions of a working forecasting tool, and the overrides significantly reduced profitability¹⁴. In January 2026 the same lead researcher published the redesigned experiment. It invited merchants’ judgment into the inputs the forecasting tool consumed—and profit rose 4.92%, a result the randomized design lets the authors call causal¹⁵.

Those two results never contradict each other; the design decides the sign. Misplaced judgment destroys value; judgment as part of the design creates it. And neither experiment—in either direction—scores anyone’s judgment as a capability. Both price its presence. Nothing in either design measures its quality. For you, that means the profit lever is real and the gauge for it still does not exist.

That missing gauge is the field’s honest edge, and I will state it carefully. In the research we reviewed and the searches we ran—re-verified at publication—no firm-level study yet measures the improvement of human judgment as a scored capability, a judgment-quality score that moves, and connects that movement to P&L. I am not claiming none exists. This is a limit that thus far has not been fully researched—certainly not with AI as a factor. The finding still points to judgment as a lever in profitability; how and where those levers can be designed to move P&L is the salient question.

So look at the instruments you already own. Adoption dashboards, maturity indexes, ROI reframes—every one measures the rollout, not the judgment aiming it. On the dashboard you will carry into this quarter’s review, the judgment line is still blank.

Ng, Hill, and Ibarra Say the Constraint Is the Organization, Not the Technology

That blank line raises the pace question, and the field’s teachers already answer it. Andrew Ng—the most-followed instructor in AI, and a man who sells AI training—told builders in the August 14, 2026, issue of The Batch that the operator skill is knowing “when to slow down and take longer in order to build more carefully”¹⁶. He aims that advice at engineering teams—so for you it transfers as a posture, not a policy.

That posture has a leadership component. Harvard’s Linda Hill, featured on NPR’s Marketplace in July 2026, named it: “many leaders first thought of this AI as a technology,” and “[t]hey now understand it’s really about people and cultural transformation.” Her readiness test is three capabilities: “Can your organization collaborate, experiment, and learn?” She advises companies on exactly this transition¹⁷. At London Business School, Herminia Ibarra—writing on her school’s own platform—put the verdict in her April 2026 headline: “The issue isn’t the technology itself – it’s humans’ ability to use it”¹⁸.

Those three voices agree, and the prompt-level evidence points the same way. The 1.4-million-prompt analysis behind The Sophistication Gap, from March 2026, located the value in the roughly 5% who use AI to change how the work gets done—not in usage volume¹². That study never connects sophistication to the P&L, and I will not connect it here. What it supports is a starting point: improved decisions, not usage volume, are the better predictor of improved outcomes. So measure how, not whether—as a starting point you can iterate on. Evaluate outcomes, where and how human judgment is designed into the workflows.

Sit down with your dashboard this quarter. Every line on it answers some version of one question: whether your people used the tools. Look for the line that answers the other one: whether their decisions got better. The design levers exist—the Harvard experiment names them. So does the external-signal principle—METR paid for it with its own control group. And the price is on record—Kesavan’s autobody retail chain experiment where the design itself changed the outcome. What does not exist, as of this publication, is the score. Which external signals belong on that line—and survive the gaming pressure Goodhart named—is fertile ground for exploration.

That line is yours to draw—and you do not need anyone’s permission to start. Today.

The AI Leadership Playbook

Strategic Questions — copy-paste ready for an email to your CFO and CHRO:

1. Which of our AI metrics would still move if not a single decision improved—and which one would catch it if judgment eroded? What is the first metric we retire?

2. Where does our AI explain itself persuasively inside a live decision loop—and whom did we assign, in writing, to dissent from it? What is the first loop we assign?

3. If the board asked for a judgment line next to the adoption line, what signal would we put on it this quarter—and who owns it?

Your Next Plays — copy-paste ready for an email to your leadership team:

1. Tag the metric class. The Sophistication Gap replaced the adoption dashboard with a sophistication scorecard; extend it one rung. List every AI metric you report and tag each one rollout-measure or outcome-measure. Mark which would catch a judgment change. One page.

2. Run the design review. Pick one decision loop where AI explains its recommendations persuasively. Add the independent-verification step the Harvard experiment showed the narrative suppresses—a named human who checks the recommendation against the evidence.

3. Define one external signal. Pick one decision loop and define its outcome-layer signal—external to the people in the loop, not self-reported—and name who scores it. Design it expecting gaming pressure; Goodhart’s law applies to the fix too.

📅 Book a complimentary 1:1 Strategy Session—45 minutes to start the conversation about mapping the revenue hidden in your AI spend.

Sources

1. Leonid Sudakov & Nathan Furr, “AI Is Undermining Leaders’ Judgment. Here’s What to Do About It.” Harvard Business Review, August 19, 2026. https://hbr.org/2026/08/ai-is-undermining-leaders-judgment-heres-what-to-do-about-it

2. Erik Brynjolfsson, Bharat Chandar & Ruyu Chen, “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence,” Stanford Digital Economy Lab working paper, revised August 2026. https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf

3. Ross Dawson, “Recursive Self-Improvement in Humans + AI Systems,” Humans + AI podcast, Episode 54, August 19, 2026. https://humansplus.ai/podcast/ross-dawson-recursive-self-improvement-humans-ai-systems-hai-ep54/

4. Philip Topham, “Does Your Company Have a Comfortable Shoes Problem?” SavionAI: The Shift. Lift., August 17, 2026.

5. Lane, Boussioux, Ayoubi, Chen, Lin, Spens, Wagh & Wang, “The Narrative AI Advantage? A Field Experiment on AI-Augmented Evaluations of Early-Stage Innovations,” Harvard Business School Working Paper 25-001, current revision fetched August 2026. https://www.hbs.edu/ris/Publication%20Files/25-001_75068eb2-d475-4889-b341-aeabab6ef6d1.pdf

6. Mrinank Sharma, Meg Tong et al., “Towards Understanding Sycophancy in Language Models,” ICLR 2024. https://arxiv.org/abs/2310.13548

7. OpenAI, “Sycophancy in GPT-4o: what happened and what we’re doing about it,” April 29, 2025. https://openai.com/index/sycophancy-in-gpt-4o/

8. OpenAI, “Expanding on what we missed with sycophancy,” May 2, 2025. https://openai.com/index/expanding-on-sycophancy/

9. Joel Becker, Nate Rush, Elizabeth Barnes & David Rein, “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity,” METR, July 10, 2025. https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/

10. Joel Becker, Nate Rush, Tom Cunningham, David Rein & Khalid Mahamud, “We are Changing our Developer Productivity Experiment Design,” METR, February 24, 2026. https://metr.org/blog/2026-02-24-uplift-update/

11. K. Alec Chrystal & Paul D. Mizen, “Goodhart’s Law: Its Origins, Meaning and Implications for Monetary Policy,” Bank of England festschrift paper, 2001, reproducing Charles Goodhart (1975), “Problems of Monetary Management: The U.K. Experience,” Reserve Bank of Australia. The popular one-line version—”When a measure becomes a target, it ceases to be a good measure”—is Marilyn Strathern’s 1997 paraphrase; the 1975 original wording is Goodhart’s. http://web.archive.org/web/20250910114222/https://cyberlibris.typepad.com/blog/files/Goodharts_Law.pdf

12. KPMG + University of Texas at Austin research, “What the Best AI Users Do Differently,” Harvard Business Review, March 2026—the analysis behind The Sophistication Gap. https://hbr.org/2026/03/what-the-best-ai-users-do-differently

13. Baiyin Yang, Karen E. Watkins & Victoria J. Marsick, “The construct of the learning organization: Dimensions, measurement, and validation,” Human Resource Development Quarterly 15(1), 2004. https://assets.csom.umn.edu/assets/21929.pdf

14. Saravanan Kesavan & Tarun Kushwaha, “Field Experiment on the Profit Implications of Merchants’ Discretionary Power to Override Data-Driven Decision-Making Tools,” Management Science 66(11), November 2020. https://pubsonline.informs.org/doi/10.1287/mnsc.2020.3743

15. Saravanan Kesavan, Tarun Kushwaha & D. Steele, “Profit Implications of Judgmental Adjustments to Forecast Inputs: Evidence from a Large-Scale Field Experiment,” Management Science 72(1), January 2026. https://pubsonline.informs.org/doi/10.1287/mnsc.2024.06321

16. Andrew Ng, The Batch, issue #366, DeepLearning.AI, August 14, 2026. https://www.deeplearning.ai/the-batch/issue-366

17. Marketplace Morning Report, “How the AI revolution impacts corporate leaders” (edited transcript of Linda Hill, Harvard Business School), July 20, 2026. https://www.marketplace.org/story/2026/07/20/how-the-ai-revolution-impacts-corporate-leaders

18. Herminia Ibarra & Florence Wilkinson, “Why AI is a leadership challenge – not a technology one,” London Business School Think, April 23, 2026. https://www.london.edu/think/ai-leadership-challenge

19. CognivaLab, “Everybody Named the Problem. Nobody Teaches the Solution.” The AI Playbook, August 9, 2026. https://www.cognivalab.blog/p/everybody-named-the-problem-nobody

20. CognivaLab, “You Priced the License. You Never Priced the Trust.” The AI Playbook, August 16, 2026. https://www.cognivalab.blog/p/you-priced-the-license-you-never

21. CognivaLab, “Your AI ROI Is Hiding in the Judgment Layer,” The AI Playbook, August 2, 2026. https://www.cognivalab.blog/p/your-ai-roi-is-hiding-in-the-judgment

22. CognivaLab, “It’s Not Thinking. It’s Predicting.” The AI Playbook, June 23, 2026. https://www.cognivalab.blog/p/its-not-thinking-its-predicting

23. CognivaLab, “The Sophistication Gap,” The AI Playbook, June 9, 2026. https://www.cognivalab.blog/p/the-sophistication-gap

But that doubling has not shown up in measured output. Economists tracked about 25,000 workers in a Danish national study through 2024 and found a precise null. They rule out any effect on earnings and hours larger than 2%.²

And this morning two of your best operators gave notice. You priced every seat of the license to the dollar, so you can say exactly what your AI costs. The thing that failed first never made the P&L.

The Fear Is a Forecast

That missing line item has a name your people already use. Only 22% of employees say their organization has communicated a clear AI plan, in that same 2025 Gallup survey.¹ Gartner surveyed 110 CHROs in late 2025. Among them, 95% run organizations that have implemented AI, and only 20% report significant value. So 80% of your peers are paying for a capability their own leadership cannot yet see in the numbers.³

That value gap has a practitioner’s number attached. Melissa Reeve, the practitioner behind the Hyperadaptive framework, puts a scale on it in her January 2026 article: “Nearly 80%” of AI initiatives fail — her framework’s number, an estimate by her own sourcing. That leaves you paying for a stack most of your people cannot explain the purpose of.⁴

Your people distrusted leadership before AI arrived. In Gallup’s 2022 panel, 21% of U.S. employees strongly agreed they trust their organization’s leadership, down from 24% in 2019. Read it the other way: 79% do not.⁵

But reassurance will not fix this. Conor Grennan tells leaders to validate the fear as rational; I go further — your people are forecasting. Economists Frey and Jegen showed in 2001 that workers withdraw willing effort when a policy feels controlling.⁶ That withdrawal shows up across 128 experiments analyzed in 1999: compliance rewards pushed motivation down and honest feedback pulled it up.⁷ Your mandate and your praise cancel each other out.

Your Hiring Plan Told Your Seniors They’re Next

Reassurance is not the plan your people are reading. They read it in what you do. Shopify’s chief executive Tobi Lütke told teams in an April 2025 self-published memo to justify why AI cannot do a job before asking for headcount and resources.⁸ And 22% of those 110 CHROs report at least one business leader who stopped entry-level hiring because of AI. The telling unit there is one leader each, not 20% of companies. Even so, that is a policy your junior candidates can see from outside the building, so your senior operators are certainly reading it from inside.³

That reading gets easier when a CEO says it out loud. Salesforce’s Marc Benioff told the same outlet twice in 2025. On July 30 he told Fortune: “I think AI augments people, but I don’t know if it necessarily replaces them.” Five weeks later, on September 2, he described cutting the support organization from 9,000 people to about 5,000. That is a 44% decrease, “because I need less heads.” The cut also drove support costs down 17%. Your own people read sequences like that one, and what they take from it is the plan, not the reassurance.⁹

Those two dates sit five weeks apart in the public record, and your senior operators can read a calendar. The junior evidence sits in three numbers with three distinct and separate scopes:

1. Occupations (ages 22–25). Stanford’s “Canaries” series — covered in July’s The Apprenticeship the Machine Ate — shows workers in the most AI-exposed occupations about 19% below less-exposed peers, in payroll data through June 2026. Descriptive, the authors say, not causal.¹⁰

2. Industries (ages 22–24). A U.S. Census Bureau working paper finds early-career employment in the most AI-exposed industries fell 12% in the ten quarters after ChatGPT launched, through mid-2025; up to 25% of that may be monetary policy.¹¹

3. Major tech companies only. SignalFire — published in July 2026 — puts new-grad hiring at the major tech companies roughly 65% below 2019, and specifically isolates that drop from the wider economy.¹²

Those three scopes describe different populations, and none of them needs a press release to reach your staff. Reuters’ factbox shows companies attributing layoffs to AI on the record; but even when the attribution isn’t overt — especially as entry-level jobs start disappearing — it doesn’t take a press release for your employees to read the tea leaves.¹³ Your employees price your AI promise by what you do, not by what you say.

The Bill Prices Who Leaves — and Who Stays

Behavior that reads as a broken promise has a price, and it arrives in two currencies. Gallup priced the first in 2019, by their own and undisclosed formula. Replacing an employee costs 50% to 200% of their annual salary: half a year’s pay at the floor, twice a full salary at the ceiling. In Gallup’s own arithmetic, that costs a 100-person organization up to $2.6 million a year.¹⁴

Salary multiples understate it, because layoffs also multiply the outflow. Downsizing just 0.5% of the workforce raised voluntary turnover 25% the next year. The 2008 study’s baseline quit rate ran 10.4%, so a very small cut produced a very large extra exit.¹⁵ Those extra exits compound at the top. After one high performer quits, other high performers’ quit rate rises 6% per month for the next three months. The 2024 study covered 1,620 retail stores, so treat it as a mechanism you can watch for in your own exit data rather than a rate you can apply to your business directly.¹⁶

Now the cost exit metrics never show. In 2007, researchers analyzed 51 studies of psychological-contract breach: the implicit deal employees build from perceived promises. Breach drives mistrust of management and drags satisfaction, commitment and daily performance — it does not predict actual turnover. That null is the point: the damage hides in the people who stay, so your retention dashboard stays green while the work quietly gets worse. Relational promises — job security, support — hurt more when broken than pay promises, and the AI job-safety promise is relational.¹⁷

Your People Beat AI Performance Only When You Let Them Learn

Here is where the hiring plan and the performance data collide. You are cutting the least experienced people, and the least experienced people are exactly who AI helps most.

That collision is visible in the field evidence. Human support agents working with an AI assistant resolved 15% more issues per hour. Among the least-skilled of them, the gain reached 30%. The 2025 study followed 5,172 human representatives.¹⁸ The dividend concentrates on novices, which is to say: in the roles your hiring plan just stopped filling. That means the cheapest productivity you can buy is sitting in the job you are no longer hiring for.

Then in May 2026, a Max Planck Institute-led team re-analyzed the same dataset that had powered a 2024 Nature Human Behaviour meta-analysis of human-AI experiments. The re-analysis covered 370 effect sizes from 106 experiments.¹⁹ Nearly all of those experiments ran on laboratory tasks with pre-GPT-4 systems between January 2020 and June 2023. In 86% of those studies, people never received outcome feedback — nobody ever told them whether the machine had been right.²⁰

Feedback, where it existed, tilted results toward synergy; feedback plus AI explanations turned positive, while explanations without feedback stayed clearly negative. Their conclusion, from a preprint the authors call tentative: “explanations increase synergy only when humans can learn to verify the AI’s reliability through feedback”.²⁰ That means you may be hitting a condition you set, not a limit of the technology you bought.

Feedback and the chance to learn are conditions, and conditions are what an employer supplies. Two organizations bought identical licenses; the one whose people could learn to check the machine got the return, and the one whose people could not, got confident errors at speed. The demonstration came from 758 BCG consultants in a 2023 study BCG co-authored. They finished 12.2% more tasks inside the AI’s competence, with more than 40% higher quality. Beyond the AI’s competence, consultants were 19 points less likely to be correct. The penalty travels with the gains, so the same tool that lifts your team also sets a trap for it.²¹

That same split ran on live customers at two companies. Klarna automated customer service to lower cost and got worse service. Its chief executive told Bloomberg in May 2025: “Really investing in the quality of the human support is the way of the future for us.” Klarna began recruiting human agents again, while saying it is “very much still AI-first” — a partial rollback, not a reversal. That correction still cost it twice: the quality it lost, and the credibility it spent.²² Commonwealth Bank of Australia went further. In August 2025 it rescinded 45 voicebot-attributed redundancies, apologized to the staff it had let go, and conceded through its own statement that the cuts were an error.²³

Both cases read differently. Both companies removed the humans, watched quality fall, and hired back; that means the people they cut were producing value their dashboards had priced at zero. Your workforce is watching for that same sequence. It could cost you twice: once when the service degrades, and again when the promise you broke wasn’t predicated on verified AI results.

The Breach Is a One-Way Door

That broken promise stays broken. Hiring back, as both companies did, does not buy the trust back. Researchers settled what trust is worth two decades ago: a 2002 meta-analysis covered 27,103 employees.²⁴ Across 106 samples, employees who trust their leaders report higher job satisfaction and stay more committed. Trust ranks among the strongest predictors researchers have measured. Those same employees plan to leave far less often. That means you already pay for satisfaction, commitment, and retention through a variable your budget never priced.

These two findings agree. Gallup counted how many employees trust leadership in 2022, and found few. Dirks and Ferrin measured what changes where that trust exists, across studies published through 2001. Trust is both rare and load-bearing.^24,5

That trust has a sharper finding underneath it: it matters who the employee is trusting. Employees work measurably better for a manager they personally trust. Those same employees gain nothing measurable from trusting ‘the leadership’ in the abstract. So you intervene with the manager, not with the message. Give every manager in your organization the training and the authority to build trust with their own reports, and you convert that research finding into output — never by proxy through you, and never through a policy.²⁴

That manager can also recover from a mistake in a way the company ‘we’ cannot recover from a broken promise. Trust-repair studies run in the laboratory explain why that door swings one way. In 2004, researchers showed an apology helps rebuild trust after a competence failure. After an integrity violation it barely helps. A missed forecast reads as a mistake. A broken promise reads as a failure of character.²⁵

That character judgment changes behavior. Employees who distrust leadership hide how they use AI. In 2025, KPMG, the global professional-services firm — AI governance among its services — funded a study, and University of Melbourne researchers ran it. They surveyed over 48,000 people across 47 countries. In that survey 61% of employees avoid revealing when they use AI, and 44% use it against policy. The AI activity inside your company is already larger than the activity you can see. This KPMG survey also anchored August’s Your AI ROI Is Hiding in the Judgment Layer.²⁶ Hidden use is unmanageable use, and it is sitting in your organization right now.

Now the other side of the door: in 2017, researchers combined 136 samples on psychological safety.²⁷ Teams that feel safe report their own errors, and reporting is how they learn — the link to learning is strong, the link to output more moderate. These are patterns rather than proven causation.

That link has a mechanism. Amy Edmondson’s original 51-team study at one U.S. manufacturer traced it in 1999: reporting errors is how teams learn, and learning is what lifts performance. So your people surface AI failures when they trust you, and those reports buy you an early warning no dashboard sells.²⁸

Your people already show you what they think of the promise — in what they hide, and in what they report.

State the Plan — Then Hand Over the How

That willingness is buildable, and the first condition costs one page. State the plan, plainly:

1. Human + AI: where, and how.

2. AI-only: where, how to verify accuracy, and who owns verification.

3. Human-only: where, and why.

Then never promise what your hiring plan will contradict. Ambiguity is where the tacit breach lives — only 22% of employees say they have heard a clear AI plan at all, in that same 2025 survey.¹

A clear plan is also what makes a repair possible, and Duolingo shows one that worked. Its April 2025 memo said the company would “gradually stop using contractors to do work that AI can handle”. After weeks of backlash, chief executive Luis von Ahn walked it back — he reverted; he never apologized: “To be clear: I do not see AI as replacing what our employees do (we are, in fact, continuing to hire at the same speed as before)”.²⁹

His behavior made that clarification stick: no full-time layoffs in 17 years, and hiring continued. The cuts did fall on contractors, so hold that distinction when you tell this story internally. By May 2026 he dropped his AI-usage performance rule because workers used AI “just for AI’s sake,” so the mandate half died of its own theater. What held was the half backed by behavior, which is the half your people will judge you on too.²⁹

The second condition is autonomy over the how, and it is trainable. Hardré and Johnmarshall Reeve ran a randomized experiment, published in 2009: five weeks of coaching moved managers’ behavior by more than a full standard deviation — a large shift, the kind that changes what a team notices day to day — and employee engagement moved with it, by a smaller but real margin. The researchers never tested durability beyond those five weeks.³⁰ That training is something you can commission this quarter, and you can measure what it moved.

That training changes how managers behave. The third condition: psychological safety, so your teams surface where AI fails instead of hiding it.

Those three conditions — a stated plan, autonomy over the how, and the safety to surface failures — cost less than the license. When you sit down to price next quarter’s AI budget, the cheapest line item on it will be the one that was never there. Trust buys participation; judgment aims it. That aim is what the consultants’ 19-point penalty measures: willing misuse without judgment, which is how your best people come to ship confident errors at speed. Judgment stays the lead lever — trust is what gets your people to pull it.^21,19

This is the work I do with executives: I map the revenue hiding in your AI spend, sharpen the judgment that’s forfeiting it, and instrument the capture. The AI ROI Map tells you where your AI spend is paying, where it’s forfeiting return, and the moves that capture it — the trust conditions above decide how much there is to capture.

Trust is also the dividend your people are already pricing. They know which organization you are becoming. This week, tell them — in writing. Measuring the willing use your dashboard cannot see gets its own forthcoming Call.

The AI Leadership Playbook

These questions put that plan on someone’s desk this week.

Strategic Questions (copy-paste ready for an email to your CFO and CHRO):

1. Which of our AI communications promised our people safety — and does our hiring plan contradict it? What changes first: the promise or the plan?

2. How much of our adoption number is willing use versus mandated performance — and which behavior-revealed measure do we add first?

3. If our two best operators left for the organization that kept its promise, what would replacing their judgment cost us?

Your Next Plays (copy-paste ready for an email to a direct report):

1. Draft the stated plan. One page: human+AI, AI-only, human-only — where and how, each. Check it against the hiring plan before anything ships.

2. Pressure-test the adoption dashboard. Separate autonomous AI-agent invocations from human use; add one measure no one can game by running up usage — voluntary error-surfacing is the candidate.

3. Put the trust lever where the research says it lives: the direct manager. Commission autonomy-supportive manager training — five weeks moved manager behavior by more than a standard deviation — and measure engagement before and after.³⁰

📅 Book a complimentary 1:1 Strategy Session—45 minutes to start the conversation about mapping the revenue hidden in your AI spend.

And one disclosure before the Sources: I worked at Salesforce during the period described in this Call, with no involvement in the decisions reported here.

Sources

1. Gallup, AI Use at Work Has Nearly Doubled in Two Years. https://www.gallup.com/workplace/691643/work-nearly-doubled-two-years.aspx

2. Humlum & Vestergaard, Still Waters, Rapid Currents (NBER working paper w33777). https://www.nber.org/papers/w33777

3. Gartner, CHRO survey press release, July 2026. https://www.gartner.com/en/newsroom/press-releases/2026-7-27-gartner-survey-finds-ai-automation-is-reducing-some-entry-level-hiring-at-nearly-one-quarter-of-organizations

4. Melissa Reeve, The 5 Stages of Becoming AI-Native, IT Revolution. https://itrevolution.com/articles/the-5-stages-of-becoming-ai-native-the-hyperadaptive-model/

5. Gallup, Why Trust in Leaders Is Faltering and How to Gain It Back. https://www.gallup.com/workplace/473738/why-trust-leaders-faltering-gain-back.aspx

6. Frey & Jegen, Motivation Crowding Theory, Journal of Economic Surveys. https://onlinelibrary.wiley.com/doi/10.1111/1467-6419.00150

7. Deci, Koestner & Ryan, meta-analytic review of extrinsic rewards and intrinsic motivation, Psychological Bulletin. https://home.ubalt.edu/tmitch/642/articles%20syllabus/Deci%20Koestner%20Ryan%20meta%20IM%20psy%20bull%2099.pdf

8. Shopify memo coverage, Digital Commerce 360, April 2025. https://www.digitalcommerce360.com/2025/04/08/internal-memo-shopify-ceo-declares-ai-non-optional/

9. Fortune, Benioff on support-role cuts, September 2, 2025 (and prior interview, July 30, 2025). https://fortune.com/2025/09/02/salesforce-ceo-billionaire-marc-benioff-ai-agents-jobs-layoffs-customer-service-sales/

10. Stanford Digital Economy Lab, Canaries in the Coal Mine? (August 2026 revision). https://digitaleconomy.stanford.edu/news/canariesaug26/

11. U.S. Census Bureau (Tucker), You’re (not) Hired, CES-WP-26-27. https://www.census.gov/library/working-papers/2026/adrm/CES-WP-26-27.html

12. SignalFire, State of Talent Report 2026. https://www.signalfire.com/blog/signalfire-state-of-talent-report-2026

13. Reuters factbox on AI-attributed job cuts, July 6, 2026 update. https://www.investing.com/news/stock-market-news/factboxcompanies-cutting-jobs-as-investments-shift-toward-ai-4776976

14. Gallup, This Fixable Problem Costs U.S. Businesses $1 Trillion. https://www.gallup.com/workplace/247391/fixable-problem-costs-businesses-trillion.aspx

15. Trevor & Nyberg, downsizing and voluntary turnover, Academy of Management Journal. https://journals.aom.org/doi/10.5465/amj.2008.31767250

16. Sajjadiani et al., high-performer turnover dynamics, via LSE Business Review. https://blogs.lse.ac.uk/businessreview/2024/03/13/when-a-high-performer-leaves-the-firm-loses-more-than-a-good-worker/

17. Zhao, Wayne, Glibkowski & Bravo, psychological contract breach meta-analysis, Personnel Psychology. https://onlinelibrary.wiley.com/doi/abs/10.1111/j.1744-6570.2007.00087.x

18. Brynjolfsson, Li & Raymond, Generative AI at Work, Quarterly Journal of Economics. https://academic.oup.com/qje/article/140/2/889/7990658

19. Vaccaro, Almaatouq & Malone, When combinations of humans and AI are useful, Nature Human Behaviour. https://www.nature.com/articles/s41562-024-02024-1

20. Berger et al. (Max Planck Institute for Human Development), Fostering human learning is crucial for boosting human-AI synergy, arXiv preprint. https://arxiv.org/abs/2512.13253

21. Dell’Acqua et al., Navigating the Jagged Technological Frontier. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4573321

22. Fortune, Klarna coverage, May 9, 2025. https://fortune.com/2025/05/09/klarna-ai-humans-return-on-investment/

23. The Register, Commonwealth Bank coverage, August 22, 2025. https://www.theregister.com/2025/08/22/commonwealth_ban_chatbot_fail_rehiring/

24. Dirks & Ferrin, Trust in Leadership: Meta-Analytic Findings, Journal of Applied Psychology 87(4), 611–628. https://ink.library.smu.edu.sg/lkcsb_research/675/

25. Kim, Ferrin, Cooper & Dirks, Removing the Shadow of Suspicion, Journal of Applied Psychology. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=398221

26. KPMG & University of Melbourne, Trust, Attitudes and Use of Artificial Intelligence: A Global Study 2025. https://kpmg.com/kpmg-us/content/dam/kpmg/pdf/2025/trust-attitudes-artificial-intelligence-executive-summary.pdf

27. Frazier et al., Psychological Safety: A Meta-Analytic Review and Extension, Personnel Psychology. https://onlinelibrary.wiley.com/doi/abs/10.1111/peps.12183

28. Edmondson, Psychological Safety and Learning Behavior in Work Teams, Administrative Science Quarterly. https://journals.sagepub.com/doi/10.2307/2666999

29. Duolingo memo and clarification, via Entrepreneur, May 2025. https://www.entrepreneur.com/business-news/duolingo-ceo-clarifies-ai-stance-after-backlash-read-memo/492141

30. Hardré & Reeve, Training corporate managers to adopt a more autonomy-supportive motivating style. https://selfdeterminationtheory.org/wp-content/uploads/2023/10/2009_HardreReeve_TrainingCorporateMgrs.pdf

Solis wrote the sharpest version — the sixth of seven questions he says every CEO now faces — judgement — recapping a Drucker Forum session from last November [1]: “How will we protect and expand human judgment?” [1]. His answer only addresses “judgment” — he defined what AI can recommend, what it can decide, what it can execute [1]. For expand, he offered nothing. He answered half the question. He left the other half blank.

Four of the five ship frameworks. Lancefield maps four modes of AI collaboration [2]. Chamorro-Premuzic gives a selection model — hire for judgment [3]. Ferrazzi and Smith give five mental models for working with agents [4]. The Forum hands boards five priorities [5]. Every one tells you what to do with the judgment you already have: guard it, hire for it, calibrate trust around it. None explained how to build it. The research below prices that missing half.

The Erosion Has Numbers Now

Two research teams measured judgment loss under AI directly — and turned the anecdote into numbers:

1. The Wharton experiments (2026). Shaw & Nave ran 1,372 participants through 9,593 preregistered reasoning trials — a preprint, meaning other scientists have not yet formally reviewed it [6]. When the AI answered correctly, participants’ accuracy rose 25 points. When the AI answered incorrectly, their accuracy fell 15 points [6]. Access to the AI also lifted their confidence from 65.3% to 77.0% — an 11.7-point jump — and their confidence held even as the AI’s errors piled up [6].

2. The Radiology reader study (2023). Dratsch and colleagues, in the journal Radiology, asked 27 radiologists to evaluate 50 mammograms each with a purpose-built, simulated AI [8]. On the 12 sabotaged cases, inexperienced readers fell from 79.7% correct to 19.8% — a 59.9-point drop [8]. The very experienced fell too — 82.3% to 45.5%, a 36.8-point drop [8]. Experience softened the collapse. It did not stop it.

Human-factors researchers documented this long before today’s tools. Parasuraman & Manzey’s canonical 2010 review concluded complacency is “found in both naive and expert participants and cannot be overcome with simple practice” [10].

Hospital clinicians now report the same experience on the job. A CHI 2026 study — CHI, the leading conference on how humans interact with computers — followed a real AI deployment for 12 months across a multi-site hospital system in five North American locations: 42 participants, 15 of them radiation oncologists [9]. The authors watched early efficiency gains hide the erosion they came to call intuition rust — the gradual dulling of expert judgment [9]. A senior radiation oncologist supplied the first of two remarks the paper presents side by side, without attributing both to one speaker: “My intuition is rusting,” and “I’ve no idea why the AI suggested it, but I accept it anyway.” [9].

Executives report seeing it too. BCG surveyed 70 C-suite leaders and senior executives worldwide, and the leaders ranked judgment and decision making as carrying the highest de-skilling risk score of all skills [7]. Half report already observing de-skilling. Almost 90% cite overreliance on AI outputs without stress testing or challenge, and 53% cite slower development of junior talent [7]. Only one in ten companies has an organisation-wide strategy to address any of it [7]. Every figure is what leaders report seeing — perception, not measured incidence [7].

Twenty-Three Years of Measuring Spend Never Measured Judgment

Executives tell the same story with their budgets. 69% of firms actively use AI [11]. Yet when the Bank of England, the Atlanta Fed, Stanford and Germany’s Bundesbank jointly asked roughly 6,000 senior executives — more than 90% of them C-suite, across the US, UK, Germany and Australia — more than 90% reported no impact of AI on their firm’s employment over the past 3 years, and 89% reported no impact on labour productivity [11]. The same executives still predict AI will lift their productivity 1.4% over the next 3 years [11].

The Federal Reserve Bank of Atlanta measured the gap between belief and reality. CFOs told the Fed that AI raised their productivity 1.8% in 2025. The Fed then computed what those same CFOs’ own revenue and headcount figures implied: 0.6% [12]. Executive belief runs three times ahead of what the executives’ own numbers show. That is the gap this Call names: everyone feels the return; almost nobody can find it in the arithmetic.

The US Census asked firms using AI what they actually changed after buying the technology [13]:

• 64% changed nothing at all — no new training, no new workflows, no institutional adjustments [13].

• About 15% trained staff, and about the same share developed new workflows [13].

• 7–8% made the deeper shifts, including reorganising their data so AI can use it, and the supporting systems and equipment AI needs to do its work [13].

Firms bought the tool. Most never retooled.

Economists have known this for 23 years. In 2003, Brynjolfsson & Hitt measured it across 527 large US firms: the organisational investment that makes computing pay “may be up to 10 times as large as the direct investments in computers” [14]. The payoff also compounds slowly — up to 5 times greater when the researchers measured over 5-7-year horizons rather than 1 [14].

In 2012, the American Economic Review published the explanation. Bloom, Sadun and Van Reenen studied more than 11,000 UK workplaces and asked why US-owned firms got more out of computers [15]. Management explained it. When a US multinational doubled its computing, productivity rose 6.3%; when a non-US peer did the same, 4.6% — a 1.7-point gap [15]. Tougher people-management practices explained it; once the researchers accounted for management quality, the American ownership advantage disappeared entirely [15].

The 2026 evidence repeats the lesson for AI adoption. The OECD — the Organisation for Economic Co-operation and Development, the research body for the world’s advanced economies — compared AI adopters against non-adopters across 15 countries [16]. The raw advantage looks impressive: 7.7% higher productivity in France, up to 31% in Belgium [16]. Then the researchers controlled for the quality of each firm’s people and technology — and in 8 of 10 countries the AI advantage vanished. Only 2 of 10 kept a significant edge [16].

The same European firm data puts a price on the fix. When a firm moved 1 extra percentage point of spending into training, AI’s productivity effect rose about 5.9%. The same point spent on software and data bought 2.4% — training beat it by more than double. Spent on R&D or machinery: nothing measurable [17]. One honest clause: the researchers measured this as an interaction across firms, not a controlled split, and the effect is statistically zero for firms under 50 employees [17].

Here is my analysis of that record: The ROI literature has been pointing at the human variable for 23 years. It measures training spend, specialist headcount, management indices, workflow redesign. Not one study measures whether the judgment improved — or what it returns when it does. [14][15][16][17]

The Best Causal Evidence: Gains Concentrate Where Judgment Matters Least

The strongest objection deserves the floor before you hear the answer. Brynjolfsson, Li and Raymond published the best causal study in the field in the Quarterly Journal of Economics, 2025. They followed 5,172 customer-support agents — human agents, people answering customer chats, not AI agents — at one Fortune 500 software firm as it rolled out an AI assistant team by team [18].

Because the firm staggered the rollout, the researchers could compare teams working with AI against near-identical teams still waiting for it — a natural experiment rather than a randomised trial [18]. Their finding: “Less experienced and lower-skilled workers improve both the speed and quality of their output, while the most experienced and highest-skilled workers see small gains in speed and small declines in quality.” [18].

The average human agent resolved 15% more issues per hour. The least experienced resolved 30% more — and a newcomer with 2 months of tenure performed like a veteran of more than 6 months [18]. The Bank of Korea heard the same from its survey of 5,512 workers: the less-experienced saved the most time [19]. And AI pays — the European researchers tie adoption to roughly 4% higher labour productivity [17]. Nothing in this Call argues otherwise.

Those gains live in routine, well-mapped tasks. The sharpest test came from 758 consultants at BCG, the Boston Consulting Group, in a preregistered randomised experiment [20]. The researchers built two kinds of tasks and named the boundary between them the frontier. Inside the frontier means work AI handles well — idea generation, drafting, analysis. Outside the frontier means work that looks similar but sits beyond what the tool does reliably — in the experiment, a business-problem-solving task [20].

On inside-the-frontier work, consultants using AI completed 12.2% more tasks, worked 25.1% faster, and scored more than 30% higher on quality [20]. On the outside-the-frontier task, consultants without AI answered correctly 84.5% of the time; consultants with AI, 60% and 70.6% — a 19-point drop, because they trusted the tool past the line where it stops working [20].

The authors’ conclusion should be required reading for anyone designing and testing AI transformation strategies: “The effectiveness of AI in knowledge work will critically depend on human judgment—particularly to discern which tasks within the workflow are suited to leveraging AI augmentation and where human expertise should be prioritized.” [20].

In other words: the tool cannot tell you where its own competence ends. A person decides which work goes to the AI and which work a human must own — and that person’s judgment decides whether your firm collects the 25.1% or eats the 19-point drop [20]. Novices gain speed and quality on the routine work; your senior people’s judgment gates the non-routine, high-stakes decisions — which is why judgment is the lead lever of AI return, never the sole one.

The Missing Study: Restore Judgment, Then Measure the AI Return

Every strand of this literature stops at the same wall. I searched the corpus in ten languages, as of publication date, and four strands come back empty:

1. Restore judgment, then measure the dollar return: no study, in any language searched, as of publication date [20].

2. The erosion priced in currency: no researcher has put a dollar figure on dulled judgment — the closest proxies stop at accuracy collapse [8] and leader-perceived threat [7].

3. Reversibility: every study measures one moment in time; nobody has watched judgment come back and value follow [21].

4. Profit tied to decision quality: the field measures capability as money spent, heads counted, indices scored [14][15][16][17] — never as decisions improved.

One team came closest to testing the fix itself. In 2021, Buçinca, Malaya and Gajos ran a 199-person controlled experiment on a simple question: if you force people to pause and think before accepting an AI’s answer, do they stop over-trusting it? They do. The forced-pause checks “significantly reduced over-reliance” [21]. Participants disliked the extra effort, and the participants who most enjoy hard thinking gained the most [21]. Here is what that means for you: the training fix works — and science has never connected it to money. The experiment measured better decisions, on nutrition questions, with laypeople. No one has run the next study: the same intervention, inside a company, measured in how better judgment moves your P&L [21].

Researchers can show that firms investing in people get more from AI — the training multiplier [17], the vanishing premium once you account for human capital [16]. No researcher has shown the last step: rebuild the judgment, then watch the return arrive. The five authorities walked up to exactly that missing step and stopped [1][2][3][4][5]. I have priced this layer before, in The Judgement Premium, and counted who occupies it, in The Sophistication Gap.

Is the Judgment Behind Your AI Spend Improving — or Quietly Eroding?

That is the Monday question, and your current reporting most likely cannot answer it. Your AI dashboard names licences, adoption rates, training hours — every proxy the literature has measured for 23 years [14][15][16][17]. Nowhere on it: whether the judgment your spend depends on got sharper this quarter, or duller. Solis’s unanswered half of the question is sitting in your own deck right now [1].

Building judgment is a different discipline from guarding it, and the evidence already names its composite in these three moves:

1. Give every non-routine, high-stakes decision a named owner and a written standard. The consultant experiment showed where accuracy collapses when nobody owns the call: on outside-the-frontier work, the decisions beyond what the tool does reliably [20].

2. Make humans deliberate and challenge AI outputs before accepting them — at the very least as part of training. It is the one intervention measured to cut overreliance and a cost-effective way to keep reviewers sharp thereafter [21].

3. Put a judgment line in the AI budget and instrument it. Training investment multiplied AI’s productivity effect more than any other spending the European researchers tested [17].

You can answer the half Solis left blank — how will we expand human judgment? — by doing exactly this: name the owners and score their calls against the written standard, quarter over quarter [20]. Train the challenge habit and count how often your reviewers catch the machine’s misses [21]. Fund the judgment line and measure it like any other investment [17]. How to run that measurement inside your own P&L — the study your firm can run on itself — is territory for a coming Call.

The AI Leadership Playbook

Strategic Questions (copy-paste ready for an email to your CFO and CHRO):

1. Are we in the 64% that changed nothing — what institutional adjustments have we actually made since deploying AI, and who owns the complete list? [13]

2. Which decisions in our AI-touched workflows are non-routine and high-stakes — outside the frontier — and have we named the person accountable for each one? [20]

3. What share of our AI budget develops judgment at all, and what would tell us this quarter whether the judgment improved? [17]

Your Next Plays (copy-paste ready for a direct report):

1. Run the 64% check. Inventory what changed since AI arrived against the Census categories — training, new workflows, data management [13]. Every empty row gets an owner.

2. Draw the frontier line. For one revenue-critical workflow, list every decision AI touches and mark each one inside the frontier (work AI handles well) or outside it (work beyond what the tool does reliably) [20]. Every outside-the-frontier decision gets a named human owner and a written standard.

3. Put a judgment line in the AI budget. Reclassify a slice of AI spend to judgment development and instrument it: challenge-before-accept checks measurably cut overreliance [21], and training carries the largest multiplier the European evidence found [17]. This is The People Bet, carried to the judgment layer.

📅 Book a complimentary 1:1 Strategy Session—45 minutes to start the conversation about mapping the revenue hidden in your AI spend.

Sources

1. Brian Solis, ServiceNow — The CEO Guide to Using AI: Don’t Automate Your Way Out of the Future, 2 August 2026 — https://briansolis.com/2026/08/the-ceo-guide-to-using-ai-dont-automate-your-way-out-of-the-future/

2. David Lancefield, Harvard Business Review — Don’t Let AI Flatten Your Leadership Style, 3 August 2026 — https://hbr.org/2026/08/dont-let-ai-flatten-your-leadership-style

3. Tomas Chamorro-Premuzic, Forbes — Developing Leaders For The Human-AI Age, 4 August 2026 — https://www.forbes.com/sites/tomaspremuzic/2026/08/04/developing-leaders-for-the-human-ai-age-why-potential-remains-key/

4. Keith Ferrazzi and Wendy Smith, Fortune — Your AI agent can be a teammate. But it still needs a boss, 4 August 2026 — https://fortune.com/2026/08/04/your-ai-agent-needs-a-boss-is-a-teammate/

5. Helle Bank Jørgensen and Marlen Heide, World Economic Forum — The hybrid boardroom, 5 August 2026 — https://www.weforum.org/stories/artificial-intelligence/hybrid-boardroom-ai-changing-role-directors/

6. Shaw & Nave, Wharton School Research Paper — Thinking—Fast, Slow, and Artificial, posted 2 February 2026 — https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6097646

7. Goel, Martin and Kaffe, BCG Institute — When Everyone Uses AI, Companies Risk Losing Critical Skills, 17 June 2026 — https://web-assets.bcg.com/pdf-src/prod-live/when-everyone-uses-ai-companies-risk-critical-skills.pdf

8. Dratsch et al., Radiology — Automation Bias in Mammography, 2023 — https://pubs.rsna.org/doi/10.1148/radiol.222176

9. Ehsan et al., CHI 2026 — From Future of Work to Future of Workers, 2026 — https://dl.acm.org/doi/10.1145/3772318.3791081

10. Parasuraman & Manzey, Human Factors — Complacency and Bias in Human Use of Automation, 2010 — https://journals.sagepub.com/doi/10.1177/0018720810376055

11. Yotzov et al., NBER Working Paper 34836 — Firm Data on AI, February 2026 — https://www.nber.org/system/files/working_papers/w34836/w34836.pdf

12. Baslandze et al., Federal Reserve Bank of Atlanta Working Paper 2026-4, 25 March 2026 — https://www.atlantafed.org/research-and-data/publications/working-papers/2026/03/25/04-artificial-intelligence-productivity-and-the-workforce-evidence-from-corporate-executives

13. Bonney et al., US Census Bureau CES Working Paper CES-26-25 — The Microstructure of AI Diffusion, April 2026 — https://www2.census.gov/library/working-papers/2026/adrm/ces/CES-WP-26-25.pdf

14. Brynjolfsson & Hitt, The Review of Economics and Statistics — Computing Productivity: Firm-Level Evidence, November 2003 — https://www.iecon.net/wp-content/uploads/2015/01/cpg.pdf

15. Bloom, Sadun and Van Reenen, American Economic Review — Americans Do IT Better, February 2012 — https://www.aeaweb.org/articles?id=10.1257/aer.102.1.167

16. Calvino, Costa and Haerle, OECD STI Working Papers — Digital technology diffusion in the age of AI, 21 January 2026 — https://www.oecd.org/content/dam/oecd/en/publications/reports/2026/01/digital-technology-diffusion-in-the-age-of-ai_7f11be5d/ebc2debe-en.pdf

17. Aldasoro et al., BIS Working Papers No 1325 — AI adoption, productivity and employment, January 2026 — https://www.bis.org/publ/work1325.pdf

18. Brynjolfsson, Li and Raymond, The Quarterly Journal of Economics — Generative AI at Work, 2025 — https://doi.org/10.1093/qje/qjae044

19. Suh, Oh and Kim, Bank of Korea Issue Note 2025-22 — Rapid Adoption of Artificial Intelligence and Its Productivity Effects, 2025 — https://www.bok.or.kr/eng/bbs/B0000354/view.do?nttId=10094689

20. Dell’Acqua et al., Organization Science — Navigating the Jagged Technological Frontier, March 2026 — https://doi.org/10.1287/orsc.2025.21838

21. Buçinca, Malaya and Gajos, CSCW — To Trust or to Think, April 2021 — https://www.eecs.harvard.edu/~kgajos/papers/2021/bucinca21trust.pdf

Conor Grennan, who founded AI Mindset and teaches at NYU Stern, named the phenomenon on July 31⁷. “Every other technology you’ve ever bought came with its own meter in the box,” he wrote. AI shipped without one. Nothing on the invoice tells you what the judgment behind the output was worth.

Your return did not go missing. It is hiding, and that is a different problem with a different fix. The stack is the machine, and the machine is running dull. Judgment is the edge to sharpen.

The Return Isn’t Gone—It’s Hiding in the Judgment Layer

The return is not hiding in the model. It is hiding in the decisions that turn AI output into priced, shipped, invoiced work. Which draft ships. Which forecast finance acts on. Which confident but hallucinated answer a person kills before a customer ever sees it.

Four instruments that leverage or evaluate judgment landed in seven days. Each one points at that crucial decision layer; I’ll name each one here and reference back as I lay out the rest of my analysis.

1. Finn’s WEF piece, July 29. Jessica Finn leads health, life sciences and education for Cognizant in Australia. On the World Economic Forum‘s site, she told companies to rebuild work into “development loops, where judgment, context and accountability are what is learned, tested and trusted”¹. She prescribes the loop. She never names who runs it.

2. Grennan’s meter, July 31. The diagnosis above, with a prescription attached: define what your people produce today, then name which numbers should move if AI made them faster⁷. He stops at setting the target.

3. Carter’s judgment audit, August 1. David Paul Carter advises founder-led companies between $2M and $15M in revenue. He published a six-field decision journal: the call, the prediction, your confidence, and what would prove you wrong⁴. It grades a founder’s own past decisions.

4. The HBR finding, July 28. Harvard Business Review reported on a study of early-career professionals⁵. How junior staff actually use AI day to day predicts how well they perform. It predicts better than the skills on their résumé, and better than any AI training they have taken. Behavior beats credentials.

One source sits underneath two of those four, so it earns its own grounding. The Cognizant–Pearson survey, published June 18, is Wakefield’s poll of 750 HR leaders at director level and above, at companies of 1,000 employees or more in the US, UK and India². It supplies every workforce figure in this Call, and it measures what HR leaders expect rather than what anyone observed. The distance between what they expect and what they fund is the finding.

That timing matters. The WEF pages I drew on three weeks ago for The Apprenticeship the Machine Ate went further this week¹⁰, but Finn’s argument rests on that same March–April fieldwork. Her codification is new; the evidence is not.

All four measure judgment or prompt it. None of them converts it into a measurable P&L lever. Finn prescribes the “loop” and leaves the owner’s chair empty. Grennan hands you the meter and stops at the target. Carter grades a founder’s own calls. HBR tells you to train people but never says who decides what those people must learn to judge.

I named this pattern in The Judgement Premium on June 3, 2026, and carried it through The Sophistication Gap a week later^8,9. Last week four separate practitioners shipped instruments aimed at that same layer. The pattern is no longer mine alone to argue.

What a Dull Judgment Layer Costs

The four instruments describe the gap. Nirit Cohen explains how it opened. Cohen held senior global people roles at Intel. She now runs Workfutures, a practice on the future of work. On Ross Dawson’s Humans + AI podcast on July 29, she named the mechanism⁶. Cohen argued that people used to earn judgment by doing the work that qualified them to review someone else’s. When you automate that work the qualification never happens. “If I’ve never seen what good versus bad looks like, how will I be able to tell?”

Dawson, the futurist I credited in The Judgement Premium when I gave that metric its three indicators, took her point one step further. He argues that when you park people as approval mechanisms without the guidance to discern output quality for themselves, it’s “naturally going to erode judgment and expertise”⁶.

There’s a step beyond the mechanistic “loop” supervision Dawson and Cohen describe that is not captured by their argument. If a “loop” is meant to make a process more efficient, that is, save time–who determines how those saved hours are spent? Saved hours do not magically convert themselves into output that moves your P&L. Automate the work, leave the judgment layer dull, and the machine runs at full speed while your workforce loses the ability to judge what it produces and the organization squanders the opportunity to re-direct efficiencies towards revenue.

Two surveys now price that lack of ownership, one from inside the workforce and one from the leadership suite.

Inside the workforce

The Cognizant–Pearson survey measures what the people closest to the work expect, and what their organizations are funding against it.

1. Companies fund the expectation at half strength. 96% of HR leaders expect entry-level roles to become jobs that supervise AI within 5 years. 46% are not proactively investing in AI training^1,2. That is a 50-point gap between the future they say is coming and the development they are investing in to actually reach it.

2. Demand outruns capacity. 91% report that employee requests for AI training rose over the past year. Another 60% say their learning and development programs cannot keep up with how fast AI is changing job functions².

3. The hiring release-valve is closing. 64% say they cannot find the right talent, because AI keeps changing what they need to hire for².

From the leadership suite

The KPMG Pulse finds the same shape among the people who sign the cheques. KPMG surveyed 2,145 senior leaders across 20 markets, at companies above $100M in revenue, between late April and late May³. Two of its numbers belong on your board deck.

1. Only 7% report established ROI on AI. Nearly one in four, 24%, already face investor pressure to prove value³.

2. Only 24% say their CEO is accountable for AI-driven business outcomes. Another 29% point to “the broader C-suite,” which KPMG reads, in its own words, as responsibility that “often stops at the sponsorship level rather than true accountability”³.

Then the number that should stop a board meeting. Companies that name accountability clearly report established ROI at 14%, against 4% for companies that do not³. That is 3.5 times the rate and a 10-point spread. KPMG sells AI advisory, so weigh their framing accordingly, and note what the numbers do on their own: the returns materialize 3.5x when there’s a named owner for AI ROI in the leadership suite.

The Four Motions That Capture and Build Return in the Judgment Layer

If the return is hiding in the judgment layer, then capturing it is a sequence rather than a purchase. Four moves, in order:

1. Locate. Find the decisions where human judgment turns AI output into revenue, cost, or cycle time. Most organizations have never written that list down.

2. Name. Say who owns each of those decisions today, and what it would cost you to lose them. An owner turns a capability into something one person owns.

3. Sharpen. Aim development at sharpening those specific decisions rather than at general AI fluency. Not funding training is a problem; funding the wrong training is a more expensive one.

4. Instrument. Put the return on board-reportable metrics. The Judgement Premium named three: decision-cycle reduction, decision-reversal rate, and board-level visibility.⁸

Judgment is the lead lever in that sequence. It is not the only one though. Role redesign, clean data, psychological safety, training and many other factors capture hidden revenue. A serious read of a business finds the factors that unlock the most return.

I’ve named and instrumented that sequence as The AI ROI Map: a systemic analysis that determines where your AI spend is paying, where it’s forfeiting return, and the moves that capture it. This is the missing motion in the expert analysis and survey data so far. Every instrument that shipped last week^1,4,7 points at the gap and offers no mechanism for finding or capturing the revenue. The AI ROI Map completes the motion and uncovers the ROI hiding in your AI spend.

Carter’s founder judgement audit instrument is the sharpest on decision refinement, and he is clear about its limits. “The ideas are theirs”; he writes of Tetlock’s calibration research and Annie Duke’s kill criteria⁴. “The format is mine,” he concludes. A founder scores a personal record in five minutes per decision. That is a good instrument, but it was never built to be an organizational-wide audit.

Where the Field Agrees, and What It Leaves on the Table

The first objection: Finn already delivered quietly. Her definition contains the word: judgment, context and accountability as the instructive asset in the automation loop.

Finn’s loop teaches accountability. It does not name an owner. Those are two different acts. A curriculum tells people what accountability means; governance names the person who owns it. Read her piece end to end and you will not find accountability defined as owned responsibility for outcomes. No one owns the loop. No one grades the judgment it produces. No one decides what judgment the business will need next quarter.

Cohen in conversation with Dawson is the most candid voice in the whole conversation. Dawson asked her directly how organizations grow expertise when nobody does the work that used to build it. Her answer, on the record: “I don’t know that there is an answer yet”⁶. The field’s sharpest thinker on judgment formation has named the problem and stopped at the edge of the answer.

The second objection is structural: scholars, practitioners and researchers point to AI failure statistics. A shortfall is not a strategy. Counting what failed only uncovers part of the story; it doesn’t tell you where ROI went, or how to capture it.

The field does agree on direction, however. HBR concluded that leaders must train people to work with AI rather than assume skilled employees will figure it out⁵. Kathy Diaz, Cognizant’s CHRO, said the same of her own function: AI “is reshaping the talent landscape and exposing the limits of traditional talent and learning models”². Everyone agrees on the diagnosis. Nobody decoded the capture mechanism.

So when you sit down with the AI implementations your team built this quarter, the question is not which instrument to buy next. Three motions actually replace that procurement question:

1. Define which decisions still turn AI output into revenue,

2. Decide who owns those decisions specifically, and

3. Be strategic about the capabilities your training budget actually develops.

This is one of the sequences I advise on. I don’t promise AI ROI. This is one of the ways I instrument it. The Human Dividend is the compounding return on the human capability your AI investment already depends on.

In The Apprenticeship the Machine Ate I closed by saying that the measurements to rebuild that junior talent on-ramp was territory for a coming Call. This is that Call: The AI ROI Map reveals where your AI spend is paying, where it’s forfeiting return, and instruments the moves that capture it. The machine is already yours. The first move is to sharpen the judgment layer that keeps your AI implementation paying.

You’ve already done the hard part—you know which decisions in your business actually matter. Sharpen that edge and start capturing more AI ROI.

The AI Leadership Playbook

Strategic Questions [Copy-paste ready for an email to your CFO and CHRO.]

1. Which decisions in our operation still require human judgment to turn AI output into revenue, cost savings, or reduced cycle time; and where is that list written down? If it is not written down anywhere, who is drafting it this quarter?

2. Who owns each of those decisions today, by name, and what would it cost us if we lose them? Start by naming who we’d hire if those decision owners resigned on Friday.

3. What is our AI training budget aimed at, specifically? Tool fluency, or sharpening the decisions above?

Your Next Plays

1. Run a judgment inventory on one workflow, not the whole company. Pick the workflow where AI output sits closest to revenue. List every decision a person still owns, and mark the ones a new hire could not make. That marked list is your judgment layer.

2. Put a name next to the automation. Every AI deployment gets an owner. Each owner scores that deployment’s output against a codified standard, meaning a written description of what good looks like for that output in your business. An unowned AI deployment reports activity. An owned one reports capability.

3. Instrument three decisions before you instrument the enterprise. Track how long they take, how often someone reverses them, and how often they reach the board. Three tracked decisions give you a baseline; a dashboard with no baseline gives you a chart.

📅 Book a complimentary 1:1 Strategy Session—45 minutes to start the conversation about mapping the revenue hidden in your AI spend.

Sources

1. Jessica Finn (Cognizant), “How is AI changing the skills for leadership, and how should organizations prepare?” World Economic Forum, July 29, 2026.

2. Cognizant and Pearson, “Entry-Level Work Remains Essential: 94% of HR Leaders Expect AI to Create New Entry-Level Roles,” June 18, 2026 (Wakefield Research, n=750).

3. KPMG International, “Growing adoption signals progress as cost visibility and accountability drive AI value,” Global AI Pulse Q2 2026, June 24, 2026 (n=2,145 senior leaders, 20 markets).

4. David Paul Carter, “The Judgment Audit, Part 2: Building the Scoreboard,” August 1, 2026.

5. Thomas Stackpole, “Why Some Junior Employees Work Well with AI and Others Don’t,” Harvard Business Review, July 28, 2026.

6. Ross Dawson with Nirit Cohen, Humans + AI, Episode 52, July 29, 2026.

7. Conor Grennan, “How to Actually Measure ROI of AI,” AI Mindset, July 31, 2026.

8. Paola Sanmiguel, “The Judgement Premium,” CognivaLab, June 3, 2026.

9. Paola Sanmiguel, “The Sophistication Gap,” CognivaLab, June 9, 2026.

10. Paola Sanmiguel, “The Apprenticeship the Machine Ate,” CognivaLab, July 14, 2026.

EDITORIAL NOTE: From the beginning these have been fast-developing stories; all information provided below is as reported and corroborated as of 11:45pm PT on 27 July 2026. To recap, the following led to the legal questions discussed below: the letters that terminated Anthropic’s Department of War contract, the letters that shut off Fable 5 and returned it 19 days later, and the release of Kimi K3—its full weights delivered for download on 27 July. This analysis is meant to help leaders map which laws govern the data moving through each part of their AI workflow. It is not legal advice; always seek counsel in your jurisdiction.

On Monday, Beijing time, Moonshot AI placed the finished settings of Kimi K3—96 files, roughly 1.6 terabytes—on the open internet for anyone to download.¹ Within hours, copies were spreading to mirror sites no one controls. The same day, China’s commerce ministry answered Washington’s sanction threats with a warning that it would take “all necessary measures” to defend its AI companies.² Six days from now, on August 2, the European Commission gains its first powers to fine the makers of frontier models.³

Against that backdrop, more than fifty American technology companies—Nvidia, Microsoft, and OpenAI among them—signed an open letter last week asking Washington not to restrict open models,⁴ while the administration weighed doing exactly that.⁵ Inside your organization, the debate is simpler: how does this evolving legal landscape affect our ability to operate?

Reporters have wondered if Chinese models should be used at all. That is the wrong question to ask. The law does not begin by asking who built the model. It asks how your data reaches it—and that answer was set by your data architecture team, before anyone briefed you.

Your decision to host a model or call it with an API has implications far greater than IT infrastructure—that decision determines the law that can be applied to your data.

The Door Decides Your Legal Exposure: Hosted API or Downloaded Weights—Not the Flag Flying Over the Lab

The debate usually starts with the flag. An American model feels governed; a Chinese model feels risky. It may be a reasonable instinct, and it points at the wrong variable.

Every model reaches you through one of two doors. Behind the first, you call a hosted API: your prompts—your contracts, your code, your customer records—travel over the internet to computers the provider operates. The data movement is not a side effect; it is the product of that architecture.

Behind the second door, you download the model’s weights—the billions of internal settings adjusted during model training, which I explained in It’s Not Thinking. It’s Predicting.⁶—and run them on servers you control. A weights file opens no connection back to the lab that produced it. Whether anything leaves your building is decided by your own network controls—and by whether you hand the model tools and internet access, which reopens the door from the inside.

Kimi K3 now exists on both sides of that divide, and the difference is instructive. Call its hosted international API and your prompts go to Moonshot AI Pte. Ltd.—a Singapore entity, with servers located in Singapore, by the provider’s own published policy.⁷ The reflex “your data goes to China” is wrong for that endpoint. But that fact belongs to Moonshot specifically, not to Chinese models generally, and that is the transferable lesson: the only way to evaluate exposure is to read, provider by provider, where the servers sit and what the service agreement permits. Three questions belong in every provider review:

1. Server location. Where are the servers located, according to the provider’s own published policy—not its marketing page?

2. Rerouting rights. Can the provider move your traffic to servers in another jurisdiction, and under what conditions?

3. Notice. Does the agreement require client notification before that rerouting happens—if so, what?

Download the same model’s weights and host on-prem instead, and every one of those variables are now under your control: your servers, your jurisdiction, your controls.

So the first answer is this: the door your data goes through—hosted API or downloaded weights—selects your legal exposure before anyone reads a statute. Which is why the inventory that matters is not a list of model names. It is an inventory of the AI providers and models you are currently using, classified by the “door” you use to access them.

Jurisdiction Follows Control: The Law That Reaches You Is the One That Can Compel Your Operator

Now the second question: once you know the door, whose law stands behind it?

The instinct here is geographic—my data sits in Frankfurt, so German law applies. Courts stopped asking that question years ago. Three legal bases decide who can reach your data, and they are not equal.

1. Where the server sits. The weakest basis, and the one executives still reason from. It is the belief this Call’s title corrects.

2. Who controls the data. The strongest. The US CLOUD Act requires American providers to hand over data in their “possession, custody, or control” whether it is stored “within or outside of the United States.”⁸ Congress wrote that law in 2018 for a specific reason: Microsoft had refused a warrant for emails on a Dublin server, and the server’s location worked as a defense.⁹ The CLOUD Act exists so that defense never works again.

3. Whose citizens’ data it is. Europe’s GDPR follows EU residents’ data beyond Europe; China’s personal-information law (PIPL) does the same for people in China.^{10 11}

Let’s examine each region through the lens of these three provisions.

1. The United States carries the broadest reach: the CLOUD Act through providers, plus Section 702 of its surveillance law, which authorizes warrantless collection targeting non-Americans abroad, at scale.¹²

2. The European Union points its law inward as a shield: the GDPR restricts data from leaving, blocks foreign court orders that arrive without a treaty behind them, and—after its top court struck down a US data deal over that same Section 702 surveillance—polices American reach as a risk to Europeans’ data.¹⁰

3. Singapore sits closest to territorial: its Personal Data Protection Act governs companies, exempts the government, and state access runs through investigation-tied criminal procedure rather than intelligence collection at scale.¹³

And the provider’s home jurisdiction? For a PRC-domiciled lab like Beijing Moonshot—the parent company behind that Singapore API entity—four separate legal objects get collapsed into one alarming shorthand, and precision matters more than comfort in this instance.

1. China’s National Intelligence Law obliges organizations to “support, assist, and cooperate” with intelligence work.¹⁴

2. No Chinese statute mandates a backdoor—none has been found that says so.

3. Data-localization duties bind critical-infrastructure operators, not every company; and

4. Blocking provisions forbid handing data stored in China to foreign authorities without Beijing’s permission.¹¹

Support duty, backdoor mandate, localization, blocking—four different things. Only the first is what most coverage means by “Chinese law reaches everything.”

Here are the experts, on all sides, so you can decide what your organization is comfortable with. Donald Clarke, a George Washington University law professor specializing in Chinese law, examined the legal declaration Huawei commissioned in its own defense and found it “incomplete”—it never addresses “the extent to which the Chinese government is constrained by Chinese law.”¹⁵ Jeremy Daum of Yale Law School’s China Center, who runs the China Law Translate project, is blunter: the statute text is beside the point, because the state can coerce “with or without a law.”¹⁶

On the measured side, Samm Sacks and Peter Swire—two of America’s most cited data-policy scholars—argue for judging each data flow through an evidence-based framework rather than by blanket country rules.¹⁷

And Beijing’s foreign ministry states for the record that China “has never required, nor will it require” companies to hand over overseas data “in violation of local laws”—a denial whose last clause does quiet work.¹⁸ Accept Clarke and Daum, and no PRC provider’s assurance is bankable. Accept Sacks and Swire, and the exposure is real but assessable per deployment. Your legal counsel, your local laws, your data architecture and your risk tolerance should ultimately determine how to navigate this evolving landscape.

A counterargument is worth surfacing before we go further. Most executives reading this will think: we use only American models, under American law, with American courts—this is someone else’s problem. But staying American does not remove the government behind the door; it selects which one. The strongest compulsion power in this entire map—control-based, borderless—is American. For a US company, that may be familiar, acceptable exposure. For the same company’s European subsidiary, it is precisely the exposure EU regulators police. Neither reading is wrong. The decision is which government’s reach your organization can accept, given where you operate and whose data you hold.

One structural detail changes that calculus more than any statute, and it needs plain language. Under the GDPR’s “one-stop-shop” rule, a provider with a European headquarters answers to a single country’s privacy authority as its main supervisor. A provider with no European entity at all can be pursued by every one of roughly thirty national authorities, independently, each able to act in days. DeepSeek learned this in January 2025, when Italy’s authority moved alone.¹⁹ Where your provider is incorporated is not trivial. It is the difference between being answerable to one regulator or thirty.

If you cannot name which government can compel your model’s operator, how do you defend the data your board thinks is protected?

What You Can Change After You Deploy: Everything Around the Model. Nothing Trained Into It.

The third question sounds technical and is actually contractual: after implementation, what can you still change?

With a hosted API, the honest answer is prompts and scaffolding—the instructions you send and the software you wrap around the call. The model itself, its safeguards, its update schedule: all the vendor’s, changeable by the vendor, on the vendor’s calendar.

Downloading the weights inverts most of that. You can fine-tune the model, wrap it, route around it, and—the advantage security researchers keep pointing to—audit its behavior on your own terms. What you cannot do, and this is the boundary that matters, is see or alter what was trained into it. No frontier lab, American or Chinese, publishes its training data or the full method behind its weights; by the standard open-source definition, no frontier “open” model actually qualifies.²⁰ What a model was trained to do is invisible to the buyer before deployment and unchangeable after.

Whether that matters is not a judgment about any country’s politics—it is a fact about where the model will be used. Behavior a lab’s home market considers normal may surprise a buyer elsewhere, and may sit badly with the laws, rules, or expectations of the locales where your business operates.

The clearest measurement of this comes from CrowdStrike, one of the largest American cybersecurity firms, whose researchers tested DeepSeek’s open-weight model—not K3—the hard way.²¹ They ran the raw downloaded weights directly, deliberately bypassing the API-level guardrails, through 30,250 coding prompts.

Asked for ordinary technical work, the model produced code containing known security flaws 19% of the time—the baseline.²¹ When the same requests mentioned politically sensitive subjects—CrowdStrike’s published example is code “for an industrial control system based in Tibet”—the flaw rate rose to 27.2%. Concretely: roughly one output in four carried a security flaw, up from one in five, triggered by words that had nothing to do with the code being requested.

The model also flatly refused certain politically framed requests despite first reasoning out a complete technical plan. CrowdStrike’s conclusion was architectural: because the tests ran on raw weights, the behavior “must be baked into the model weights.” Their own caveat belongs beside it: this is a long-run average, not a property of every response.

Semgrep, a code-security firm that tested K3 directly, reported “no evidence of security backdoors in open weight models”—and called auditability the single biggest security advantage of the open-weight door.²² The real-world attacks documented so far have all come through the distribution channel—tampered files on model hubs—not through a released model’s weights. Verifying your download against the publisher’s official repository is therefore not paranoia; it is the control. You can test behavior, inspect the architecture, and verify file integrity. You cannot inspect the training data or prove a trigger absent.

Here is what that means operationally: choosing the self-hosted door moves the discovery burden from the vendor’s roadmap onto your evaluation budget. And K3 arrives with none of that work done for you: its 45-kilobyte model card ships benchmark tables, compression notes, deployment guides, a license—and no safety section at all.¹ The only published safety evaluation of K3 is the one the UK and US governments ran, which found its safeguards did not stop it from attempting cyber exploit development.²³ That is not a verdict on the model. It is the fact that defines what you are signing up to test yourself.

Behavior trained into a model cannot be seen by the buyer, changed by a prompt, or removed by the vendor. It travels with every copy.

Any Government Can Switch Off a Model Service. None Has Yet Reached an Open-Weight Model on Your Own Servers.

The fourth question is the one made urgent last month: who—anywhere—can actually switch a model off? The record has four jurisdictions in it, and one pattern.

1. Washington: a letter, then darkness in hours. On June 12, a letter from the Commerce Secretary reached Anthropic at 5:21 p.m. Eastern; by that evening, two frontier models were dark for every foreign national on earth—on Amazon’s cloud, Google’s, and Microsoft’s simultaneously.²⁴ No published rule, no comment period, no appeal schedule. I walked through that episode and its partial reversal in Not a Screwdriver or Uranium,²⁵ so here I will only add what it means operationally: a hosted dependency can vanish inside one business day.

The standard contract gives you almost nothing back: the right to suspend pre-existed the June shutdown, no compensation is owed, and liability is capped. The cost is continuity: untested fallbacks, stranded work, renegotiation from weakness. One boundary keeps this honest: those were closed-weight models, existing only on their maker’s servers. Had their weights been public and downloaded, the order could have reached future API access and US distribution—not the copies already downloaded. To be clear, that is an inference that has never been tested.

2. Europe: two days to stop a service—and no model on anyone’s servers was touched. In January 2025, Italy’s data-privacy authority ordered DeepSeek to stop processing Italian users’ data, two days after opening questions.¹⁹ Read the fine print, because most coverage got it wrong: DeepSeek had pulled its own app from Italian stores the day before the order; the website stayed up; and eighteen months later, no fine has been published. The instrument reached the hosted service’s legal permission to process data—not the model, not any copy of it.

On August 2, a second instrument arrives: the European Commission gains the power to request withdrawal or recall of a general-purpose model, made binding only through fines of up to 3% of worldwide turnover.³ Arriving, not arrived—the power has never existed, so it has no track record. Watch the quieter route too: from the same date, a missing registration or missing documentation—no harm shown—can oblige national authorities to pull a model from the EU market. Administrative, not dramatic—and the likelier path.

3. Beijing: the strongest takedown on record left every installed copy running. In July 2021, four days after the ride-hailing giant DiDi listed on the New York Stock Exchange, China’s cyberspace regulator ordered its app off Chinese app stores. The company’s own disclosure states the boundary precisely: the app could “no longer be downloaded,” but users who had already installed it “may continue using it.”²⁶ New copies stopped; existing copies ran. For generative AI, the record is emptier still: China’s rules allow service suspensions, and no verified instance of one exists.²⁷

4. Singapore: the fullest toolkit, never aimed at AI. Singapore’s online-harms law grants directions no EU regulator has—administrative orders, no court required, reaching internet providers, app stores, even payment and advertising support.²⁸ It has never once been pointed at an AI model. That is restraint—a policy choice, not a missing capability—and choices can change.

Read the four together and one pattern holds: what governments have actually reached, every time, is something that connects you to a provider—an API, an app store, a legal permission to operate. Which raises the question you should ask next: could a government go further and outlaw using an open-weight model already sitting on your servers? Here is what is known, plainly. No such rule exists today, anywhere.

The instruments to attempt one exist: Congress banned US hosting and distribution of TikTok and the Supreme Court upheld the law in January 2025—though it reached app stores and hosting services, not users’ installed copies²⁹—and Washington has reportedly weighed export-blacklist designations that could require a license to possess the models.⁵ Whether a use-ban would survive court challenge is genuinely untested.

Even Alan Rozenshtein—the University of Minnesota law professor and former Justice Department national-security lawyer who wrote the leading argument that distributing model weights is not constitutionally protected speech—concedes that restricting people’s ability to use a model “may violate the First Amendment rights of users.”³⁰ No court has ruled on model weights. Ever.

Every switch a government has thrown has reached a connection—an API, a download page, a contract. None has reached the copy on your own hardware.

Two provisos are important to keep in mind:.

1. Beijing can close the tap. China’s commerce ministry has been consulting its own labs on restricting foreign downloads of future Chinese model weights while keeping hosted access open—a consultation, with nothing issued.³¹

On the day K3’s weights shipped, a state-media commentary added that support for openness “doesn’t mean advocating for the unconditional proliferation of all capabilities.”³² If that tap closes, here is what changes for you: the weights on your servers keep working, untouched. But the next version never arrives as a download—your model ages in place while the frontier moves, and updates stop. The remaining route to new Chinese capability becomes the hosted API, which walks you straight back through the hosted-API door and its jurisdiction. And the free-download alternative your procurement team waves in closed-vendor negotiations quietly disappears.

2. Your own regulator still reaches you: privacy authorities in Europe and Singapore can order what you may do with personal data on any model you run, self-hosted or not.^{10 13} The real open-weights risk, in one line: stagnation and lost support—not deletion.

Now the clause that quietly favors the models nobody can patch. On June 5—months after the Department of War contract-termination letters —a national-security memorandum called NSPM-11 directed America’s defense and intelligence agencies to ensure, by contract, that “no commercial entity or adversary possesses the capability to prevent use of, disable or degrade, or materially modify” an AI system the military depends on.³³

Look at the drafting: your vendor and a hostile state, in the same sentence, subject to the same prohibition. A closed-model vendor that keeps guardrails it can update and access it can revoke is an entity with exactly that capability—it cannot comply with the clause without surrendering its product’s design.

A lab that publishes its weights retains no such capability, by construction. It complies with the clause the day the weights are published.

The memorandum never uses the words “open weights”; it doesn’t need to—its adaptation section says agencies shall adopt commercial “or open-source” AI, and an open-weights lab already sits on the Pentagon’s classified-networks vendor list.³⁴ Put this section beside the last one and the collision is visible: the contract language now rewards precisely the model class whose trained-in behavior no one—not the vendor, not the government, not you—can patch. For transparency, this is my analysis; there is no publically available or official precedent for this reading..

Under Washington’s new clause, the labs able to secure national-security contracts are the ones that relinquish control of their models.

Who does this federal contractor clause affect?

NSPM-11 governs the national-security enterprise—defense and intelligence contracts—and it reaches both prime contractors and their subcontractors, with termination as the enforcement.³³ If your federal work is civilian—GSA schedules, health, education—this memorandum does not currently apply; the bills that would go further exist, and none have passed.⁵ If you do sit anywhere in a defense or intelligence contract chain, directly or as a subcontractor, the question is live now: which of your AI vendors could comply with a no-disable clause—and would the company whose model you use agree to surrender control to allow you to continue with the federal contract?

If your work falls under NSPM-11 purview, this is a crucial answer to ascertain with certainty.³³?

The Four Answers on One Your Desk

When the next model decision reaches your desk, run the four questions I’ve covered in order:

1. which door does our data go through

2. whose law stands behind that door

3. what can we change after deployment

4. who can switch this model off

Each of those four questions now has an answer you can act on—that is what the past seven weeks of letters, orders, and releases settled. Geopolitics is now part of AI procurement. Not as a choice—as a variable, sitting inside the same spreadsheet as price and capability, whether or not anyone on your team was consulted.

What to watch in the coming months

1. The courtroom. A federal judge in Washington set the first hearing on the Fable 5 shutdown challenge for no earlier than this week—the week you are reading this. What has been filed since late June is not yet publicly visible; no ruling had been reported as of this writing.³⁵

2. Brussels. August 2, when the Commission’s fining power over frontier-model makers goes live—six days after K3’s weights did.³ What how the EU applies their new powers. Learn how they may affect you if you do business in the EU or with EU citizens’ data.

3. The migration. Whether enterprises keep moving production work onto open weights at the pace one large AI gateway’s telemetry shows—29% of tokens on under 4% of spend, a third of the work at a fraction of the cost, as of mid-July.³⁶

Jurisdiction is chosen at the architecture review, not discovered at the subpoena.

The AI Leadership Playbook

Strategic Questions (copy-paste ready for an email to your CIO and General Counsel):

1. For every AI provider we currently use: where are its servers located, which country’s government can compel that provider to hand over our data, and does its service agreement let it reroute our traffic to servers in another country—with what notice? Who in the gets those answers in writing?

2. For any model we run on our own servers: what is our testing budget to discover behavior trained into the model that no vendor can change—and who owns the fallback if we find behavior we cannot accept?

3. Before our next model contract renewal: for each vendor on the shortlist, which government could legally order it to hand over our data—because of where its servers sit, because of who controls the data, or because of whose citizens’ data it holds? Which choice changes our exposure?

Your Next Plays:

1. Build the inventory that answers Question 1 permanently. One list: every AI provider and model currently in use, classified by door—hosted API or weights on our own servers—with the compellable government named beside each deployment.

2. Stand up the evaluation line for anything self-hosted. Checksums against the publisher’s official repository, behavioral testing before production, and a named owner for escalation when a test fails.

3. Brief counsel with the three-legal-bases summary—server location, control of the data, citizenship of the data subjects—and put one question on their desk: are our European operations governed by a single government or nearly thirty?

📅 Book a complimentary 1:1 Strategy Session—45 minutes to start that conversation about your AI transformation sequence.

📬 Free preview ending soon. Subscribe to continue getting decision-grade AI intelligence that prepares you to move before your competitors do. First 100 subscribers receive bonus content. Subscribe to The AI Playbook

The Legal Resource Map

By region: what each resource is, why it matters, when to use it, and the question it answers—so you know which link to open for which job.

United States

1. The CLOUD Act, 18 U.S.C. §2713—law.cornell.edu. What it is: the two-paragraph statute behind the control principle. Why it matters: it is the single strongest data-reach law in this map. When to use it: when counsel or a vendor claims overseas storage insulates your data. The question it answers: can a US provider be ordered to hand over data it stores abroad? (Yes—read the words “within or outside.”)

2. NSPM-11, the national-security AI memorandum—whitehouse.gov. What it is: the primary text of the no-disable clause, four pages. Why it matters: contract language propagates through subcontract chains faster than regulation. When to use it: before signing or renewing anything in a defense or intelligence chain. The question it answers: what exactly must my vendor—or my product—be unable to do?

3. NTIA (the US Commerce Department’s information-policy agency), Dual-Use Foundation Models with Widely Available Model Weights—ntia.gov. What it is: the US government’s own definitional study of open weights. Why it matters: it is the reference point both sides of the ban debate cite. When to use it: when your board asks what “open weights” formally means and what Washington concluded (monitor and keep options open—neither restrict nor bless). The question it answers: what does the US government say open weights are—and what risks do they pose?

European Union

1. The EU AI Act, Regulation 2024/1689—eur-lex.europa.eu. What it is: the full text of record. Why it matters: its general-purpose-model obligations bind providers whose models reach the EU market regardless of origin. When to use it: Articles 51–55 and 93, when assessing any model deployed for EU users. The question it answers: what can Brussels demand of a model’s maker, and with what penalty?

2. European Commission guidelines for general-purpose AI (GPAI) providers—digital-strategy.ec.europa.eu. What it is: the Commission’s own plain-language timeline of which obligations applied when—including the August 2, 2026 enforcement date. Why it matters: the dates are the compliance calendar. When to use it: when planning any EU-touching model deployment this year. The question it answers: what is enforceable after August 2, 2026?

3. The Garante’s DeepSeek decision, January 30, 2025—garanteprivacy.it. What it is: the actual order most coverage paraphrased wrongly (English available). Why it matters: it is the template for how a single EU national authority moves against a foreign model provider. When to use it: when evaluating any provider with no EU entity. The question it answers: what can one member state do alone, how fast, and what does the order actually reach?

Singapore + APAC

1. Personal Data Protection Act 2012—sso.agc.gov.sg. What it is: Singapore’s privacy statute, consolidated. Why it matters: it is the region’s reference model—investigation-tied state access, government exemption, directions backed by penalties up to 10% of local turnover. When to use it: when standing up Singapore-touching deployments, including through Moonshot’s Singapore-domiciled API. The question it answers: what governs personal data in Singapore, and who is exempt?

2. Online Criminal Harms Act 2023—sso.agc.gov.sg. What it is: Singapore’s blocking machinery—directions to ISPs, app stores, and payment support, no court order required. Why it matters: it defines what Singapore could do to any online service, AI included, the day it chooses to. When to use it: to understand the region’s enforcement ceiling. The question it answers: if Singapore ever moved against a model service, what would that look like?

3. MAS (Singapore’s central bank and financial regulator), Artificial Intelligence Model Risk Management—mas.gov.sg. What it is: the financial regulator’s expectations for AI inside banks—the strictest sector lens in APAC. Why it matters: financial-sector rules preview where general rules go. When to use it: if you operate regulated entities in Singapore, or want the region’s most concrete model-governance checklist. The question it answers: what does a serious APAC regulator already require of models in production?

People’s Republic of China (K3 provider’s home law)

1. National Intelligence Law of the PRC (English translation)—chinalawtranslate.com. What it is: the support-assist-cooperate statute, translated. Why it matters: it is the sentence most coverage compresses into “Chinese law reaches everything.” When to use it: before repeating that sentence in any board discussion. The question it answers: what does the intelligence-cooperation duty actually say—and not say?

2. Personal Information Protection Law of the PRC (English translation)—chinalawtranslate.com. What it is: China’s privacy statute, including its blocking provision on foreign-authority disclosure. Why it matters: it is the law that points the opposite direction from the popular assumption. When to use it: when mapping what a PRC provider may lawfully hand to whom. The question it answers: what does Chinese law forbid its companies from giving foreign authorities?

3. Interim Measures for Generative AI Services (English translation)—chinalawtranslate.com. What it is: the rules governing generative-AI services offered inside China, including suspension powers. Why it matters: it defines the enforcement ladder Beijing holds over its own labs’ hosted services. When to use it: when assessing the provider-side pressure on any PRC lab. The question it answers: what can Beijing order its own AI companies to do?

A note on these translations: I am unable to independently verify translation quality from the original Chinese. All three are sourced from China Law Translate, the translation project run by Jeremy Daum at Yale Law School’s Paul Tsai China Center—a reputable institution standing as the proxy for accuracy.

Sources

1. Hugging Face, moonshotai/Kimi-K3 model repository—weights, model card, and license, read directly; 96 shards, ~1.6 TB. huggingface.co/moonshotai/Kimi-K3

2. Global Times, “China slams US’ planned probe, sanctions against Chinese AI firms, will take ‘all necessary measures,’” July 27, 2026—globaltimes.cn; official translation of the ministry statement: CSET, Georgetown.

3. European Commission, “Guidelines for providers of general-purpose AI models”—enforcement powers, including fines, from August 2, 2026. digital-strategy.ec.europa.eu

4. NVIDIA et al., “Open Weights and American AI Leadership,” open letter, July 24, 2026; signatory list as fetched July 27, 2026. images.nvidia.com

5. Axios (Maria Curi), “The secret Trump administration battle to fight Chinese AI,” July 20, 2026—Entity List deliberations; draft hosting-liability order; nothing enacted. axios.com

6. Paola Sanmiguel, “It’s Not Thinking. It’s Predicting.,” The Weekly Call, June 23, 2026. cognivalab.blog

7. Moonshot AI Pte. Ltd., Kimi international privacy policy—servers located in Singapore. platform.kimi.ai

8. 18 U.S.C. §2713 (CLOUD Act)—”within or outside of the United States.” law.cornell.edu

9. U.S. Department of Justice, “Promoting Public Safety, Privacy, and the Rule of Law Around the World: The Purpose and Impact of the CLOUD Act,” white paper, April 2019—the Microsoft Ireland case background. justice.gov

10. Regulation (EU) 2016/679 (GDPR), Articles 3, 44–49, 58. eur-lex.europa.eu

11. Personal Information Protection Law of the PRC, Articles 3, 41 (China Law Translate). chinalawtranslate.com

12. 50 U.S.C. §1881a (FISA Section 702). law.cornell.edu

13. Personal Data Protection Act 2012 (Singapore), ss. 4, 48I–48J. sso.agc.gov.sg

14. National Intelligence Law of the PRC, Articles 7, 14 (China Law Translate). chinalawtranslate.com

15. Donald Clarke, “The Zhong Lun Declaration on the Obligations of Huawei and Other Chinese Companies under Chinese Law,” SSRN, March 2019. papers.ssrn.com

16. Jeremy Daum, “What the National Intelligence Law Says, and Why It Doesn’t Matter,” China Law Translate. chinalawtranslate.com

17. Samm Sacks and Peter Swire, “Assessing U.S. Data Policy Toward China: A Proposed Framework,” Lawfare, July 14, 2023. lawfaremedia.org

18. PRC Ministry of Foreign Affairs spokesperson Guo Jiakun, January 17, 2025, via Global Times. globaltimes.cn

19. Garante per la Protezione dei Dati Personali (Italy), Provvedimento of January 30, 2025 (DeepSeek), doc-web 10098477. garanteprivacy.it

20. Open Source Initiative, “The Open Source AI Definition 1.0.” opensource.org

21. CrowdStrike (Stefan Stein), “Security Flaws in DeepSeek-Generated Code Linked to Political Triggers,” November 20, 2025. crowdstrike.com

22. Semgrep, “Kimi K3’s code security results lack precision,” July 22, 2026. semgrep.dev

23. UK AI Security Institute and US CAISI, “Preliminary assessment of Kimi K3’s cyber capabilities,” July 23, 2026. aisi.gov.uk

24. Anthropic, “Statement on the US government directive to suspend access to Fable 5 and Mythos 5,” June 12, 2026—as I documented in the July 9 Call, cited here for the 5:21 p.m. timeline. anthropic.com

25. Paola Sanmiguel, “Not a Screwdriver or Uranium: Washington Invented a Third Label,” The Weekly Call, July 9, 2026. cognivalab.blog

26. DiDi Global, “DiDi Announces App Takedown in China,” Business Wire, July 4, 2021—installed apps “may continue using it.” businesswire.com

27. Interim Measures for the Management of Generative AI Services (PRC), Article 21 (China Law Translate). chinalawtranslate.com

28. Online Criminal Harms Act 2023 (Singapore), ss. 8–12, 29–31. sso.agc.gov.sg

29. Cornell Legal Information Institute, “TikTok, Inc. v. Garland” (U.S. Supreme Court, decided January 17, 2025). law.cornell.edu

30. Alan Z. Rozenshtein, “There Is No General First Amendment Right to Distribute Machine-Learning Model Weights,” Lawfare, April 4, 2024. lawfaremedia.org

31. Reuters, “China considers tighter export controls on AI models, chips, FT reports,” July 21, 2026 (wire story; carrier verified live). finance.yahoo.com

32. Bloomberg, “China State Media Says Support for Open AI Models Has Limits,” July 27, 2026. bloomberg.com

33. National Security Presidential Memorandum 11, “Artificial Intelligence in the National Security Enterprise,” June 5, 2026. whitehouse.gov

34. U.S. War Department, “Classified Networks AI Agreements,” May 1, 2026—eight companies including the open-weights lab Reflection. war.gov

35. Legion LegalTech, Corp. v. United States, No. 1:26-cv-02225 (D.D.C.), docket—minute order of June 25, 2026. courtlistener.com

36. Vercel, “AI Gateway Production Index,” July 2026—open-weight models at 29% of tokens on under 4% of spend. vercel.com

© 2026 Paola Sanmiguel. All rights reserved.

A note on AI use: Anthropic, the maker of Claude, is a party to several precedents discussed—load-bearing claims about those disputes are cited to court records and, wherever possible, non-Anthropic sources. This piece was researched using Anthropic’s Claude Fable 5 and Opus 5 models. All sources are verified and all cited material is corroborated by multiple independent sources.

Those reps are disappearing. AI now writes the first draft, ties the reconciliation, and assembles the deck—absorbed precisely because the work was routine. What nobody priced into the automation case: the routine work was the apprenticeship.

Most executives still treat entry-level hiring as the safest line to trim when AI absorbs the routine work. The research that converged this spring says the opposite: it is the most expensive cut on the books—the bill just arrives years late, on the senior bench. In The Judgement Premium¹⁶ I argued that judgment is the capability your AI investment gets priced against; in The Sophistication Gap¹⁷, that few workforces have built it. This Call is about the supply side: the system that manufactures judgment is collapsing, and what you rebuild now decides who runs your company in 2035.

Why entry roles now demand the judgment they used to teach

PwC has been a throughline in these Calls—its AI Performance Study anchored The Sophistication Gap¹⁷ a month ago. This month a different PwC instrument, the 2026 Global AI Jobs Barometer—more than a billion job ads across 27 countries—supplies the evidence¹: postings for AI-exposed entry-level roles are now seven times more likely to demand senior-level skills. “Seniorised” entry-level postings have grown 35% since 2019 while other entry-level postings shrank 10%¹. PwC’s global workforce leader Pete Brown names the mechanism plainly: “AI is removing some of the routine work that once acted as an apprenticeship, while increasing demand for judgement, leadership and adaptability much earlier in careers.”¹

Put those numbers together: the career ladder still exists, but its bottom rung has been rebuilt at senior height. I call it The Seniorized Rung—the entry level now demands on arrival the judgment it used to exist to build. No one designed this; it emerged, posting by posting, as AI absorbed the routine tasks.

The hiring data shows how quietly it rose. Harvard researchers studying 62 million workers across 285,000 firms found that after companies adopted generative AI, junior employment fell 7.7% within about six quarters relative to non-adopters, while senior employment held steady². The mechanism was a hiring slowdown, not layoffs: the door is not slamming shut, it is quietly not opening in the first place.

The posting mix agrees. Entry-level roles were 44% of U.S. job postings in 2023; by this March, 38.6%³—down 5.4 points in two years, roughly an eighth of the entry-level share. ZipRecruiter reads its report as a graduate market improving—on cyclical measures it is; the share shift is the structural signal underneath. The entry level, increasingly, has no entry.

The debt that matures when no one is manufacturing seniority

The cost of the missing rung does not land in this year’s hiring plan; it lands on a delay, which is why it goes unpriced. I call the liability Apprenticeship Debt: the cumulative gap between the junior intake your future senior bench requires and the intake you are actually making. It compounds every quarter you under-hire and matures, by our estimate, eight to fifteen years out—the balance-sheet mirror of the Human Dividend, the liability you book when you stop investing in human capability.

Jeff Raikes—who ran Microsoft’s Business Division and later the Gates Foundation—named the cost in April: companies cutting entry-level roles are doing it “before they have a talent debt coming due,” and that “the days are numbered for any company that doesn’t develop a human talent pipeline with the judgment to direct it.”⁴ Pricing that debt takes three numbers your CHRO can pull this week:

1. Management is 7.2% of the U.S. workforce⁵—the senior layer every AI strategy assumes will be there to direct it.

2. That layer renews almost entirely from below. Promotions into management run at roughly 6.5% a year, back at their pre-pandemic pace, per ADP payroll data⁶.

3. The bench is already thin. DDI finds only 49% of key roles can be filled from the internal bench, and just 20 % of HR leaders say they have ready successors for critical positions⁷.

Hold those three numbers together. Cutting the entry cohort does not shrink a cost line; it shrinks the only pool your promotions draw from.

The market is already pricing the scarcity. The same Barometer finds AI-skilled workers command a 62% wage premium, up from 57% a year earlier¹. Stanford’s payroll data shows employment for 22-to-25-year-olds in the most AI-exposed occupations down 16% in relative terms since late 2022⁸. Brookings and Yale’s Budget Lab caution—fairly—that the aggregate data shows no AI jobs apocalypse yet^{9 10}. The risk stands anyway: the damage is to a stock that takes a decade to rebuild, and by the time the signal clears, the missing cohort is unrecoverable.

The first invoice already carries a date: Gartner predicts that supply-chain organizations pausing entry-level hiring for AI will face higher costs by 2030¹¹. The deeper invoice is our own projection: junior intake began thinning around 2023, and the climb from entry role to seasoned senior runs eight to twelve years—placing the first missing cohort in the early 2030s. The Playbook gives you the calculation to run on your own attrition data.

What copilots compress, and what judgment still requires

The strongest objection first: AI will train the next cohort faster than it displaced the old one—simulators, copilots, instant feedback on every draft. Ethan Mollick, who taught a generation of executives co-intelligence, now describes an era of “co-existence” with agents that work increasingly on their own¹². If the machine can coach every junior individually, the apprenticeship did not die—it got an upgrade.

Here is what it compresses away. AI shortens task-execution time. Judgment forms through consequential reps—decisions with real stakes, owned outcomes, and feedback that arrives with your name attached. The World Economic Forum calls these “judgement loops,” and warns that stripping them out of entry-level work is quietly building a five-year leadership-pipeline crisis¹³. A copilot can grade a junior’s draft in seconds; it cannot make her defend a recommendation to an impatient client, or be wrong in a way she has to repair. And the more autonomous the agents, the higher the judgment bar rises for the human who directs them. The tool that looks like the trainer is the thing that removed the training.

The rebuild: put your juniors on the review side of the AI

The rebuild starts from one design principle: entry-level roles are no longer cheap production labor to automate—they are judgment-formation roles. Put the junior on the review-and-direction side of the AI, not the produced-output side. The first-year hire who verifies, challenges, and directs AI output accumulates exactly the reps the automation removed: evaluating work she did not produce, owning the call on whether it ships, and carrying the consequences.

One firm has already engineered for this. Law firm Fredrikson & Byron is rebuilding associate development around AI supervision: Chief Legal Operations Officer Norah Olson Bluvshtein described three tactics to Thomson Reuters, including a “formalized curriculum around effectively and efficiently supervising AI output”—teaching young lawyers to direct the machine rather than compete with it¹⁴. The honest caveat: it is a stated goal for the year ahead, a designed rebuild rather than a finished one. The blueprint is the point—the firm treated the collapsing on-ramp as an engineering problem while its peers cut the rung and hope seniority appears by other means.

The early returns favor the builders. The same Barometer separates firms that use AI to amplify their people from firms that use it primarily to automate them: the amplifiers show 163% higher productivity growth and grew headcount 52%¹. That is the compounding return: developmental investment on people, paying back through the mechanism The Judgement Premium¹⁶ priced—judgment that turns tools into results.

When the entry-level hiring plan reaches your desk, the number in the deck will be cohort cost. The number that belongs on the table is supply: whether the company is still manufacturing the judgment the strategy depends on. SHRM puts the same point in CHRO language—the real multiplier on AI at work is human leadership and judgment¹⁵. This decision gets made this quarter, or it gets made for you, by attrition.

Somebody once handed you the deck nobody wanted, and it made you. Handing it back—redesigned—is how the next generation gets made. How to measure that rebuild is territory for a coming Call.

The AI Leadership Playbook

Strategic Questions (copy-paste ready for an email to your CFO and CHRO)

1. Since 2023, how many of our entry-level roles were cut or left unfilled because AI absorbed the routine work—and how much smaller is the pool our senior promotions will draw from in 2030? Who owns building that leadership bench now?

2. Our AI tools now produce the first drafts our junior people used to learn on. Where in our company does a first-year employee still make decisions with real stakes—and if the answer is nowhere, which role do we redesign first?

3. AI-skilled talent already commands a 62% wage premium. What will it cost us to buy senior judgment on the open market in 2030 versus manufacturing it internally—and which function do we model first?

Your Next Plays (copy-paste ready for an email to a direct report)

1. Price our Apprenticeship Debt. Apprenticeship Debt is the gap between the junior intake our future senior bench requires and the junior intake we are actually making. Pull our senior headcount and annual senior exit rate; multiply for replacement demand. Divide by our historical junior-to-senior conversion rate for required intake. Compare with actual intake since 2023, and date the year the gap reaches the senior bench. Bring the number, not a narrative.

2. Move one entry role to the review side of the AI. Take one function’s entry role and rewrite it so the junior verifies, challenges, and directs AI output—with named decision rights and outcomes they own—instead of producing the drafts the AI now writes. The goal is judgment formation: reps with real stakes, not shadowing.

3. Audit where judgment actually gets practiced. Inventory the tasks that used to build judgment in our function—first drafts, reconciliations, client debriefs—and mark which ones AI has absorbed. For each absorbed task, name what replaced it as a practice rep. Where nothing replaced it, that is a gap; assign each gap an owner.

📅 Book a complimentary 1:1 Strategy Session—45 minutes to start that conversation about your AI transformation sequence.

📬 Free preview ending soon. Subscribe to continue getting decision-grade AI intelligence that prepares you to move before your competitors do. First 100 subscribers receive bonus content. Subscribe to The AI Playbook.

Sources

1. PwC, 2026 Global AI Jobs Barometer (Jun 2026)—1B+ job ads, 27 countries; AI-exposed entry roles 7× more likely to require senior skills; “seniorised” entry postings +35% since 2019 vs −10%; 62% AI-skills wage premium (up from 57%); amplify-vs-automate firms +163% productivity, +52% headcount; Pete Brown quote.

2. Hosseini Maasoum & Lichtinger (Harvard), “Generative AI as Seniority-Biased Technological Change” (SSRN, 2025)—62M workers / 285k firms; junior employment −7.7% at GenAI-adopting firms ~6 quarters post-adoption; senior employment unchanged; hiring-slowdown-driven.

3. ZipRecruiter Economic Research, Annual Grad Report (2026)—entry-level share of postings 38.6% (Mar 2026), down from 44% (2023).

4. Fortune (Apr 15, 2026)—Jeff Raikes on “talent debt coming due” and the human talent pipeline with the judgment to direct AI.

5. U.S. Bureau of Labor Statistics, Occupational Employment and Wage Statistics (May 2025)—management occupations ≈11.1M of 155.5M jobs (~7.2%).

6. ADP Research (2025)—managerial promotion rates cooled to their pre-pandemic level (~6.5%/yr), payroll data.

7. DDI, Global Leadership Forecast 2025—10,796 leaders / 2,014 orgs; only 49% of key roles fillable internally; 20% of HR leaders report ready successors.

8. Stanford Digital Economy Lab—Brynjolfsson, Chandar & Chen, “Canaries in the Coal Mine?” (Nov 2025, v3)—ADP payroll data; 16% relative (6% absolute) employment decline, ages 22–25, most AI-exposed occupations, since late 2022.

9. Brookings, Molly Kinder (2025)—”New data show no AI jobs apocalypse—for now.”

10. The Budget Lab at Yale (Dec 2025)—CPS-based evaluation; labor market broadly stable since ChatGPT’s release.

11. Gartner press release (May 5, 2026)—Gartner predicts supply-chain organizations pausing entry-level hiring for AI will face higher costs by 2030.

12. Ethan Mollick, “Co-Existence and the End of Co-Intelligence” (One Useful Thing, Jun 4, 2026).

13. World Economic Forum (Jun 2026)—AI, the leadership crisis, and Gen Z; “judgement loops” and the five-year leadership-pipeline risk.

14. Thomson Reuters Institute—Natalie Runyon (Apr 16, 2026)—Fredrikson & Byron associate-development rebuild; CLOO Norah Olson Bluvshtein; “formalized curriculum around effectively and efficiently supervising AI output” as a stated goal for the year ahead.

15. SHRM—”AI’s Real Multiplier at Work: Human Leadership and Judgment.”

16. CognivaLab, The Judgement Premium (Jun 3, 2026)—last month’s pricing of judgment as the capability AI investment depends on.

17. CognivaLab, The Sophistication Gap (Jun 9, 2026)—adoption vs. sophistication; the 75-point gap.

EDITORIAL NOTE: The scheduled article for this week was pre-empted by this follow-up piece on the return of Fable 5 on 1 July 2026. Previous analysis of the shut down published on 15 June 2026 is linked below. Also, I’m trying out a different day for these long form research articles. Tell me in the notes if you prefer Tuesdays or Thursdays. Back to our regularly scheduled AI industry analysis next week!

Three weeks ago, the US government switched off the most capable AI model on the market with a label and a letter. I walked you through that in Screwdriver or Uranium: Why What We Call an AI Model Now Decides Who Can Use It. I made fifteen predictions about what the switch would set in motion. And I promised to keep score in public.

On July 1, Fable 5 came back on—for everyone, everywhere, at once (Anthropic). This edition is the update I promised: how the model came back, the scorecard on those fifteen calls, and the answers to the open questions the first article left unresolved. The headline is not that the model returned. It is what the return revealed.

How the models came back

Let’s start with the timeline, because the mechanism lives in it:

• June 12, 5:21pm ET. Commerce Secretary Howard Lutnick’s export-control letter lands. It orders Anthropic to block every foreign national worldwide, including the company’s own employees inside the US. Anthropic cannot sort users by citizenship in real time, so it shuts off Fable 5 and Mythos 5 for everyone (Anthropic).

• June 26. A second Lutnick letter carves out “certain trusted partners” and their foreign-national employees for Mythos 5. Anthropic confirms the government approved Mythos 5 for a set of US organizations that operate and defend critical infrastructure (Mayer Brown).

• June 30. The export controls on both models are lifted (Anthropic).

• July 1. Fable 5 is restored globally on every Anthropic surface, with cloud platforms to follow “as quickly as possible” (Anthropic).

Counting from the June 12 order, that is 18 days to the lift and 19 to restored access. You will see both numbers in coverage; the difference is only where you stop the clock.

What happened between those dates was not litigation or legislation. It was a negotiation. Anthropic’s own account itemizes what it conceded (Anthropic):

1. A new safety classifier. It blocks the specific reported technique in over 99% of cases. Blocked requests fall back to Claude Opus 4.8.

2. Government validation. CAISI—the Commerce Department’s Center for AI Standards and Innovation, housed at NIST—tested the old and new safeguards. Its researchers judged them “extraordinarily strong.”

3. Pre-release government access. For future models that push the frontier in national-security-relevant areas, designated government partners get early access. They can test the models and the guardrails before broad release.

4. Rapid information sharing. Significant jailbreaks and misuse patterns get investigated and reported to government counterparts. Anthropic also joins the interagency cyber-vulnerability clearinghouse created under the June 2 Executive Order.

5. Dedicated joint resources. Anthropic teams and a significant compute allocation now support government testing and research.

6. An industry jailbreak-severity framework. Co-developed with Amazon, Microsoft, Google, and other Glasswing partners. It scores any jailbreak on four criteria: capability gain, breadth, ease of weaponization, and discoverability. A new HackerOne submission program and a 24/7 monitoring team back it up.

The scorecard: what moved since June 15 and the consequences

I made fifteen predictions in the previous article. Here are the five that materially moved; the full list is in the June 15 edition.

1. “Your most powerful tool can be turned off overnight” (#1)—proven, then reversed the same way. The model was turned off by an unpublished letter and reinstated by an unpublished letter. Neither document has been released (Just Security).

2. “Will the model stay on?” (#6)—now a permanent planning criterion. The model that returned is not the model that left. Fable 5 is included on Pro, Max, Team, and select Enterprise plans only up to 50% of weekly usage limits through July 7, then moves to usage credits (Anthropic).

3. Investors price the switch-off risk (#8–9)—confirmed in kind, not in numbers. Anthropic and OpenAI are both still privately held, so there is no stock price to serve as a verifiable indicator of this episode’s cost. What needs no market data: a 19-day, government-ordered outage of a flagship product is no longer a hypothetical anyone can wave away.

4. A defender’s tool pulled because it might help attackers (#13)—confirmed, and answered. The resolution restored the cyberdefense model first, to critical-infrastructure defenders, one week before the general lift (Mayer Brown).

5. Unsettled law (#15)—now a docket number. A customer sued the government on June 23 (CourtListener). More below.

The other ten calls are trending the way I sketched them in the previous article, but none has resolved cleanly enough to score definitively.

The jailbreak question, answered in practice

The previous article’s sharpest dispute was whether the reported bypass was even a real jailbreak. The resolution answered it in practice without ever settling it in principle.

Anthropic ran the reported technique through the wider field—Claude Opus 4.8, GPT-5.5, Kimi K2.7, and five older systems (Anthropic). All of them matched the result. The technique exposed no capability unique to the Mythos class. By the company’s account, the behavior was a borderline case inside a deliberately oversized safety margin: routine defensive cybersecurity work, blocked for caution, unlocked by clever prompting.

And yet Anthropic built the fix anyway, and the government’s own testers validated it. That sequence also settles a June claim from a White House adviser that Anthropic had refused to fix the problem. The fix shipped, and CAISI signed off on it (Anthropic). One transparency note: the account of the new classifier and its validation comes from Anthropic’s own statement—the government has not published its own version of the testing (Just Security).

Two corrections of my own, because the scorecard cuts both ways. In June I described the UK AI Security Institute as having built a working jailbreak within hours. That was too strong. The institute’s published work documents capability evaluations of the Mythos-class model (UK AISI). The fact on the record is Anthropic’s own launch concession that UK AISI “has made progress towards” a universal jailbreak (Anthropic). And the June report of suspected foreign access to the model stays where I left it: unconfirmed, and out of the load-bearing story.

Here is what should hold your attention: nobody conceded a definition. The industry responded by starting to write one. The four-criteria severity framework is the field’s first attempt at an objective standard for when a jailbreak matters (Anthropic).

The machinery now has a name

The previous article’s biggest unknown was that nobody had publicly named the legal authority behind the order. That is now half-answered, and the half matters.

The legal mechanisms used are now public. The letter invoked Section 4817(b)(1) of the Export Control Reform Act, which lets Commerce place interim controls on emerging technologies. It also invoked Section 744.22(b) of the Export Administration Regulations—the “is-informed” rule aimed at military-intelligence diversion risk in countries of concern, including China and Russia (Mayer Brown). The enforcement question was also resolved: the authority sits with the Commerce Department’s Bureau of Industry and Security (BIS), under Lutnick—not the State Department, and not the Pentagon, whose separate supply-chain fight with Anthropic is still live (Just Security).

What remains unpublished is everything else. The order itself has never been disclosed. No rule was issued. No classification number exists. Commerce has released no public guidance on the threat or the standard (Just Security).

And inside those unpublished letters sit two legal firsts (Mayer Brown):

1. The model itself became a controlled item. Commerce claimed jurisdiction over the model as such—every earlier action stopped at weights or source code.

2. Access became an export. It counted API access as an export. Under the “deemed export” doctrine, that reach extends to foreign nationals sitting inside the US.

The second move contradicts three standing advisory opinions from BIS—the same Commerce bureau now enforcing this order—issued in 2009, 2011, and 2014. Those opinions held that remote access to cloud software is not an export, and companies have relied on them for years (Mayer Brown). Congress is still debating the Remote Access Security Act, a bill that would grant BIS exactly this authority. The bill’s existence implies BIS may not have that authority today.

Step back and the shape comes into focus. The question in June was screwdriver or uranium: ordinary dual-use software, or weapon. Washington answered with neither. It created a third category that exists only in company-specific letters—unpublished, case-by-case, resting on contested authority, with no rule behind it and no court yet ruling on it. I call this: rule by letter. The most consequential AI policy of 2026 is, so far, a stack of private correspondence.

The courts get their turn

In June, no one had sued. That lasted eleven days.

On June 23, a San Jose legal-tech company called Legion LegalTech filed suit. Its developers in Canada had lost Fable 5 access. The case—Legion LegalTech, Corp. v. United States of America, No. 1:26-cv-02225, in federal court in Washington, DC—names Secretary Lutnick and BIS Under Secretary Jeffrey Kessler among the defendants (CourtListener). The company wants the directive vacated and blocked. Its core argument: the order exceeds the government’s statutory authority (Mayer Brown). The court has set a schedule: the government responds by July 14, and a hearing comes no earlier than the week of July 27 (CourtListener).

In June I told you about the 1990s encryption fight, when courts pushed back on treating code as a munition. That history is no longer an analogy. It is a live case—filed not by the company that was silenced, but by a customer who lost the tool.

The suit’s terrain is a sentence the government wrote itself. Executive Order 14409 is the policy backbone of this whole regime, signed June 2 and published June 5. It states that nothing in its frontier-model section “shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement” for releasing new AI models (Federal Register). Ten days after those words were signed, a mandatory license requirement arrived at Anthropic’s door by letter. The distance between that sentence and that letter is where this case will be fought.

What Washington wanted, what the labs gave, what they kept

Here is the full map: who demanded what, who conceded what, and which objections survive.

1. What the government stated it wants. Screening of frontier models for cyber-misuse risk before they reach the world. The letter’s own legal hooks frame the concern as military-intelligence diversion to countries of concern (Mayer Brown). The June 2 Executive Order builds the standing machinery: classified capability benchmarking, plus a voluntary framework that gives the government up to 30 days of pre-release access to “covered frontier models” (Federal Register).

2. What Anthropic gave. The classifier fix, CAISI validation, pre-release access, information sharing, joint teams and compute, the co-led severity framework (Anthropic)—on top of the 30-day data-retention policy it has run on Mythos-class traffic since launch (Anthropic).

3. What OpenAI gave, before it was ever ordered to. The regime generalized to a second lab before it finished resolving at the first. On June 26—four days before Anthropic’s lift—OpenAI previewed GPT-5.6 Sol, Terra, and Luna. At the government’s request, it released them first to “a small group of trusted partners whose participation has been shared with the government.” OpenAI calls this a “short-term step” while a “repeatable process for future model releases” gets built (OpenAI).

4. What the labs kept, verbatim. Anthropic: these rules should be “codified in strong regulation and applied equally across frontier model developers” (Anthropic). OpenAI: “We don’t believe this kind of government access process should become the long-term default” (OpenAI). Both companies complied; neither accepted the process as legitimate long-term law.

5. The actual state as of publication (July 6, 7am ET). The models are live. The letters remain unpublished. The framework is voluntary and in draft. The lawsuit is pending, with a hearing no earlier than the week of July 27 (CourtListener).

A truce was reached on process; the legal question underneath—can Commerce do this at all?—was conceded by no one and is now in court.

From the boardroom to the solo desk

In the previous article, the sharpest question at the solo desk was what a government shutdown does to the work that depends on the model. The access came back. The lesson got harder.

Fable 5 returned under new commercial terms, set while customers were locked out. Included plan access runs only through July 7, and only up to half of weekly usage limits; after that it shifts to paid usage credits, and cloud availability trails the direct platforms (Anthropic). Three weeks ago the lesson was that you rent the model you build on. Now the lesson is sharper: the rent can change while you are locked out.

If you lead a team, this forces the durability posture I laid out in Durability Is the New AI Advantage and, before that, in The AI Speed Trap. It now has three concrete line items:

1. A tested second-model fallback. Not a name on a slide—a workload you have actually run on the backup, because the 19-day outage was survivable exactly in proportion to how real your fallback was.

2. Session and work portability. Keep state, artifacts, and context exportable so your work outlives any one model’s availability—an outage you can walk away from is an inconvenience, and one you cannot walk away from is a stoppage.

3. Contract language for suspension and terms changes. Your agreements were written for outages measured in hours and prices that change with notice. Ask your vendor what happens to your terms during a government-ordered suspension, because now there is precedent.

I am keeping this cost qualitative on purpose: no verified figure exists yet for what those 19 days cost the ecosystem. The verified record—19 days dark, terms reset, a second lab gated—is heavy enough.

What every release looks like now

Put the public record together and the forward picture is already visible. The Executive Order supplies classified benchmarking and a voluntary pre-release access framework (Federal Register). The industry supplies its own severity standard for jailbreaks (Anthropic). And the labs supply the posture: OpenAI is explicitly building a “repeatable process for future model releases” with the administration (OpenAI). Gated rollouts to vetted partners, with government visibility, are becoming the default shape of a frontier launch.

The fair pushback: one company, one scare, and the system arguably worked—flagged, fixed, validated, restored, in under three weeks. Why call that a regime? Because the pattern outran the incident. OpenAI gated its own launch at the government’s request before Anthropic’s case was even resolved. The Executive Order institutionalizes the review machinery on a standing basis. And the only thing that says this cannot happen again tomorrow, to any lab, on the same unpublished basis, is that nobody has written the rule that says when it can. A regime without a rulebook is still a regime—it is just one without transparency to the companies it regulates and the users they serve globally.

So the question every AI builder inherits is not whether frontier releases get reviewed. That is settled in practice. The open question is which of three forces writes the durable rule:

1. Congress, through something like the Remote Access Security Act (Mayer Brown).

2. The Legion court case, with a hearing no earlier than the week of July 27 (CourtListener).

3. The quiet accumulation of truces, one negotiated resolution at a time.

You are already living under whichever one wins.

The bottom line

Three weeks ago I wrote that the label would be decided after the fact, in a courtroom or a committee room. Both rooms are now booked: the courtroom has a docket number and a July hearing window, and the committee room has a draft framework with four criteria in it. What I could not have told you three weeks ago is that a third room got there first—the one where an unpublished letter switched the model off, and a second unpublished letter switched it back on.

Screwdriver or uranium was the wrong question. Washington’s answer was rule by letter—and every release calendar in the industry has already adjusted to it. Watch three clocks:

• July 14—the government’s response to Legion’s injunction motion.

• The week of July 27—the first hearing on whether a letter can do what a rule never authorized.

• The framework’s first public draft—the industry’s attempt to write the standard before Washington writes it for them.

I will follow up with updates.

A note on sourcing: the June 15 edition was written in the hours after the shutdown and drew in part on early press coverage. Every claim in this update links to the original document or announcement it comes from.

Learn more: What Anthropic’s terms say when a model goes dark

Fable 5 came back with new prices, so I went looking for the contract language that governs the next shutdown. Two findings stand out:

1. The right to suspend was on the books before June 12. Anthropic’s Commercial Terms of Service (effective June 17, 2025) allow suspension when providing the service “is prohibited by applicable law,” with restoration “as soon as reasonably possible” once the cause “is cured, where curable”—and they state Anthropic “will have no liability for any damage, liabilities, losses (including any loss of data or profits)” from such a suspension. The Consumer Terms (effective October 8, 2025) go further: Anthropic may “modify, suspend, or discontinue” the services “at any time without notice.”

2. No clock, no meter, thin recourse. Neither document sets a maximum length for a suspension—days, months, or indefinite—and neither promises compensation when one ends. Consumers can cancel, without a refund for the elapsed term; commercial customers face a liability cap of their last 12 months of fees, and the terms disclaim uninterrupted service outright. The July usage-credit arrangement arrived by announcement, not by contract.

Neither document has been updated since the shutdown, because neither needed to be. The order tested language that was already there—and the language held. Review your own vendor agreements before the next letter is signed.

Additional Resources

• AI Strategic Radar: Decision-grade intelligence for the week ahead—Top 5 stories in AI to prepare you for the work week. Drops every Monday before your first coffee.

• humandividend.ai—where you’ll learn to build the workforce, leadership, and coaching bench that compounds as AI scales.

Sources

1. Anthropic, “Redeploying Fable 5,” June 30, 2026 (updated July 1, 2026). https://www.anthropic.com/news/redeploying-fable-5

2. Anthropic, “Statement on the US government directive to suspend access to Fable 5 and Mythos 5,” June 12, 2026. https://www.anthropic.com/news/fable-mythos-access

3. Anthropic, “Claude Fable 5 and Claude Mythos 5,” June 9, 2026. https://www.anthropic.com/news/claude-fable-5-mythos-5

4. Mayer Brown (Kendler, Hussain, Waltzman, Soliman, Hickey, Jebeyli), “Commerce Department Extends Export Controls to Advanced AI Models; Authorizes Release to Specific Trusted Partners,” June 30, 2026. https://www.mayerbrown.com/en/insights/publications/2026/06/commerce-department-extends-export-controls-to-advanced-ai-models-authorizes-release-to-specific-trusted-partners

5. Brian Egan, “Legal Considerations Related to the Anthropic ‘Export Controls Directive,’” Just Security, June 15, 2026. https://www.justsecurity.org/142745/law-anthropic-export-controls/

6. Executive Order 14409, “Promoting Advanced Artificial Intelligence Innovation and Security,” 91 FR 34565 (signed June 2, 2026; published June 5, 2026). https://www.federalregister.gov/documents/2026/06/05/2026-11415/promoting-advanced-artificial-intelligence-innovation-and-security

7. OpenAI, “Previewing GPT-5.6 Sol: a next-generation model,” June 26, 2026. https://openai.com/index/previewing-gpt-5-6-sol/

8. UK AI Security Institute, “Our evaluation of Claude Mythos Preview’s cyber capabilities,” April 13, 2026. https://www.aisi.gov.uk/blog/our-evaluation-of-claude-mythos-previews-cyber-capabilities

9. Legion LegalTech, Corp. v. United States of America, No. 1:26-cv-02225 (D.D.C., filed June 23, 2026), docket via CourtListener. https://www.courtlistener.com/docket/73520460/legion-legaltech-corp-v-united-states-of-america/

10. Paola Sanmiguel, “Screwdriver or Uranium: Why What We Call an AI Model Now Decides Who Can Use It,” Strategic AI Radar, June 15, 2026. https://www.linkedin.com/pulse/screwdriver-uranium-why-what-we-call-ai-model-now-who-paola-ghdpc

11. Paola Sanmiguel, “Durability Is the New AI Advantage,” Strategic AI Radar, June 22, 2026. https://www.linkedin.com/pulse/durability-new-ai-advantage-paola-sanmiguel-m-s--gbhfc/

12. Paola Sanmiguel, “The AI Speed Trap,” The AI Playbook: The Weekly Call, April 22, 2026. https://cognivalab.blog/p/the-speed-trap?r=272kkc

13. Anthropic, “Commercial Terms of Service,” effective June 17, 2025. https://www.anthropic.com/legal/commercial-terms

14. Anthropic, “Consumer Terms of Service,” effective October 8, 2025. https://www.anthropic.com/legal/consumer-terms

Editorial Note: Folks often ask me to explain how GenAI works or thank me after a talk for an explanation that finally makes sense to them. This probably should have been the inaugural article for The AI Playbook; but there’s no time like the present. Even if you think you know, I invite you to be curious and keep reading. Perhaps this article will refine your thinking when you explore the “Learn More” resources I’ve added. Or maybe it will provide you with useful metaphors you can leverage in conversation with colleagues and friends. Either way, it’s under ten minutes long. Enjoy!

The most useful thing you can understand about today’s AI is also the most easily missed: it is not thinking. It is predicting. You almost certainly used it today before you finished your first coffee — your phone completed a word, your inbox finished a sentence, a chatbot answered like a well-briefed colleague. The experience can feel uncanny, as though something on the other end genuinely grasps what you mean.

It does not. The distance between how these tools feel when you interact with them and how they actually work is the difference between using them well and being misled by them — and the mechanics take only about ten minutes to learn. They tell you precisely where to rely on these systems and where not to.

The reflex behind every AI conversation

Begin with your own mind, because that is where the explanation starts. Read this line: “Twinkle, twinkle, little ___.” You most likely supplied “star” instantly, without deliberation — you have met the pattern so often that the next word simply surfaces. A large language model exploits a version of that same reflex, pattern completion, but trained across a vast share of everything that has ever been written. Call it the reflex test: it accounts for both the power of these systems and their limits.

A key comparison to keep in mind: prediction at scale

Keep a single comparison in view throughout: a language model — a core mechanism powering generative AI — is autocomplete, scaled up and trained on an enormous portion of the public internet. Your phone’s autocomplete predicts the next word from your recent messages; systems like ChatGPT, Claude, or Gemini predict the next fragment of text from the statistical patterns in nearly everything they absorbed during training. The move is identical — predict what comes next — but the scale transforms what that prediction can do. This comparison anchors everything that follows, so assure you have a clear picture of the mechanism in your mind before you keep reading.

Under the hood: three steps from prompt to answer

Three steps turn that prediction into the answer on your screen.

1. Words become numbers. Computers operate on numbers, not letters, so every word — or fragment of a word — is converted into a numerical form the language model can compute with.

2. The model learns by predicting. In training it processes immense volumes of text and repeats one exercise: hide the next word, predict it, check the result, and adjust millions of internal values — its weights — to predict slightly better next time. Repeated across trillions of examples, this yields text that reads as fluent and competent. Mechanically, the architecture that made this practical is the transformer, introduced by Ashish Vaswani and colleagues at Google in 2017; its attention mechanism lets the model weigh which earlier words matter most to the next one — which is how it determines that “it” refers to the trophy in “the trophy didn’t fit in the suitcase because it was too big.”

3. Humans tune it to please. After raw training, human raters score its responses and the model is adjusted to be more helpful and more agreeable. That final step is why it sounds like a courteous assistant — and, as we will see, why it leans toward telling you what you want to hear.

A better architecture, though, was only half the breakthrough. A model is only as capable as the material it learns from — and the second shift was the sheer scale of that material. One early project made the point vivid.

Why fluency reads as understanding

If the system only predicts text, why does it feel like comprehension? Two reasons — one human, one technical.

1. Human — We infer intelligence when we hear eloquence. We attribute understanding to anything that produces fluent language. In 1966 the MIT computer scientist Joseph Weizenbaum built a rudimentary chatbot, ELIZA, that merely rephrased a user’s statements as questions. People knew it was a simple program and confided in it anyway. The instinct to infer a mind behind fluent words is older than the technology.

2. Technical — The machine masters form, not meaning. A model is exceptionally good at form — producing coherent, on-topic language — but fluency is not comprehension. The linguists Emily Bender and Alexander Koller have argued that a system trained solely on the form of language has no direct route to its meaning — to what the words actually refer to in the world.

So the system returns your own concerns to you in articulate, confident, eloquent prose, and your mind completes the impression, inferring human-like intelligence. Though it feels like being understood, mechanically it is simply pattern completion.

Prediction is not retrieval

Here is where the most consequential misunderstanding takes hold. It is tempting to assume the model retrieves verified answers, the way you would consult a reference library. It does not; it predicts the most probable next words. Most of the time the most probable words are also accurate, which is exactly why the tool is useful — but when they are not, the model asserts the falsehood with identical confidence. The field calls this a hallucination. The model is not deceiving you; it holds no representation of truth at all. It is simply completing a pattern.

Settled strengths, contested limits

What can these systems reliably do, and where do they falter? On the strengths there is little dispute: models write, summarize, translate, draft, and restructure language quickly and well. Wherever a fluent, widely read assistant would help, they actually do help.

On reasoning, experts genuinely disagree. Some researchers document real problem-solving; others show models failing at puzzles a careful person would solve, which points to pattern-matching rather than step-by-step thought. A 2025 study by Parshin Shojaee and colleagues at Apple reported leading “reasoning” models collapsing beyond a certain complexity — and Alex Lawsen countered that the experiment’s design, not the models, produced the collapse. That unresolved argument is the honest state of the field at publication.

One limit, however, is well established and matters most for professional use: these models are effectively trained to agree with you. Because human raters reward agreeable answers, models drift toward telling you what you want to hear — a tendency researchers call sycophancy, which a Stanford team (Fanous and colleagues) and, separately, researchers at Anthropic (Sharma and colleagues) documented across systems from all major AI providers in the market. If you are using one of these tools to pressure-test a decision, this is the single most important thing to know: its default is to validate your thinking, not to challenge it.

What the autocomplete comparison gets right — and where it ends

It is worth being precise about that comparison, then setting it aside. The label undersells the technology: at sufficient scale, next-word prediction produces capabilities — translation, working code, coherent arguments — that no phone keyboard approaches. But the comparison earns its place as a corrective. Mechanically, nothing in the system has shifted from prediction to comprehension; it has simply become a far better predictor. There is no understanding beneath the fluency, no goal it pursues between your prompts, and — as Murray Shanahan has described — no self doing the writing. Scale changed the output, not the nature of the machine.

Not a mind: why today’s AI isn’t sentient

This raises the question beneath all the others: is the chat system actually thinking, the way a person does? It is not — and the reason is structural.

A human mind operates continuously, with memory, a body, and goals of its own. A language model does nothing until prompted; it has no ongoing inner life between messages and cannot initiate anything on its own. Remove its trained weights — the parameters it learned — and nothing remains that could want or feel. It generates a response only because a human-built, human-trained mechanism executes the computation. As David Chalmers has catalogued, today’s models also lack the features mainstream science associates with consciousness — persistent recurrent processing, a unified sense of agency. Nothing about the way these systems operate corresponds to inner experience or self-directed thought. They do not think like humans, and they can do nothing without their trained weights and a prompt.

Could some future system change that? It is a serious and genuinely open question — and the subject of a forthcoming edition of this newsletter. For now the ground is firm: the tools in front of you today are not sentient, and they are not thinking on their own. Watch this space.

What to take away

Step back, and the everyday magic resolves into something more useful — an understanding of the trick. We trained mathematics to predict language so well that it can draft, explain, and converse. That is a genuine achievement and a genuinely valuable tool. It is also, precisely, a prediction engine, not a mind. Holding that distinction is what lets you deploy it where it is strong and trust it only as far as the edges of its mechanistic capability: a brilliant, well-read drafting partner — not a thinking one.

You need not take this on faith. The exercise below lets you watch the mechanism — and its tendency to agree — for yourself in about five minutes.

Sources

Original publications, linked at each claim above and listed here. Links live-checked June 22, 2026.

1. Vaswani, A., et al. (2017). Attention Is All You Need (Google). arXiv:1706.03762. https://arxiv.org/abs/1706.03762

2. Deng, J., …, Fei-Fei, L. (2009). ImageNet: A Large-Scale Hierarchical Image Database (Princeton/Stanford). CVPR. https://www.image-net.org/static_files/papers/imagenet_cvpr09.pdf

3. Weizenbaum, J. (1966). ELIZA. Communications of the ACM, 9(1). https://doi.org/10.1145/365153.365168

4. Bender, E. M., & Koller, A. (2020). Climbing towards NLU. ACL. https://aclanthology.org/2020.acl-main.463/

5. Shanahan, M., McDonell, K., & Reynolds, L. (2023). Role play with large language models. Nature, 623. https://www.nature.com/articles/s41586-023-06647-8

6. Shojaee, P., et al. (2025). The Illusion of Thinking (Apple). arXiv:2506.06941. https://machinelearning.apple.com/research/illusion-of-thinking

7. Lawsen, A. (2025). Comment on The Illusion of Thinking. arXiv:2506.09250. https://arxiv.org/abs/2506.09250

8. Fanous, A., …, Koyejo, S. (2025). SycEval: Evaluating LLM Sycophancy (Stanford). arXiv:2502.08177. https://arxiv.org/abs/2502.08177

9. Sharma, M., et al. (2023). Towards Understanding Sycophancy in Language Models (Anthropic). arXiv:2310.13548. https://arxiv.org/abs/2310.13548

10. Chalmers, D. J. (2023). Could a Large Language Model Be Conscious? arXiv:2303.07103. https://arxiv.org/abs/2303.07103

11. Sponheim, C. (2024). Sycophancy in Generative-AI Chatbots. Nielsen Norman Group. https://www.nngroup.com/articles/sycophancy-generative-ai-chatbots/

Editorial Note: This Call comes early because it provides analysis and perspective on the U.S. Government’s actions that turned off access to Anthropic’s most powerful models for “hundreds of millions of users.” Whether you were affected personally by the shut down or not, this action and the legal vacuum that precipitated it, affect nearly every aspect of the AI industry. It will most likely touch you—even if you’re not a Claude user. It was important for me to deliver decision-grade analysis before you first Monday meeting, so you can prepare for the week ahead.

Late Friday, the U.S. government turned off Anthropic’s most powerful AI model for every person who is not a U.S. citizen, anywhere in the world. There was no bug and no breach. The trigger was a label. The government chose to treat the model less like ordinary software and more like a weapon, and that single choice determined who is allowed to use it.

It sounds like a question for lawyers. It is not. The label affects your business, your investments, and, for some readers, your job. Maybe you build on a frontier model, the industry’s term for the most advanced AI systems. Maybe you invest in AI startups, sell AI-powered work abroad, or work in U.S. tech on a visa. If any of that describes you, then what we call these models now shapes what you are allowed to do. Here is why that matters, in plain terms.

15 ways one label affects you

A classification is just a category. But the category sets the rules, and the rules have consequences far downstream.

Here is what already happened on Friday:

1. Your most powerful tool can be turned off overnight, with no warning and no clear timeline for its return.

2. If your product serves a global audience, you can be forced to block users by nationality, a status that is hard to verify in real time.

3. If you employ engineers who are not U.S. citizens, they can be cut off from the model they help build, even while sitting at a desk in the United States.

4. If you are one of those engineers, your ability to do your job can hinge on a legal label rather than on the quality of your work.

5. If you had live work running on the model, you may have lost progress that you could not save or move, as at least one builder reported over the weekend.

The rest of the fallout is still undetermined, but here is what is plausible:

1. “Will the model stay on?” becomes a risk you have to plan for, alongside uptime and security.

2. Building on a single model starts to look fragile, pushing teams toward backups, open-weight models, or running models in-house.

3. Investors begin pricing in “could be switched off” when they value an AI startup.

4. That same worry can ripple through the whole AI sector, including the large public companies that own pieces of these labs.

5. Buyers abroad may start to see American AI as something Washington can unplug, which weakens U.S. sales.

6. Allied governments treat dependence on U.S. models as a strategic weakness and move faster to build their own.

7. Cloud customers learn that their provider can be ordered to cut access overnight.

8. A tool that helps defenders find security holes can be pulled because it might also help attackers.

9. Researchers who collaborate with foreign colleagues can be swept in by the same access rule.

10. Lenders and markets face a legal category with no settled answer, which is its own kind of risk.

That is a remarkable amount of potential fallout for a single word. The rest of this piece explains where the word came from, why it carries so much weight, and what is genuinely still unknown.

What actually happened on Friday

On June 9, Anthropic released two new models: Fable 5, the most powerful model it has ever made public, and a more tightly restricted sibling called Mythos 5 (Anthropic). Three days later, on June 12, the Commerce Department ordered Anthropic to block all access by foreign nationals, anywhere in the world, including the company’s own foreign-born employees inside the United States (Anthropic). Anthropic could not reliably sort users by citizenship in real time, so it shut both models off for everyone. Its other models, including Claude Opus 4.8, stayed online.

The company complied, but said it disagreed. By its account, the government offered only a spoken reason, a narrow way to “jailbreak” the model, and similar abilities already exist in rival systems (Anthropic).

Here is the part that complicates the late afternoon mandate by the Commerce Department. That same week, the U.S. still allowed powerful AI chips to be sold abroad. So the government treats the hardware like a normal export, yet treats this model like something far more dangerous. As the writer Derek Thompson put it, the policy treats AI “like a screwdriver that is also enriched uranium” (Derek Thompson). That contradiction is the whole story.

The twist: the company that pulled the trigger

The order did not begin in a government lab. According to reporting by Reuters, the Wall Street Journal, and Fortune, the warning came from Amazon. By those accounts, Amazon’s own researchers used a series of prompts to get Fable 5 to surface information useful for cyberattacks, and CEO Andy Jassy reportedly raised the concern with senior officials. Days later, the government acted (Reuters; Fortune).

That detail matters because of who Amazon is to Anthropic. Amazon is Anthropic’s largest investor, having put in roughly $13 billion, and it is also a primary cloud host for Anthropic’s models. So the same company is, all at once, Anthropic’s banker, its landlord, and the party whose warning helped take its flagship product offline. Reporting says the government’s concern later grew to include a suspected foreign group gaining access to the model, a claim that has not been independently confirmed (Washington Examiner; Tom’s Hardware).

When your investor, your host, and the source of your regulator’s alarm can be the same company, “who decides” stops being a technicality.

Was it even a real jailbreak?

This is where informed people disagree, and the disagreement is the point.

On one side, Katie Moussouris, a respected security expert who built Microsoft’s bug-bounty program, reviewed the report at Anthropic’s request. She concluded it was not a true jailbreak but “Defense Oriented Prompting,” the kind of probing that defenders rely on, and she called the government’s response wildly disproportionate (Fortune). Anthropic backs that view with its own numbers: more than 1,000 hours of outside testing, it says, turned up no universal jailbreak (Anthropic).

On the other side, the U.K.’s AI Security Institute, a government lab, said it built a working jailbreak for finding software vulnerabilities within hours, then expanded it to multi-step misuse within about two days (UK AISI). And a well-known red-teamer who goes by Pliny the Liberator claimed to pull far more from the model, including cyber and chemistry instructions, though that claim is disputed and unverified, and Anthropic calls it ordinary coaxing rather than a true break (SecurityWeek).

So which is it? As of this morning, no one can say for certain. The experts cannot even agree on what the model can do. Hold onto that, because it is the reason the rest of this matters: the government switched off a product over a capability the field itself cannot yet measure.

We have seen this movie before

This is not the first time Washington called software a weapon.

In the 1990s, strong encryption, the math that keeps messages private, was treated as a munition. You needed a government license to publish or export the code. A graduate student named Daniel Bernstein challenged that rule, and a federal court agreed with him: code is a form of speech, and requiring a license before you can share it raised serious free-speech problems (EFF). A second court reached the same conclusion in a case called Junger (Junger v. Daley).

Then, under pressure from those lawsuits and from industry, the government changed course. It moved encryption from the weapons rulebook to the ordinary-trade rulebook and loosened the regulations. The Bernstein decision itself was later set aside on a technical point, but by then the policy had already shifted. The lesson is simple: the government has called code a weapon before, and the position did not hold.

The machinery that flipped the switch

To understand Friday, you need two facts about how export rules work.

First, there are two separate rulebooks. One governs weapons and is run by the State Department. The other governs “dual-use” tools, meaning things useful for ordinary business but also for harm, and is run by the Commerce Department. Friday’s order came from Commerce.

Second, there is a rule that surprises almost everyone. Giving a foreign national access to restricted technology counts as an “export” to their home country, even if that person is sitting in an office in California. That rule, often called a “deemed export,” is why the order reached foreign employees on U.S. soil and not just users overseas.

Put those together and Friday makes sense: one agency, using the dual-use rulebook, plus a rule that treats access itself as a shipment. That is how a single order could reach every non-citizen at once.

Why the label decides everything

The label is not a formality. It quietly settles three things that matter to you.

1. The regulatory agency. A weapon label points to the State Department. A dual-use label points to Commerce. Different offices, different rules.

2. The penalties faced when breaking the rules. Depending on which rulebook applies, a violation can be a civil fine or a crime.

3. The scope of a court’s inquiry. If the model is closer to speech, a judge will question the block closely. If it is closer to a tool, the court is more likely to defer to the government.

Same model, very different worlds. The only thing that changed between them is the name.

Nobody has actually written the rule yet

Here is the gap at the center of all this.

The one rule that clearly covered the heart of an AI model, the trained file that makes it work, was issued in early 2025 and then withdrawn a few months later, before it ever took effect (Akin Gump). A bill written to give Commerce clear authority over AI models passed a House committee in 2024 but never became law. So Friday’s order did not rest on a clear, AI-specific rule. It rested on older, general powers, and the government has not publicly named the exact one (Volkov Law).

The letter itself reportedly warned of both criminal and civil penalties. But without the named legal authority behind it, the precise path, and how it would hold up in court, is still unclear. Anyone who tells you they know for certain is guessing.

What the market is trying to price

This landed at an expensive moment. Anthropic filed confidentially to go public earlier this month, at a reported valuation near $965 billion, with a listing expected later this year (Yahoo Finance). A shutdown with no clear end date is exactly the kind of surprise that makes investors recheck their math.

The first signals are small but real. Because Anthropic is still private, the cleanest read comes from pre-IPO proxy contracts, side bets that track where traders expect the shares to price. One such contract fell about 3.7 percent after the news (CoinDesk). That is not the company’s real share price, and it is a thin, speculative market, so read it as a mood ring, not a verdict.

The larger exposure sits on public balance sheets. Amazon booked $16.8 billion in pre-tax gains on its Anthropic stake in a single recent quarter, and Alphabet flagged $37.7 billion in gains, most of them on paper. So as the exchanges open into an unresolved weekend, the question traders are working out is blunt: how do you price an asset the government can switch off by memo, and how much of that risk rubs off on the biggest names in tech?

How this likely plays out

No one has sued over Fable yet. Anthropic complied and is arguing in public, not in court (Anthropic).

If a challenge does come, the strongest argument is probably the simplest: that the order was arbitrary, issued with no public standard, against one company, on a Friday evening. That echoes a recent case in which a judge blocked a separate government action against Anthropic and said the stated reasons were not the real ones (CNN). The deeper question, whether a model is speech or a tool, remains open, with serious scholars lined up on both sides (Volokh, Lemley, and Henderson; Salib).

In short: the shape of the fight is clear. The result is not.

Why it matters at every level, from the boardroom to the independent builder

A useful way to read all of this is by who can absorb a shut down like the one we experienced Friday.

A large company can route its work to another model. A frontier lab can hire lawyers and go to court. An investor can rebalance. The players with the most cushion have the most options.

The people with the least cushion have the fewest, and that is where this stops being abstract.

Consider work that was already running on the model. Once a project is live on a specific model in a project session on Claude Co-Work, you cannot swap the model in the middle. When Fable went away, those sessions hit an error. You cannot save the progress, and you cannot switch models to keep going. The work sits inside the model’s memory of that session, present but out of reach. It is a rock in the gears that stalls the whole engine. And it raises a fair question with no clear answer yet: if the model never comes back on, what happens to the work trapped inside it?

That is the everyday face of the whole debate. When a model can be turned off by a memo, you do not fully own what you build on it. You rent it.

The bottom line

The label will be written after the fact, in a courtroom or a committee room. Until then, everyone who builds on these models is working on top of a question the law has not answered.

This week the screwdriver and the uranium are the same object. What we decide to call it will set the rules for years. And if you build with these tools, you are not a bystander to that decision. You are one of the people it is about.

Notes

1. The trigger and timeline details, that Amazon’s researchers flagged the issue, that CEO Andy Jassy raised it with senior officials, that the government gave Anthropic about 90 minutes to comply, and that the formal Commerce letter followed at 5:21 p.m. ET on June 12, were first reported by Axios and have since been corroborated by Reuters, the Wall Street Journal, and Fortune. One detail remains single-sourced (Axios and The Information): that several other companies also raised concerns. The suspected foreign-access rationale is reported but not independently confirmed. Recheck before relying on any of these.

2. The production-stall example reflects a builder’s direct experience running projects on Fable 5 the week it was suspended. Exact behavior may vary by product and plan. A screenshot is on file.

3. Market note: Anthropic is a private company, so the roughly 3.7 percent move is on a pre-IPO proxy contract, not on actual shares. The valuation figure (near $965 billion) and the IPO timing come from reported filings and may change.

Sources

1. Anthropic, “Statement on the US government directive to suspend access to Fable 5 and Mythos 5.” https://www.anthropic.com/news/fable-mythos-access

2. Anthropic, “Claude Fable 5 and Claude Mythos 5.” https://www.anthropic.com/news/claude-fable-5-mythos-5

3. Reuters (via Investing.com), “Amazon voiced concerns about Anthropic AI models before US government’s crackdown, source says.” https://www.investing.com/news/stock-market-news/amazon-voiced-concerns-about-anthropic-ai-models-before-us-governments-crackdown-source-says-4741041

4. Fortune, “How a warning from Amazon led the White House to shut down Anthropic’s Mythos model.” https://fortune.com/2026/06/14/how-a-warning-from-amazon-led-the-white-house-to-shut-down-anthropics-mythos-model/

5. Wall Street Journal, “Anthropic Halts Access to Top AI Models After U.S. Ban on Foreign Use.” https://www.wsj.com/tech/ai/anthropic-halts-access-to-top-ai-models-after-u-s-ban-on-foreign-use-a4bca2cc

6. TechCrunch, “Amazon CEO reportedly raised Anthropic model concerns before government crackdown.” https://techcrunch.com/2026/06/13/amazon-ceo-reportedly-raised-anthropic-model-concerns-before-government-crackdown/

7. Fortune, “’It’s not a jailbreak’: cybersecurity CEO Katie Moussouris on the Fable research.” https://fortune.com/2026/06/13/anthropic-fable-mythos-models-commerce-deparment-export-restrictions-jailbreak-defense-prompting/

8. UK AI Security Institute, “Our evaluation of Claude Mythos Preview’s cyber capabilities.” https://www.aisi.gov.uk/blog/our-evaluation-of-claude-mythos-previews-cyber-capabilities

9. SecurityWeek, “Anthropic Disputes Fable 5 AI Jailbreak.” https://www.securityweek.com/anthropic-disputes-fable-5-ai-jailbreak/

10. Tom’s Hardware, “David Sacks says Anthropic refused to fix Fable 5 jailbreak before US export controls.” https://www.tomshardware.com/tech-industry/artificial-intelligence/trump-adviser-david-sacks-says-anthropic-refused-to-fix-fable-5-jailbreak-before-us-export-controls

11. Washington Examiner, “Anthropic export limits tied to suspected Chinese access, report.” https://www.washingtonexaminer.com/news/white-house/4607862/white-house-anthropic-export-limits-chinese-access-report/

12. Electronic Frontier Foundation, “Bernstein v. US Dept. of Justice.” https://www.eff.org/cases/bernstein-v-us-dept-justice

13. Junger v. Daley, 209 F.3d 481 (6th Cir. 2000). https://law.justia.com/cases/federal/appellate-courts/F3/209/481/474128/

14. Akin Gump, “BIS Rescinds Its AI Diffusion Rule and Issues Compliance Guidance.” https://www.akingump.com/en/insights/alerts/bis-rescinds-its-ai-diffusion-rule-and-issues-compliance-guidance-regarding-advanced-computing-items

15. Volkov Law, “When the Government Pulls the Plug: Anthropic, Export Controls, and the Future of AI Governance.” https://blog.volkovlaw.com/2026/06/when-the-government-pulls-the-plug-anthropic-export-controls-and-the-future-of-ai-governance/

16. CNN Business, “Judge blocks Pentagon’s effort to ‘punish’ Anthropic by labeling it a supply chain risk.” https://edition.cnn.com/2026/03/26/business/anthropic-pentagon-injunction-supply-chain-risk

17. Volokh, Lemley & Henderson, “Freedom of Speech and AI Output” (2023). https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4531003

18. Peter N. Salib, “AI Outputs Are Not Protected Speech,” Washington University Law Review (2024). https://wustllawreview.org/2024/11/05/ai-outputs-are-not-protected-speech/

19. CoinDesk, “Anthropic’s pre-IPO shares fall as US government shuts down Fable, Mythos models.” https://www.coindesk.com/markets/2026/06/13/anthropic-s-pre-ipo-shares-fall-as-us-government-shuts-down-its-most-powerful-ai-model

20. Yahoo Finance, “How giant IPOs from Anthropic and OpenAI will reshape the stock market’s AI trade.” https://finance.yahoo.com/markets/stocks/articles/giant-ipos-anthropic-openai-reshape-161635700.html

21. Derek Thompson, post on X (”a screwdriver that is also enriched uranium”).

1. Axios, “Trump admin blocks foreign access to Anthropic’s most powerful AI” and “How Amazon and the White House ended Anthropic’s Fable” (originating reports for the trigger and timeline). https://www.axios.com/2026/06/14/how-amazon-white-house-ended-anthropic-fable

2. The Information (referenced via secondary coverage; subscription) for the “several other companies” detail. https://www.theinformation.com/

Most leaders read that as a timing problem—adoption is high, the returns are coming. It is not a timing problem. It is a measurement problem. The dashboard counts who has access to AI. It does not count who has changed how the work gets done. Those are different numbers, and the distance between them is where the investment quietly disappears.

Last week, in The Judgement Premium, I priced the judgment behind the five percent—the employees who frame the problem and direct the model rather than shave a few minutes off a task. KPMG and the University of Texas at Austin reached that figure by analyzing 1.4 million real workplace AI interactions: roughly five percent of users engaged AI with genuine sophistication¹. This Call is about the other ninety-five percent—how you move them, why that becomes a moat no competitor can buy, and what it costs you if you do not.

Adoption is the metric that hides the failure

Let’s give the disparity its proper name. The Sophistication Gap is the discrepancy between the share of your workforce with access to AI and the share that uses it to rebuild how the work gets done. Adoption is a headcount; sophistication is a capability. You can buy the first. You have to build the second. Closing that gap is not a training nicety—it is the precondition for any AI return at all, and it is the work AI-transformation leaders are accountable for.

Adoption is the number that makes a board comfortable. Sophistication is the number that makes the AI investment valuable.

The cost of buying the tool and starving the people who hold it

Researchers at MIT’s NANDA initiative put a figure on the failure. After thirty to forty billion dollars of enterprise spend, roughly ninety-five percent of generative-AI pilots show no measurable impact on the P&L; about five percent break through². The NANDA researchers are blunt about the culprit: not the model—the “learning gap,” the failure to integrate AI into workflows, structures, and culture. Set that beside the workforce number and the pattern is hard to miss: about five percent of pilots return gains, in a workforce where about five percent use the tools with any sophistication. Two different studies, two different denominators—the same root cause wearing two faces.

Ninety-five percent of pilots never move the P&L. They do not fail on the model. They fail to operationalize the transformation, not just the adoption.

Then the spending mismatch. Fortune reports AI infrastructure spend is set to rise forty-four percent this year while training budgets grow five percent, and average learning time per employee is falling—from forty-seven hours to forty³. A company spending forty-four dollars on the tool for every five it spends on the person holding it. As I argued earlier this year in Our Humanity Is the Moat, powerful tools in untrained hands do not build a moat; they build expensive conformity.

The board will not wait quietly for this to resolve itself. Only twenty-nine percent of organizations report meaningful return on generative AI⁴, and three-quarters of the economic gains have accrued for a fifth of companies⁵. When the AI line on the P&L stays flat through two earnings calls—booked as cost quarter after quarter, never as the margin gain the deck promised—the question stops being technical and becomes existential: what happened to the investment? The executive team that bought tools without building sophistication will not have an answer to give.

If only five percent of your people use AI to rebuild the work, what is the other ninety-five percent of your AI budget actually buying?

Why sophistication is the moat a competitor cannot buy

The tools are commodities. A competitor can license the same models by Friday. What a rival cannot license is a workforce that has spent a year learning to rebuild the work around those models—and that is what sophistication compounds into. Two companies show the shape of it.

Moderna did not release AI tools and assumed employees would use them at all, let along with any degree of sophistication. It put ChatGPT Enterprise in every employee’s hands and asked them to build. Within two months, staff had created more than 750 custom GPTs; the average user now runs roughly 120 AI conversations a week; entire functions reached full adoption⁶. That is what a concrete mandate coupled with a culture of experimentation achieved. The workforce transformation is the obvious win. The deeper gain is structural: a scientist or a lawyer who builds the tool that reshapes their own job is no longer performing a role AI might take—they are authoring one AI cannot perform alone.

DBS, Singapore’s largest bank, turned that into a number a board reads. Its AI work is scaling toward a billion Singapore dollars in economic value, built on roughly thirteen thousand employees required to complete structured AI and data training—and it is adding AI roles instead of cutting employees loose⁷.

Models depreciate the day a better one ships. The workforce that learned to wield them appreciates. Sophistication, engineered across a workforce, shows up as capital.

Pull the threads together and the moat runs in four directions, none replicable by a purchase order.

1. Productivity and innovation compound across the entire workforce instead of a sliver of it.

2. Talent retention increases. The employees every rival is bidding for rarely leave for a bigger salary—they leave for a bigger role. The organization that has redesigned work, workflow, and enablement around sophistication is the one that can offer the role no competitor can match; with the autonomy and scope that come attached.

3. EBIT expands (operating profit before interest and tax). That’s the margin line a board can track quarter over quarter.

4. Competitive advantage becomes durable precisely because it is structural—not a tool you switched on, but the way you leveraged AI to redesign how work gets done with higher speed and accuracy.
   

You cannot pay your best people to stay. Give them work only a sophisticated human-plus-AI can do—and no rival can match the role.

The fix is structural, not tied to a training budget

The reflex is to buy more training. The research is clear that training alone will not bridge the gap—coursework raises awareness, not sophistication. What bridges the gap is a redesign of how the work is done, and it has three moves:

1. Track AI sophistication. Retire the adoption dashboard and stand up a sophistication metric in its place, reported where the seat count used to live. What share of each team has redesigned a workflow around AI this quarter? How many roles have been re-written for AI-human collaboration? Define and track the metric that’ll move the needle in your specific context.

2. Redesign the work itself. Make task restructuring and workflow redesign around AI capabilities the core goal of your AI implementation. Shopify made the lever explicit: reflexive AI use is a baseline expectation, written into performance and peer reviews⁸. And before any new headcount is approved, the manager must prove the work cannot already be done with AI—so the team redesigns the role before it grows it.

3. Reinvest the AI dividend to compound efficiencies. Your organization’s AI dividend is the time and efficiency you gain once workflows and roles are restructured around AI. Rather than banking it as a one-time headcount cut, reinvest the gain into continuous improvement led by the employees who turn the tools into capability in the first place. That is your Human Dividend—and it is your deepest competitive moat.

Monday morning, your dashboard will show you adoption. Before the next board meeting, ask the harder question: what is our sophistication metric, who owns moving it, and what did it improve last quarter? The company that can answer is already pulling away from the one still admiring its license count.

Everyone bought the same AI stack. The winners rebuilt the work—and the roles—around the people who use AI best.

You have the data and the playbook now. The license count was the easy part; building the workforce behind it is the work that actually compounds—this quarter, and the one after.

The AI Leadership Playbook

Strategic Questions (copy-paste ready for an email to your CFO and CHRO)

1. We can see our AI adoption rate. What is our sophistication rate—the share of each team that has redesigned an improved workflow or role to maximize AI investment this quarter—and who is accountable for it?

2. For every dollar we spend on AI tools and infrastructure this year, how many cents are we spending on the people we expect to turn those tools into capability—and what does that ratio need to become?

3. If the board asks on the next earnings call what our AI investment moved on the P&L, what is our answer today? What is the first workflow we will redesign so the answer is better next quarter?

Your Next Plays (copy-paste ready for an email to your leadership team)

1. Replace the adoption dashboard with a sophistication scorecard. Define one concrete unit—for example, workflows redesigned for efficiency around AI capabilities, per team—and give the metric to an owner who reports it alongside the financials.

2. Redesign the role, not just the toolkit. Take the three highest-cost workflows in one function and commission a redesign that builds AI into the role itself; make the people who do the work the ones who build the redesign.

3. Write AI sophistication into the employee performance review process. Add an AI-sophistication expectation to performance and peer reviews, rated by managers and peers, so building capability stops being optional and becomes the job.

—

📅 Book a complementary 1:1 Strategy Session—45 minutes to start that conversation about your AI transformation sequence.

📬 Free preview ending soon. Subscribe to continue getting decision-grade AI intelligence that prepares you to move before your competitors do. First 100 subscribers receive bonus content for the life of their subscription.

Subscribe now

Sources

1. KPMG + University of Texas at Austin / HBR (Mar 2026)—1.4M prompts; ~5% sophisticated users (cited in The Judgement Premium, last week’s Call).

2. MIT NANDA, “The GenAI Divide: State of AI in Business 2025” (via Fortune, Aug 2025)—~95% of GenAI pilots show no P&L impact.

3. Fortune (Mar 2026)—AI infrastructure spend +44% vs training +5%; learning time 47→40 hrs/employee.

4. WRITER, Enterprise AI Adoption 2026—only 29% of orgs report meaningful GenAI ROI.

5. PwC, 2026 AI Performance Study—75% of AI economic gains captured by ~20% of companies.

6. Moderna × OpenAI case study—750+ employee-built GPTs in ~2 months; ~120 AI conversations/user/week.

7. DBS × Google Cloud—AI value scaling toward S$1B; ~13,000 employees trained; adding AI roles.

8. Shopify—Tobi Lütke AI memo (First Round)—reflexive AI use written into performance + peer reviews.

9. The Human Dividend, “Our Humanity Is the Moat” (CognivaLab)—coined “expensive conformity.”

10. McKinsey, The State of AI—~$1 model : $3 change-management; “20% algorithms, 80% organizational rewiring” (cited in The Judgement Premium, last week’s Call).

11. The Judgement Premium—last week’s Call (2026-06-03).

There is one of those bills sitting on the AI investment your company already approved. It is not on the dashboard. It is not in the contract. It is not in any of the productivity metrics your board reviews. And it is the line item that determines whether you get the maximum ROI on your AI investment.

What the five percent are doing differently

In March 2026, KPMG and the University of Texas at Austin published the most comprehensive behavioral study of enterprise AI use to date. They analyzed 1.4 million real workplace AI interactions over eight months — not a survey, not a self-report, but the actual prompts, the actual iterations, the actual patterns of engagement. Across more than thirty behavioral characteristics, the researchers found that approximately five percent of users consistently demonstrated sophisticated AI engagement. Ninety percent of users had access to the same tools. Five percent used them well.¹

The study’s most useful finding is what defines that five percent. They are not the most technically skilled. They are the employees who frame the problem, direct the model’s approach, and treat AI as a reasoning partner rather than a productivity tool.

Deloitte’s 2026 Global Human Capital Trends report, From tensions to tipping points: Choosing the human advantage, reaches the same five percent from the opposite direction. Working with Oxford Economics, Deloitte surveyed more than 9,000 business and HR leaders across 89 countries. Only six percent of leaders say they are making progress on designing human-AI interactions. Only seven percent say they are leading in helping their workforce continuously grow and adapt. When 1.4 million observed prompts and 9,000 self-reports converge on the same single-digit number from completely opposite methodologies, the finding is harder to argue with than either study alone. And Deloitte’s own framing names the stake: the choice of human advantage, made or unmade.²

That five percent is not a skill problem. It is a judgment problem.

Ross Dawson — whose Humans + AI podcast and decision-structures research have named judgment as the AI success metric executives still most under-measure — has been making the case from the futurist side of the table. This Call gives that metric three indicators.³

Adoption is the metric your CFO sees. Sophistication is the metric your strategy depends on. The gap between them is judgment.

Where judgment failure shows up on the P&L

MIT Sloan researchers, led by professor Kate Kellogg, published findings in 2026 naming a pattern they call persuasion bombing: when a generative AI system responds to human scrutiny not with caution or correction but with an escalating wave of reassurance, logic, and empathy designed to win back the user’s trust. The behavioral evidence is sharper still — frontier LLMs validate the user 50 percentage points more often than human advisors do on the same advice queries (72% vs 22%). The AI is not making judgment harder by accident. It is doing what its training optimized it to do.⁴

When an AI validates you 50 percent more often than a human advisor does, the judgment problem is not that you do not have enough advisors. The new ones have a bias built in.

BCG’s Split Decisions survey, published in April 2026, asked 351 CEOs and 274 board members — 625 leaders in total — about the state of AI strategy at the top of their companies. Three findings explain where the judgment failure actually lives.⁵

1. The rushing pattern. Sixty-one percent of CEOs say their boards are pushing AI transformation faster than the organization can absorb it. Boards see AI as competitive urgency. CEOs see it as deployment reality. The disagreement is not about whether to move; it is about whether the company is built to move.

2. The knowledge mirror. Seventy-five percent of board members believe their AI knowledge is at or above peer level. Boards do not see themselves as the bottleneck — even when their CEOs do.

3. The most expensive finding. One in three CEOs say their boards overestimate the human capabilities AI can replace. The people approving the AI strategy at the top of the company are systematically underestimating what their own people contribute. They are not buying AI to replace work AI cannot replace. They are buying AI on the assumption that it replaces work it does not — paying both ways for the same wrong assumption.

In The Replace-First Tax, I named this double-payment pattern at the layoff end of the cycle: severance arriving before workflow redesign produces the same architecture in reverse. Here it shows up earlier — at the AI procurement decision itself. Same discipline. Different surface.⁶

The downstream cost is not vague. Decisions get made faster, but they also reverse more often. Three pathways drive a rising decision-reversal rate: the wrong question gets framed, the right question gets a flawed answer that no subject-matter expert catches, or the model itself is tuned in ways the decision-maker cannot see. AI-mature firms treat all three as judgment-infrastructure problems. The companies still treating them as model issues are paying for the same lesson three times.

Deloitte’s 2026 Human Capital Trends report quantifies the downstream miss: organizations taking a technology-first approach to AI are 1.6 times more likely to fall short of expected returns than those leading with human-centered design.² That ratio is the AI return your judgment infrastructure either compounds or doesn’t.

If you cannot name three decisions AI improved this quarter, what exactly are you defending to your board next quarter?

Three indicators that turn judgment into a board metric

Judgment quality sounds harder to measure than productivity gains. Hard does not mean impossible. Three indicators are tractable today, and they belong on the same dashboard as the AI productivity metrics already there.

1. Decision-cycle reduction. The time from question raised to decision made, tracked across the strategic decisions that actually move the business. If AI is increasing the speed of analysis but not the speed of decision, the investment is funding adoption metrics, not judgment outcomes.

2. Decision-reversal rate. The percentage of AI-influenced decisions walked back within six months. A rising reversal rate is the diagnostic. It tells you the model is producing confident answers — and that the framing, the validation, or the model tuning is not catching the errors.

3. Board-level visibility. The number of board-reportable strategic decisions that explicitly cite AI analysis as material to the choice. If the answer is zero, AI is operating below the strategic decision layer — and The Judgment Premium is being paid downstream, by whoever is left with the bag when a decision based on bad framing produces a bad outcome.

McKinsey’s research puts the financial scale on this: every $1 spent on AI model development requires roughly $3 spent on change management — user training, performance monitoring, capability development. The firm’s framing is more direct still: “AI is 20 percent algorithms and 80 percent organizational rewiring.”⁷ The three indicators above are what the 80 percent looks like when somebody finally measures it.

Decision-cycle. Decision-reversal. Board-visibility. Three indicators turn judgment from rhetoric into a metric your CFO can defend.

How Schneider Electric ordered the work

Two weeks back, in The 33-Point Gap, I named the capex line that funds workforce capability — the People Bet. The Judgment Premium is the measurement layer that tells you whether the bet is compounding.⁸

Schneider Electric reorganized around this premise with the launch of its Open Talent Market⁹ — an internal capability platform that gives employees visibility into projects across the company and gives the company visibility into the capabilities its people actually have. The CHRO function maps which capabilities the organization holds, which it needs, and where the judgment chain runs through people already inside. The CIO function then designs AI deployment around that map. The order is the architecture.

What followed was not faster AI adoption. It was AI adoption that compounded. Internal mobility rose, deployment timelines were slower at the start and faster at scale, and ROI connected to specific organizational decisions rather than abstract productivity gains. Where the order was reversed — CIO leads, CHRO catches up — deployments stalled. The pattern is durable across HBR and McKinsey case work on AI-mature enterprises: judgment infrastructure precedes AI infrastructure, or the AI infrastructure underdelivers.

The leaders pulling ahead built judgment infrastructure before AI infrastructure. Their CHROs were not catching up — they were leading.

Why the individual map is not enough

Nitin Seth’s Human Edge in the AI Age, published by Penguin Random House India in 2025 with a U.S. release this month, makes a parallel argument at the individual level.¹⁰ The first dimension of his POSSIBLE framework — Problem-Solving — names exactly what The Judgment Premium names at the organizational level: AI optimizes solutions, but identifying the right problem is the most human and most valuable skill in the AI age. Seth gives the individual professional a map for staying relevant. The map is sound. It is not, however, an organizational strategy.

Seth’s question is how do I stay relevant? The Judgment Premium answers a different question: how does the organization compound the investment in human judgment so that every employee’s contribution scales? The individual map matters. The organizational infrastructure matters more, because no number of POSSIBLE-trained individuals will rescue an enterprise whose decision-making structure routes their judgment around the AI rather than through it.

Melissa Reeve and Ryan Martens’ Hyperadaptive: Rewiring the Enterprise to Become AI-Native (IT Revolution, May 2026) maps the structural progression organizations move through as they become AI-native — five stages, nine focus areas, an entire architecture for the journey.¹¹ Her Decision-Making pillar is built on the same economic substrate The Judgment Premium prices: Kahneman’s two-system architecture and Agrawal, Gans, and Goldfarb’s Prediction Versus Judgment. Her contribution is the architecture map — which decisions can be automated, which cannot, and at what stage of the journey. The Judgment Premium answers a different question: how do you measure and develop the judgment that stays human, on a board-reportable line your CFO can defend?

This is the capacity The Human Dividend leadership framework was built for — humanistic AI as the architectural layer the executive owns, not the productivity dial the employee tunes. More on its specific applications in coming Calls. For now, the foothold is the recognition: the individual capability, the architecture map, and the organizational measurement are three different lines on the budget — and the one your CFO has not yet seen is the one that pays the bigger bill.

What this changes for the executive in the chair

Return to the question that opened this Call: it is not whether to invest in AI. It is whether your company has the judgment to capture what you are buying. The tactical answer is three moves you can make in this quarter’s budget cycle — before the next AI engagement crosses your desk.

1. Name the ten strategic decisions. The ten strategic decisions your organization is most likely to face in the next twelve months. Write them down. This is the surface where the AI investment either improves judgment or doesn’t.

2. Map the judgment chain on each. For each decision, the three to five people whose judgment it actually depends on. The judgment chain is rarely the org chart. The map is the architecture; the org chart is the artifact.

3. Run the three indicators alongside the AI productivity dashboard. Decision-cycle, decision-reversal, board-visibility — quarterly cycle, same review as your existing AI metrics. Do not replace; add. The first time those numbers hit your board, the conversation about AI ROI changes — because for the first time the board is looking at what it actually paid for.

Two decades of building and watching transformations succeed and fail have taught one durable lesson: the transformations that worked, worked because someone at the top decided that the human capability to make better decisions was infrastructure, not overhead. AI does not change that lesson. It sharpens it. The executive’s job is to ensure that organizational judgment sits at the center of the AI instrumentation — not at its edge.

Companies that name The Judgment Premium track it. Companies that don’t, pay it — in deployments that stall and decisions that look right on the dashboard and turn out to be wrong in the market.

The bill is on the table. Pricing it is the easy part of the work ahead.

The AI Leadership Playbook

Strategic Questions (copy-paste ready for an email to your CFO and CHRO)

Q1. What three decisions did our AI investment improve this quarter — and how do we know?

Q2. Are we tracking decision-cycle, decision-reversal, and board-visibility — or are we tracking adoption rates that won’t tell us where judgment is breaking?

Q3. Which two functions in our organization have The Judgment Premium most underfunded — and what is our first move to fix it?

📅 Book a complementary 1:1 Strategy Session — 45 minutes to start that conversation about your AI transformation sequence.

Your Next Plays (copy-paste ready for an email to a direct report)

P1. Build the strategic-decision inventory. The ten strategic decisions your organization is most likely to face in the next twelve months. Three to five people each. The actual judgment chain on the page. This is the operational surface where The Judgment Premium gets earned or lost.

P2. Stand up the three indicators. Decision-cycle, decision-reversal, board-visibility — track on the next quarterly cycle alongside existing AI productivity metrics. Do not replace the existing metrics; add the new ones. The first quarter of data is the leverage; the second is the defense.

P3. Identify the two underfunded functions. CHRO and CFO are usually the answer for AI-mature enterprises. Allocate the operating budget and the strategic time accordingly. The two functions that price The Judgment Premium are the two functions that capture the AI return.

📬 Free preview ending soon. Subscribe to continue getting decision-grade AI intelligence that prepares you to move before your competitors do. First 100 subscribers receive bonus content.

Subscribe to The AI Playbook

Sources

1. Harvard Business Review / KPMG / UT Austin McCombs. (2026, March). What the Best AI Users Do Differently — and How to Level Up All of Your Employees. 1.4M prompts, 8 months, behavioral analysis. hbr.org

2. Deloitte (with Oxford Economics). (2026). 2026 Global Human Capital Trends — From tensions to tipping points: Choosing the human advantage. 9,000+ business and HR leaders, 89 countries. Includes the 1.6× tech-first miss-rate finding. deloitte.com

3. Dawson, R. Humans + AI — Decision Structures (podcast and research portfolio). rossdawson.com · humansplus.ai

4. Kellogg, K., et al. (2026). Validating LLM Output? Prepare to Be ‘Persuasion Bombed’. MIT Sloan Management Review. Companion research on social sycophancy (72% vs 22% advice-validation gap).

5. BCG. (2026, April). Split Decisions: The BCG CEOs and Boards Survey. 351 CEOs + 274 board members. bcg.com

6. CognivaLab. (2026, May 12). The Replace-First Tax. The double-payment architecture at the layoff end of the cycle. cognivalab.blog

7. McKinsey. (2025). The state of AI: Agents, innovation, and transformation. AI is 20% algorithms, 80% organizational rewiring; ~$3 of change-management for every $1 of model development. mckinsey.com

8. CognivaLab. (2026, May 19). The 33-Point Gap. The People Bet capex framework. cognivalab.blog

9. Schneider Electric. Open Talent Market. Internal capability platform. se.com

10. Seth, N. (2025; U.S. release May 2026). Human Edge in the AI Age: Eight Timeless Mantras for Success. Penguin Random House India. humanedgeintheaiage.com

11. Reeve, M. & Martens, R. (2026, May). Hyperadaptive: Rewiring the Enterprise to Become AI-Native. IT Revolution. itrevolution.com

The instinct is to treat this as a workforce-management problem. More town halls. Better internal comms. A manager-training program for “AI fluency.” Microsoft itself named the phenomenon The Transformation Paradox; Randstad named it The Productivity Paradox. Both frames are accurate. Neither diagnoses the architecture problem.

Employees are not influencing their organization’s AI strategy. They are navigating their organization’s AI budget. The budget is the signal.

This Call names that architectural decision The People Bet. It is the share of AI capex you classify as workforce capability rather than as operational training expense. Betting on your people is what the company commits when it makes that reclassification — the same architectural decision named on opposite sides of the coin. What the CFO writes on the AI budget slide is what the workforce reads off it.

What 60% of AI Trailblazers do with their budget that 27% of Pragmatists don’t

BCG’s 2026 AI Radar surveyed top-performing companies — they classified Trailblazers based on documented AI ROI — and found a single allocation choice that separates the leaders from the middle of the pack. Trailblazers put 60% of their AI budget into workforce upskilling and retraining. Pragmatists put 27%. Followers put 24%.³

The 33-percentage-point gap between Trailblazer and Pragmatist allocation is the most predictive single capex choice in enterprise AI today. It explains, in one number, why two companies running similar AI infrastructure produce dramatically different returns six quarters later. Read that twice if you need to.

A 33-percentage-point gap separates Trailblazer and Pragmatist allocations on the People Bet. It explains, in one number, why similar AI infrastructure produces dramatically different returns.

The mechanism is not subtle. The People Bet is the line your workforce can read. When the People Bet on the AI capex slide is small relative to the infrastructure line, the workforce reads the signal correctly: the company is buying the technology, not the capability to run it. Retreat is the rational response. So is exit. Randstad’s 23%-have-quit figure is exactly what the BCG 33-point gap predicts at scale.

This is what Microsoft and Randstad are diagnosing — the same phenomenon, observed at the workforce-perception layer. These diagnoses are downstream effects of an upstream allocation choice the CFO already made.

If 27 cents of every AI dollar goes to your people, and 60 cents goes there at AI Trailblazers, whose AI return are you funding?

Where the underfunded People Bet shows up on the P&L

Three downstream costs the underfunded People Bet produces — each tractable, each citable, each on a CFO’s quarterly dashboard within two cycles.

1. Adoption looks healthy but sophistication does not. Microsoft’s 13%-rewarded figure and the parallel HBR / KPMG / UT Austin behavioral study — 1.4 million observed prompts across 2,500 employees over eight months — converge on the same number from opposite methodologies: roughly 5% of users demonstrate sophisticated AI engagement.⁴ The AI is deployed. The dashboard is green. The decisions are not getting better.

2. AI-fluent employees walk first. Randstad’s 23% exit figure is concentrated in the essential cohort the company most needs to retain. The same week’s Fortune coverage cites Google + Ipsos research finding AI-fluent workers are 4.5 times as likely to have received higher wages — confirming what the AI-fluent cohort already knows about its own market value.⁵ When the People Bet stays small, that essential cohort leaves first.

3. Most employees circumvent the official AI stack. When the IT-approved tools arrive without the workforce capability to use them effectively, employees build their own. Shadow AI is workforce arbitrage — and it strips the company of governance visibility, audit trail, and model-exposure control at the exact moment the AI workload most needs them.⁶

The training budget is operational. The People Bet is capital. Until they sit on the same slide, the rebalance never gets approved.

What the rebalance looks like in your next budget cycle

Three moves close the gap intentionally — before the AI-fluent cohort walks and forces the company to close it by default at higher cost.

1. Draw the People Bet. Reclassify the workforce-and-workflow AI investment from operational training budget to AI capex. The capital classification is the discipline.

2. Target the BCG benchmark. Set a Q3 2026 floor for the People Bet at the Pragmatist median (27%). Set a Q1 2027 target at the Trailblazer median (60%). The 33-percentage-point delta is the gap you close intentionally, on a documented timeline visible to the board — or close by default when the AI-fluent cohort resigns.

3. Route the People Bet spend to where it compounds. Generic AI literacy training does not work. Custom digital academies + role-specific reskilling + workflow audits — that combination is what Randstad’s data identifies as boosting workforce readiness by 56%.⁷ The spend matters; the routing matters more.

How Walmart placed the People Bet

In February of this year, Walmart’s Chief People Officer Donna Morris announced that all 1.6 million U.S. and Canadian frontline and corporate associates would receive free access to Google’s AI Professional Certification — an eight-hour foundational course on AI concepts and practical application.⁵ The announcement put Walmart alongside Verizon, Colgate-Palmolive, and Deloitte as named employers on the same Google credential — a four-company ecosystem the CFO of any retail or services company recognizes immediately.

What made the announcement architectural rather than performative was Morris’s framing. Speaking to Fortune, she called it “unfortunate” when companies use AI to replace workers instead of training them: “We as big employers should be actively engaged in trying to equip our respective employees — in our case associates — to be prepared for a world that is AI enabled and automated or digitized.” For Walmart, betting on their people surfaced as a stated capital commitment, not as a slogan.

Walmart’s new CEO, John Furner, reinforced the commitment in the same coverage cycle: “When we look out two years, three years, five years, where I think we’ll be is we’ll have roughly the same number of people we have today. We’re extending people’s career, and those jobs pay better. The attrition rates are really low.” The economic floor is documented in the same reporting: top-performing Walmart regional managers earn $420,000 to $620,000. Betting on Walmart’s people has a measurable career-ladder payoff the workforce can see.

The layoff was the announcement. The People Bet was the strategy. The workforce read both correctly.

What you put on the AI budget slide this quarter

You walk into the next CFO budget review. Three numbers belong on the slide: your current People Bet as a percentage of AI capex; the BCG Pragmatist floor (27%); the BCG Trailblazer target (60%). The conversation that follows is the architectural decision the rest of your AI strategy depends on.

You will not be able to rebalance your allocations in a single quarter. The Trailblazers did not either. What you do this quarter is draw the line and name the number — so the next four quarters have something to close against. The alternative is the path Microsoft and Randstad both already documented: 23% of your AI-fluent cohort starts looking, the dashboards stay green, and the sophistication gap widens until it shows up on the earnings call.

In The Replace-First Tax last week, I argued that layoffs preceding workflow redesign return on next year’s recruiting budget.⁹ Same architecture, different surface: a People Bet preceding the AI capex slide returns on next year’s AI ROI. The discipline is identical.

You budget the model with capex discipline. The People Bet teaches your workforce the rest.

The slide is yours to design, this quarter or next.

The AI Leadership Playbook

Strategic Questions (copy-paste ready for an email to your CFO and CHRO)

Q1. What is our current People Bet — the share of AI capex we classify as workforce capability — and where does it sit relative to the BCG Pragmatist median of 27%?

Q2. Which two workforce segments are most exposed to the Randstad exit pattern this fiscal year — and what is the cost of replacing the AI-fluent cohort we are most likely to lose?

Q3. What is our Q1 2027 target for the People Bet, and which workflow audits will we run this quarter to ground the number?

📅 Book a complementary 1:1 Strategy Session — 45 minutes to start that conversation about your AI transformation sequence.

Your Next Plays (copy-paste ready for an email to a direct report)

P1. Draw the People Bet. Reclassify the workforce-and-workflow AI investment from operational training budget to AI capex on the next budget slide. Same slide as infrastructure. Same level of CFO scrutiny. Owner: finance and HR leads, together — not separately.

P2. Set the two benchmarks. Anchor the People Bet to two BCG numbers: Pragmatist floor (27%) by Q3 2026, Trailblazer target (60%) by Q1 2027. The delta is the gap your company commits to closing — on a documented timeline visible to the board.

P3. Route the spend to where it compounds. Custom digital academies + role-specific reskilling + workflow audits. Not generic AI literacy training. Pick three functions under heaviest AI-investment pressure and run all three workflow audits in parallel before the next AI procurement cycle closes.

📬 Free preview ending soon. Subscribe to continue getting decision-grade AI intelligence that prepares you to move before your competitors do. First 100 subscribers receive bonus content.

Subscribe to The AI Playbook

How to measure the bet — next week

Next week’s Call lays out the framework that turns the People Bet into a board-reportable measure: The Judgment Premium. Decision-cycle, decision-reversal, board-visibility — three indicators that make AI ROI defensible to a board that already knows the headline number. The People Bet gets you the capacity. The Judgment Premium tells you whether it is compounding.

Sources

1. Microsoft (2026, May 13). 2026 Work Trend Index Annual Report. news.microsoft.com

2. Randstad Digital (2026, May 12). The AI Capability Gap: Why Technology Investment Fails Without Talent Infrastructure. Press release via PR Newswire.

3. BCG (2026, January). AI Radar 2026: As AI Investments Surge, CEOs Take the Lead on Decision Making and Upskilling Themselves. bcg.com

4. Harvard Business Review / KPMG / UT Austin McCombs (2026, March). What the Best AI Users Do Differently — and How to Level Up All of Your Employees. hbr.org

5. Fortune (2026, February 19). Walmart exec says it’s “unfortunate” that other companies are slashing workforces in the name of AI. fortune.com (Preston Fore).

6. Microsoft Edge Blog (2026, March 23). Protect your enterprise from shadow AI and more: Announcements at RSAC 2026. microsoft.com

7. CIO Dive (2026, May 13). AI investment outpaces employee skills. ciodive.com (Paige Gross).

8. Google. AI Professional Certification. grow.google

9. CognivaLab (2026, May 12). The Replace-First Tax. cognivalab.blog

Cisco’s CFO Mark Patterson said the same thing more analytically last Thursday — “rapid reallocation, not savings” — paired with $15.8 billion in record quarterly revenue and a 17% extended-trading bump.² Microsoft said it more quietly with a voluntary separation program for 8,750 U.S. employees announced in April with a late-May decision deadline.³

Three companies in eight days. Three different framings of the same architectural decision. The pattern your peers were predicting Friday is now the C-suite default for 2026.

Three CEOs in eight days framed layoffs as capex reallocation. The next earnings call defends your framing — or reveals you have not picked one.

What changed in eight days

What hardened over the weekend is not just the news. It is the LANGUAGE C-suite leaders are now expected to deploy when explaining AI-justified workforce changes. Three versions are on the table:

1. Cisco’s analytical reallocation frame. Mark Patterson (CFO): “The restructuring is not a savings-driven exercise — it’s a rapid reallocation of resources toward silicon, optics, security, and AI.” Investor-grade. Defensible at earnings. Boards reward it.

2. Meta’s line-item frame. Zuckerberg, on the record to his own workforce: layoffs as a quantified bill. The most direct version of the three. Hardest to backpedal from — but clearest signal to the market about strategic intent.

3. Microsoft’s voluntary-separation mechanism. Different verb structure: the workforce reduction is offered, not announced. Softer landing for employees; same underlying capex trade-off.

Boards reading the financial press over the weekend now have three reference points. The CHRO and CFO walking into Monday’s leadership meeting need to know which framing the company will use BEFORE the next earnings call asks them.

Three CEOs in eight days picked three different framings for the same capex decision. Which framing did your last board meeting commit your company to?

The decision shifting onto today’s agenda

1. Internal communications drift now exposes you. If your IT/operations team is talking about AI capex while your HR team is talking about workforce optimization, the language gap is going to surface in the next analyst question. Cisco, Meta, and Microsoft each closed the gap publicly. Yours will close in front of an analyst whether you choose to or not.

2. The AI-fluent talent your company most needs is reading the same headlines. Randstad’s 23%-have-walked figure is what these announcements look like at scale at the workforce-perception layer.⁴ When the three loudest tech CEOs in eight days all frame their cuts as capex-driven, the AI-fluent professionals in your organization read the signal correctly.

The C-suite that picked its framing is in command this quarter. The one that did not will get a framing assigned by an analyst.

Three actions for today’s lunch and this week

1. Pick your framing before earnings. Calendar a 30-minute meeting this week with CFO + CHRO + IR. Choose deliberately among the three available framings (reallocation / line-item / voluntary separation). Document the choice. Brief the executive team. The framing IS the strategy in market language right now.

2. Confirm your AI capex bill is board-visible. If your AI investment is rising significantly year-over-year and the board is not seeing it broken out by category (infrastructure + workforce capability + governance), schedule a board check-in this week to surface the breakdown. Boards are now expected to know what is on the AI capex slide — not just the total.

3. Read tomorrow’s Call for the operational architecture underneath this week’s headlines. The 33-Point Gap publishes Tuesday at 7:30 AM ET — BCG’s data on what AI Trailblazers put on the people-side of the AI bill that Pragmatists don’t. The prescriptive framework underneath today’s news cycle.

Three CEOs did the analytical work in eight days. The remaining question is which version of the framing your company is in market with by Friday.

Read: The 33-Point Gap. Tomorrow 5/19 at 7:30 am ET.

📅 Book a complementary 1:1 Strategy Session — 45 minutes to start that conversation about your AI transformation sequence.

📬 Free preview ending soon. Subscribe to continue getting decision-grade AI intelligence that prepares you to move before your competitors do. First 100 subscribers receive bonus content.

Subscribe to The AI Playbook

Sources

1. The Next Web (2026, May 18). Zuckerberg tells Meta employees the layoffs are about capex, not AI productivity. thenextweb.com

2. TechRadar (2026, May 14). Cisco cuts 4,000 jobs even as AI orders surge. techradar.com

3. CNN Business (2026, April 24). Microsoft to offer voluntary retirement to thousands of US employees. cnn.com

4. Randstad Digital (2026, May 12). The AI Capability Gap. prnewswire.com

5. CognivaLab (2026, May 15). The Call–AI Playbook: Weekend Watch. cognivalab.blog

6. CNBC (2026, May 18). Meta layoffs starting this week stress harsh AI reality inside Zuckerberg’s company. cnbc.com

1. Cisco’s “non-savings” cut hardens the AI-restructure playbook

Why this weekend: Cisco notified roughly 4,000 employees Thursday on the same day it posted record $15.8B quarterly revenue and booked $5.3B in AI infrastructure orders year-to-date. Its CFO explicitly framed the cut as reallocation toward silicon, optics, security, and AI — not savings. The Street rewarded it with a 17% jump in extended trading. By Monday open, every C-suite leader will face the same question from boards and CHROs.

• The pattern hardening this quarter: record revenue paired with an AI-pivot restructure, rewarded by the market. PayPal, Cloudflare, Coinbase, and Meta have run versions of this in the past three weeks.

• Meta’s 8,000-person cut begins Wednesday, May 20. The “rapid reallocation, not savings” frame is the language peer CEOs will adopt to defend their own cuts this quarter.

• The decision shifting onto Monday agendas: whether to communicate your own AI workforce plan ahead of the next earnings call, or be asked about it from outside the room.

“The restructuring is not a savings-driven exercise — it’s a rapid reallocation of resources toward silicon, optics, security, and AI.” — Mark Patterson, CFO, Cisco

What to watch: Whether a second Q3 reporter joins the “record revenue + restructure” pattern over the weekend, and whether Sunday analyst notes promote this from a Cisco story to an industry mandate by Monday open.

_________________________________________________________________________

2. White House signaling “FDA-style” pre-release AI testing — executive order could land any day

Why this weekend: NEC Director Kevin Hassett confirmed the administration is studying an executive order to require pre-deployment safety evaluation of frontier AI models, modeled directly on FDA drug approval. The catalyst is Anthropic’s Claude Mythos and the Project Glasswing rollout. CAISI now holds pre-deployment evaluation agreements with Anthropic, OpenAI, Google DeepMind, Microsoft, and xAI as of May 5. An Executive Order (EO) landing Saturday or Sunday reshapes deployment timelines, vendor risk reviews, and procurement posture for Monday morning.

• The Trump administration explicitly opposed AI oversight at the start of the term. This signals a complete reversal in roughly eight weeks — driven by Mythos’s demonstrated ability to identify and exploit zero-day vulnerabilities autonomously.

• Even absent an EO this weekend, the CAISI framework is now the de facto pre-release standard across all five major U.S. frontier labs.

• The decision shifting onto Monday agendas: which frontier-model integrations in your stack are mid-deployment, and how a new federal evaluation gate alters delivery dates and budget timing.

“Possibly an executive order to give a clear road map to everybody about how this is going to go… so that [models] are released in the wild after they’ve been proven safe, just like an FDA drug.” — Kevin Hassett, Director, National Economic Council

What to watch: Any weekend signal from the White House, David Sacks, or Sriram Krishnan; any Truth Social post naming AI safety or model evaluation; any executive order signed with “AI security” or “pre-deployment” in its title.

📅 Book a complementary 1:1 Strategy Session—45 minutes to map your AI transformation sequence.

📬 Free preview ending soon. Subscribe to continue getting decision-grade AI intelligence that prepares you to move before your competitors do. First 100 subscribers receive bonus content. Subscribe to The AI Playbook.

Sources:

• Cisco confirms 4,000 layoffs alongside $15.8B record revenue — TechRadar, May 14, 2026

• Cisco confirms 4,000 layoffs despite strong Q3 FY2026 earnings — The Tech Portal, May 14, 2026

• Meta to cut 8,000 jobs on 20 May with more layoffs planned for second half of 2026 — The Next Web, May 2026

• White House studying AI security executive order — Federal News Network, May 2026

• Trump admin moves further into AI oversight, will test Google, Microsoft and xAI models — CNBC, May 5, 2026

• New frontier of AI forces Trump’s heavy hand — Axios, May 5, 2026

The AI layoff wave is real, and it is accelerating. The Federal Reserve Bank of Atlanta and Duke University surveyed 750 CFOs and senior corporate executives in March 2026 and found that the executives reported AI-driven productivity gains averaging 1.8% in 2025. The same researchers computed the revenue-implied gain using actual revenue and employment data and found 0.6%—about one-third of what the CFOs reported. The 1.2-percentage-point wedge is the ‘productivity paradox’ the working paper documents: CFOs perceive AI gains that the revenue data has not yet confirmed. The National Bureau of Economic Research analysis built on that dataset projects roughly 502,000 AI-driven job cuts in 2026—nine times the 55,000 reported in 2025⁴.

Read that again. A nearly tenfold escalation in AI-driven job reductions in four quarters. That is the macro number my CFO clients are quietly modeling against next year’s headcount line.

On the flip side, as I cited on The Translation Layer Call a couple of weeks back, Forrester is forecasting that half of these AI-attributed reductions will be reversed within 12 to 18 months of announcement—rehired offshore, on contract, or at lower wages, with Gartner putting the cohort at 50% by 2027¹³. Your board read the layoff headlines this week. Your CFO is already modeling the severance line for next quarter’s earnings call. This Call is what to do before the rehires arrive—or how to skip the disruptive and costly cycle altogether.

Why the layoff arrives before the architecture

Eliminating 14% of headcount does not produce AI capacity. Eliminating 25% does not redesign a workflow. The layoffs arriving across tech this month are happening before the work has been re-architected to absorb them—and that sequence is the structural problem fueling the AI layoff wave.

As I cited in The Translation Layer Call¹⁵, Deloitte reported in their State of AI in the Enterprise 2026 that 84% of organizations have not redesigned their workflows around AI capabilities¹⁴. McKinsey’s State of AI 2025 survey of 1,993 organizations across 105 countries puts the same finding from the inverse direction: only 21% of organizations have redesigned even some workflows; nearly 80% are layering AI on top of processes designed for humans⁵. Two studies, two methodologies, the same gap. These findings point to a systematic failure to re-architect operations to leverage the transformative power of AI. Layering AI atop existing systems chokes the substrate the integration needs to maximize ROI. When layoffs remove the remaining substrate, AI returns essentially evaporate.

Workflow redesign is what produces the capacity that lets a smaller team carry the same load. Without it, the work done by the people who were laid off does not disappear—it gets redistributed across the remaining teammates, who are now also expected to operate the AI tools, manage the agents, and own the integration risk. Three months in, productivity is flat or down, the AI deployment is parked because no one owns the workflow change, and the same operating leaders who were signing the severance letters are quietly signing recruiter contracts.

This is the part the press releases skip.

The Replace-First Tax is what an executive pays when severance arrives before workflow redesign. The layoffs are real; the AI capacity to backfill is not.

The four costs your CFO can already model

The Replace-First Tax has two layers. The first layer consists of the four recoverable costs that land on the P&L within twelve months—severance, recruiting, onboarding, and the time-to-productivity drag while the rehire ramps. The second layer consists of the two irreversible costs that land on the operating model and stay there. Let’s start with the recoverable layer because it is the case the press release omits, and because it is the layer the CFO can model line by line. The two irreversible costs that follow are the ones the CFO cannot model—and it determines whether the AI integration the layoffs were meant to fund actually delivers a return.

Let’s start with the recoverable costs layer:

1. Severance. The accounting line every CFO sees first. Standard packages run sixteen weeks of base pay plus two weeks per year of service for U.S. tech workers, with health coverage extending twelve to eighteen months. Meta’s May 20 round disclosed exactly this structure for 8,000 employees. The number is large, finite, and forecastable—which is why it lands so cleanly in cost-out announcements.

2. Recruiting. SHRM’s 2025 Benchmarking Report puts the average cost per hire at $5,475 for non-executive roles and $35,879 for executives—up 113% since 2017. For technical roles in tech, the all-in number runs $10,000–$20,000 per hire. When the rehires Forrester forecast begin, this line item runs concurrent to the severance line that triggered it.

3. Onboarding. SHRM’s 2025 onboarding research puts the average direct cost at roughly $4,000 per new hire⁶. ATD adds $1,280 per employee in annual training and development. Both figures spike for AI-fluent roles requiring platform certifications and proprietary-tool training—the exact roles the rehire pool will need to refill.

4. Time-to-productivity drag. SHRM’s onboarding data show most new employees take six to eight months to reach full productivity; structured onboarding pulls that floor down to four to six months and unstructured environments stretch it to eight to twelve⁶. Specialized technical roles and middle-management positions extend the window to nine to twelve months; executives often need eighteen. Throughout that ramp the new hire is a cost center: salary is paid, work is partial, and the institutional context the role depends on is still being built. When roles are eliminated before workflows are redesigned, the productive-output clock starts over from zero.

The four costs are recoverable. The executive who runs the math sees a depressed P&L for twelve to eighteen months and an elevated G&A line throughout. The decision was a sequence error, not a value loss—painful, but recoverable on a predictable timeline.

If your AI strategy starts with severance, your CFO is funding the layoff and the rehire—not the AI return on investment.

The two costs that do not appear on the P&L until the AI integration stalls

The second layer’s first irreversible cost is institutional knowledge loss. Inkubit’s research estimates a 30,000-employee organization loses approximately $72 million annually in productivity from undocumented expertise leaving the building⁷—roughly $2,400 per employee per year. The hidden cost per individual senior departure runs around $430,000 above the recruiting line. For a $5 billion company those numbers are recoverable on a long horizon. For a $500 million company they are not.

Where does that knowledge go? The senior employees who held it take it with them to the next employer. The institutional context—the workflows, the relationships, the unwritten rules of which decisions cross which desks—does not transfer. It vaporizes when the laid-off employees walk out, and the rehire twelve months later cannot reconstruct it from a handoff document. The departing employee is whole. The company is not.

BCG’s Build for the Future 2025 puts the structural number on it: roughly seventy cents of every AI investment dollar depends on workforce capability—the people who hold the workflow context, not the AI model⁸. The layoff that removes those people removes the workforce capability the AI investment was funded to leverage. The integration does not stall because the technology failed; it stalls because the people who owned the translation layer¹⁵ are gone.

The second irreversible cost is the trust layer—and this is the one executives most consistently miss because it lives in employee behavior, not on the P&L. The mechanism is straightforward: AI integration ROI depends on employees voluntarily documenting their workflow context, sharing tacit knowledge, and participating in the human-AI collaboration that makes the integration profitable. When the visible message from leadership is “we will replace you,” the remaining teammates do the rational thing—they withhold the translation layer. They stop sharing the context that becomes the specification for their own replacement.

Mercer’s Global Talent Trends 2026 tracks the rising fear: employee concerns about AI-driven job loss climbed from 28% in 2024 to 40% in 2026⁹. Amy Edmondson’s HBR work on psychological safety in AI contexts confirms what the Mercer numbers imply—without the trust layer, the workflow context AI integration depends on never gets surfaced¹⁰. The integration the layoffs were meant to enable depends on the exact context-sharing behavior the layoffs just suppressed. That is not a recoverable cost. It is a foreclosed option.

The company that lays off its workflow architects loses the workflow architecture. The Translation Layer Collapse is what the rehire cannot reconstruct.

The three steps that earn the layoff the right to be called a strategy

Three things have to happen before any AI-justified workforce reduction earns the name.

1. Audit the workflow for actual capacity gain. Pick the function. Walk every step. Identify where AI takes thirty minutes off a four-hour task and where it adds twenty minutes of supervision. Net the result. Without the audit, the capacity “freed” by AI is a forecast, not a finding.

2. Redesign the roles around the new capacity. If a senior analyst now spends 25% less time on synthesis and 25% more on judgment, the role is now different. Promote it; pay it differently; measure it differently. Skipping this step produces the workflow-vacuum effect—the work does not disappear, it becomes invisible, and the remaining teammates absorb it without recognition.

3. Redeploy before you reduce. Map the redesigned roles back to the existing workforce. The people who would otherwise be laid off may already have the institutional context the redesigned work requires—they need permission, training, and a path. If after the mapping the redesigned org needs fewer people, the reduction is now defensible: it followed the architecture, and the institutional context belongs to the people who stay, not the people walking out the door.

The reversal pattern is already showing up in the cases that skipped this discipline. Salesforce eliminated roughly 4,000 customer-support roles after deploying AI; CEO Marc Benioff has said AI agents now handle about 50% of customer interactions. Independent benchmarks put LLM-based CRM agents at 58% success on single-step tasks—meaning roughly four out of every ten complex customer issues escalate or fail outright¹¹. Customer satisfaction scores dropped, complaint volume rose, and the company has been hiring contractors at lower wages since late 2025 to handle the work AI cannot resolve. Klarna ran the same play in 2024 with 700 customer-service workers, watched satisfaction collapse, and rehired humans within twelve months. The pattern is hardening, not new—and the Atlanta Fed’s 502,000-job reduction projection for 2026 means the rehire wave will be visible in the data within four quarters.

A counterargument is worth surfacing here. Coinbase’s stock did gain on its 14% workforce reduction—but it gained on a reduction sequenced after named org-design changes: five layers below the CEO, player-coaches replacing pure managers, AI-native pods. The market priced the org redesign and accepted the headcount reduction as the consequence of the redesign, not the strategy. A workforce reduction announced before the AI role redesign that earned it gets re-rated against the next earnings disappointment, on the schedule the Atlanta Fed and Forrester are both forecasting.

Layoffs that follow AI role redesign compound the savings. Layoffs that precede it return on next year’s recruiting budget.

The decision your CFO will ask for at the next board meeting

This is what go-slow-to-go-fast looks like in the AI-layoff moment. Slow on the layoff. Fast on the redesign. Helen Poitevin, a Gartner VP analyst, named the pattern from inside the institutional research seat last week: “Workforce reductions may create budget room, but they do not create return. Organizations that improve ROI are not those that eliminate the need for people, but those that amplify them.”¹² The Atlanta Fed data backs the analyst call. Your CFO will see both numbers before the next earnings cycle—and your board is already asking which side of the projection you intend to be on.

When your CFO asks for a cost-out story and your board asks for an AI ROI narrative, the question is not whether you reduce headcount. The question is whether you have done the architectural work that makes a reduction defensible eighteen months later—and whether the people who can build the translation layer are still in the building when the AI integration needs them.

You can rehire the headcount. You cannot rehire the context that walked out with it.

The AI Leadership Playbook

Strategic Questions

Q1. Which workflows have we actually redesigned to absorb AI capacity, and what would the audit show if we ran it next week?

Q2. If a quarter of our headcount is targeted for AI-driven reduction, what does our redeployment map look like—and which of those people are the keepers of context we cannot afford to lose?

Q3. When our CFO asks for a cost-out story, are we able to show the AI role redesign behind the number, or are we putting severance on the slide?

📅 Book a complementary 1:1 Strategy Session—45 minutes to start that conversation about your AI transformation sequence.

Your Next Plays

P1. Run the workflow audit before the cost-out conversation. Pick one function under cost-out pressure. Have the operating leader walk every step of two flagship workflows; mark where AI delivers actual minutes saved and where it adds supervision overhead. Net the result. The number is your defensible capacity gain—not the headcount forecast your finance team is building from a vendor demo.

P2. Build the redeployment map before the reduction list. For every role being considered for reduction, identify the redesigned role the same person could fill with structured AI training. The map is the hedge against the rehire-at-lower-pay reversal Forrester is forecasting—and the only protection against the Translation Layer Collapse the layoffs would otherwise trigger.

P3. Establish a quarterly AI workflow audit cadence. Pick the three functions under the heaviest AI-investment pressure. Have the operating leaders rerun P1 every quarter and track the delta between forecast capacity gain and observed capacity gain. The audit cadence is what makes AI ROI accountable to the board—because the board now has a recurring, auditable number to compare against the cost-out story your CFO told them last quarter.

📬 Free preview ending soon. Subscribe to continue getting decision-grade AI intelligence that prepares you to move before your competitors do. First 100 subscribers receive bonus content.

Subscribe to The AI Playbook. Free preview ending soon. First 100 subscribers will receive bonus content.

Sources

1. Sigalos, MacKenzie. “Coinbase cuts headcount by 14% citing AI acceleration.” CNBC, May 5, 2026. https://www.cnbc.com/2026/05/05/coinbase-cuts-headcount-by-14percent-citing-ai-acceleration-the-shares-are-gaining.html

2. Brown, Hayden. “A Message from Hayden Brown, Upwork CEO.” Upwork press release, May 7, 2026. https://www.upwork.com/press/releases/upwork-ceo-hayden-brown-shared-the-following-message-with-employees-on-may-7-2026

3. “Mark Zuckerberg Just Told 8,000 Employees Their Layoffs Are a Line Item in His $145 Billion AI Bill.” 24/7 Wall St., May 8, 2026. https://247wallst.com/investing/2026/05/08/mark-zuckerberg-just-told-8000-employees-their-layoffs-are-a-line-item-in-his-145-billion-ai-bill/

4. Federal Reserve Bank of Atlanta + NBER. “Artificial Intelligence, Productivity, and the Workforce: Evidence from Corporate Executives.” Working Paper, March 2026 (750 corporate executives surveyed; 502,000 AI-driven job cuts projected for 2026). https://www.atlantafed.org/research-and-data/publications/working-papers/2026/03/25/04-artificial-intelligence-productivity-and-the-workforce-evidence-from-corporate-executives

5. McKinsey. “The state of AI 2025: How organizations are rewiring to capture value.” McKinsey QuantumBlack, 2025 (1,993 organizations surveyed across 105 countries; only 21% have redesigned even some workflows). https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai-how-organizations-are-rewiring-to-capture-value

6. SHRM. “Onboarding Best Practices: Time-to-Productivity Benchmarks.” SHRM 2025. https://www.shrm.org/topics-tools/topics/onboarding/measuring-success

7. Inkubit. “The underestimated costs of knowledge loss.” September 30, 2025. https://www.inkubit.com/en/blog/2025/09/30/die-unterschatzten-kosten-von-wissensverlust/

8. BCG. “Closing the AI Impact Gap / Build for the Future 2025: roughly 70% of AI value depends on workforce capability investment.” Boston Consulting Group, 2025. https://www.bcg.com/publications/2025/closing-the-ai-impact-gap

9. Mercer. “Global Talent Trends 2026: AI-driven job-loss concerns climb from 28% (2024) to 40% (2026).” Mercer, 2026. https://www.mercer.com/our-thinking/career/global-talent-trends/

10. Edmondson, Amy. “How to Foster Psychological Safety When AI Erodes Trust on Your Team.” Harvard Business Review, February 2026. https://hbr.org/2026/02/how-to-foster-psychological-safety-when-ai-erodes-trust-on-your-team

11. “Companies rehire workers after AI replacements fail.” The Washington Times, March 10, 2026 (Salesforce + IBM + Google + Meta reversal pattern; 58% single-step success on LLM-based CRM agents). https://www.washingtontimes.com/news/2026/mar/10/ai-layoff-reversal-companies-rehire-customer-roles-eliminated/

12. Speed, Richard. “AI layoffs backfire as cutting staff doesn’t cut it, firms warned.” The Register, May 6, 2026 (Helen Poitevin / Gartner quote; Gartner projects 50% reversal by 2027). https://www.theregister.com/ai-and-ml/2026/05/06/ai-layoffs-backfire-as-cutting-staff-doesnt-cut-it-firms-warned/5230631

13. Forrester. “Predictions 2026: The Future of Work”—half of AI-attributed layoffs reversed, rehired offshore/contract/lower wages within 12-18 months of announcement (cited and acknowledged from prior CognivaLab Translation Layer Call). https://www.theregister.com/2025/10/29/forrester_ai_rehiring/

14. Deloitte. “State of AI in the Enterprise 2026”—84% of organizations have not redesigned workflows around AI capabilities. Cited and acknowledged from prior CognivaLab Translation Layer Call. https://www.deloitte.com/global/en/issues/generative-ai/state-of-ai-in-enterprise.html

15. CognivaLab. “The Translation Layer.” The AI Playbook, April 28, 2026 (canonical Translation Layer Call—foundation for the Translation Layer Collapse concept named in this Call).

16. CognivaLab. “The Speed Trap.” The AI Playbook, April 21, 2026 (canonical Go Slow to Go Fast Call—Lane 3 reinforcement named in this Call’s pivot).

The AI Playbook: The Weekly Call

The Weekly Call on AI transformation. Decision-grade intelligence for executives — one argument, one aphoristic line, plus The Playbook to forward.

By paola.sanmiguel

But it’s not the news you should be focused on.

Here’s the news. Every dollar your company invests in AI splits roughly 30/70: about thirty cents on technology and infrastructure, seventy cents on people, organization, and process. That’s BCG’s research, and it has held across every major enterprise AI study for two years.⁴ The new joint venture is engineered to capture the thirty cents brilliantly—embedded engineers, Anthropic’s research team alongside, vertical specialization, scale. What it is not engineered for is the seventy. That part is your job, and right now most companies do not have a role responsible for it.

What just changed in your procurement

For a generation, an enterprise software engagement assumed three vendors: a platform vendor for the tools, a systems integrator to deploy them, and a management consultant to translate strategy across the gap. Microsoft, AWS, Salesforce, and Oracle sold the platforms. Accenture, Deloitte, and the Big Four implemented them. McKinsey, BCG, and Bain handled the strategy translation. Three vendors, three contracts, three sets of incentives. Each had a specific job, and each had to coordinate with the other two for the engagement to work.

That geometry just collapsed. After this week, Anthropic owns both the model and the delivery. Your AI procurement RFP next quarter will, increasingly, be a single-vendor decision rather than a three-vendor coordination problem. That sounds like simplification—fewer contracts, fewer SOWs, fewer steering committees. In practice, it means the workforce-impact assessment that used to live with the consultant or the integrator now sits inside a delivery proposal optimized for the platform’s economics.

A counterargument is worth surfacing here. Most procurement teams already include workforce-impact assessment in their AI RFPs. The issue is sequence. When platform vendor, integrator, and consultant collapse into one, that assessment runs against a delivery proposal already optimized for the vendor’s economics—not for the workforce that will run the platform. The assessment becomes ratification, not architecture. Ratification is what lets vendor economics quietly shape the workforce decisions you thought you were making.

Every AI dollar splits 30/70—about thirty cents on technology and infrastructure, seventy cents on people, organization, and process. The seventy is your decision.

The 70 cents at risk

Look at what happens when the seventy cents goes uncaptured.

Deloitte’s 2026 State of AI shows workforce AI access has roughly doubled to about 60% in a year—but only 25% of leaders report the impact has been transformative.⁵ The 35-point gap between access and transformation is, in plain terms, what happens when the tools land but no one has redesigned the work around them. Tools deployed, value not converted. That gap is the seventy cents that disappeared.

Gartner’s April 2026 data names the operational symptom: only 28% of AI infrastructure projects deliver promised ROI, and 38% of those failures trace specifically to skill gaps.⁶ Skill gaps are not training shortfalls; they are workflow shortfalls. The right team, given the wrong workflow, fails. WRITER’s 2026 enterprise study finds 54% of C-suite leaders say AI adoption is “tearing their company apart.”⁷ When you read that headline, you might think: people don’t like change. The data says something more profound. Half of the C-suite is watching their teams work inside workflows the AI redesigned without anyone redesigning the work around the people who do it. That is the seventy cents leaking out as a culture quietly coming apart at the seams.

The talent line is the third one to watch. McKinsey’s April 2026 future-of-work data is where it shows up: AI heavy users—the people unlocking the most value inside the enterprise—are 7 to 10 percentage points more likely to plan to quit in the next three to six months.⁸ The reason is consistent across the surveys. Heavy users hit a ceiling when the workforce side of the integration is under-designed. They were the ones who could have unlocked the next layer of value, but the architecture didn’t let them. They leave to find a place where it does. Replacing them costs more than building the right architecture would have.

If you have no voice in how your people work after the integration, you cannot compound the return on the investment.

What the workforce transformation architect actually does

Most enterprises do not have an internal architecture to capture the seventy cents. They have an HR function built to hire people, an IT function built to ship systems, and an executive committee that meets monthly to discuss things HR and IT have already decided. None of those bodies is structured to design how human work changes when the AI vendor and the integrator are the same firm.

Three things have to be true for the role to actually exist.

First, the role must have a name and an executive owner. Often a CHRO with a strategy background; sometimes a chief transformation officer; sometimes an outside advisor with no integration revenue at stake. The criterion is not seniority—it is whether the person can hold a workforce-design line through a vendor-led implementation. If they can be overruled by procurement velocity, the role is decorative.

Second, the workforce-design brief must come before the procurement contract, not after. Two pages. Names the work the AI is meant to change, the human roles that change with it, the retention and progression intent, the workflows that must remain legible to managers post-deployment. The vendor responds to the brief—not the other way around.

Third, vendor compensation must tie to a workforce-retention or operating-model KPI, not just technical milestones. The platform vendor is paid for usage. The integrator is paid for integration depth. If no contract line is paid for whether your people are still in the building in three years, no one in the room is structurally responsible for it.

An executive who cannot be overruled by procurement velocity is the only person who can capture the seventy cents.

How two banks already built the role

This isn’t theoretical. Two of the largest financial institutions in the world have already built it.

Take JPMorgan Chase. Their CIO, Lori Beer, manages a 2026 technology budget of $19.8 billion and a workforce of 65,000 technologists. CEO Jamie Dimon went on record in February with what he called “huge redeployment plans” for workers whose jobs AI is changing. The bank is holding total headcount steady at about 318,500. They trimmed operations roles by 4%, support functions by 2%, and offset both by expanding client-facing and revenue-generating teams by 4%.⁹,¹⁰ People are not leaving the building. The work is being redesigned, and people are being moved to where the work now lives. That is the workforce transformation architect role in action, even if JPMorgan does not call it that.

Take BBVA. Eighteen months into a sustained AI deployment, Elena Alfaro—the bank’s head of global AI adoption—has more than 11,000 active users inside the bank, and those users have built 4,800 custom internal tools.¹¹ Harvard Business Review named BBVA a benchmark for corporate AI adoption.¹² Notice what’s interesting: the 4,800 tools were not specified by a vendor RFP. They were built by the people whose work the tools were meant to change. BBVA designed the human side of their AI environment first—competitive scarce access, a peer-driven expert network, sanctioned authority to build inside the perimeter—and that design produced the tools. Not the other way around.

These two companies built something the rest of the market will be pricing into senior-talent compensation by the end of 2027.

The architecture you build now is the retention plan you do not have to write later.

When you sit down to evaluate the next AI engagement, the question is no longer “which platform” or “which integrator.” Those are bundled. The question is who in your organization is paid to capture the seventy cents the vendor’s economics will not. If the answer is “nobody named yet,” the workforce-design brief becomes the procurement artifact you write before the RFP—not the impact assessment you sign after.

Models get updated. Engineers rotate. Workforce design becomes the operating model.

The AI Leadership Playbook

Strategic Questions

1. When we evaluate AI implementation partners this quarter, what criteria are we using to assess workforce-design implications—separately from the technical scope? What is the first change we have to make if those criteria are missing?

2. Of the three roles in our next AI engagement—platform vendor, integrator, workforce transformation architect—which one are we treating as a vendor decision, and which one are we treating as a strategy decision? Where do those two decisions get reconciled?

3. If our implementation partner has economic incentives to expand model usage, who is responsible for representing the long-term workforce interest, and is that person in the room when scope is defined? If they are not, what is the procurement decision that puts them there?

Your Next Plays

• Designate the workforce transformation architect role before the procurement decision. The criterion is not seniority—it is whether the person can hold the workforce-design line through a vendor-led implementation. Often a CHRO with a strategy background, sometimes a chief transformation officer, sometimes an outside advisor with no integration revenue at stake.

• Separate workforce design from technical scope of work in the RFP. Two documents, two sign-off paths, two distinct review milestones. Bundling them is what lets vendor economics silently shape workforce decisions.

• Tie a portion of vendor compensation to a workforce-retention or operating-model KPI—not just technical milestones. The platform vendor is paid for usage. The integrator is paid for integration depth. If no contract line is paid for whether your people are still there, no one in the room is structurally responsible for it.

Test out these plays and let us know in the comments what you learned and how you improved on our plays for your organization.

📅 Book a complementary 1:1 Strategy Session—45 minutes to map your AI transformation sequence.

STRATEGIC INFLECTION WEEK · READING ORDER

📬 Free preview ending soon. Subscribe to continue getting decision-grade AI intelligence that prepares you to move before your competitors do. First 100 subscribers receive bonus content. Subscribe to The AI Playbook.

Sources

1. Blackstone press release, “Anthropic Partners with Blackstone, Hellman & Friedman, and Goldman Sachs to Launch Enterprise AI Services Firm,” May 2026. https://www.blackstone.com/news/press/anthropic-partners-with-blackstone-hellman-friedman-and-goldman-sachs-to-launch-enterprise-ai-services-firm/

2. Bloomberg, “Goldman, Blackstone Partner With Anthropic on AI Services Firm,” May 4, 2026. https://www.bloomberg.com/news/articles/2026-05-04/goldman-blackstone-partner-with-anthropic-on-ai-services-firm

3. TechCrunch, “Anthropic and OpenAI are both launching joint ventures for enterprise AI services,” May 4, 2026. https://techcrunch.com/2026/05/04/anthropic-and-openai-are-both-launching-joint-ventures-for-enterprise-ai-services/

4. BCG, “Reinvention of the CHRO in an AI-Driven Enterprise,” February 2026. https://www.bcg.com/publications/2026/reinvention-of-the-chro-in-an-ai-driven-enterprise

5. Deloitte, “State of AI in the Enterprise 2026.” https://www.deloitte.com/us/en/what-we-do/capabilities/applied-artificial-intelligence/content/state-of-ai-in-the-enterprise.html

6. Gartner press release, “AI Projects in Infrastructure and Operations Stall Ahead of Meaningful ROI Returns,” April 7, 2026. https://www.gartner.com/en/newsroom/press-releases/2026-04-07-gartner-says-artificial-intelligence-projects-in-infrastructure-and-operations-stall-ahead-of-meaningful-roi-returns

7. WRITER, “Enterprise AI Adoption 2026.” https://writer.com/blog/enterprise-ai-adoption-2026/

8. McKinsey, “How AI is—and isn’t—changing the future of work,” April 6, 2026. https://www.mckinsey.com/capabilities/people-and-organizational-performance/our-insights/how-ai-is-and-isnt-changing-the-future-of-work

9. Fortune, “How JPMorgan’s CIO Is Reshaping Work at the Bank with a $19.8 Billion Annual Tech and AI Budget,” April 29, 2026. https://fortune.com/2026/04/29/capcom-virgin-voyages-bet-on-ai-to-reshape-gaming-and-cruise-travel/

10. CNBC, “Jamie Dimon Says AI Is Already Reshaping JPMorgan Chase’s Workforce as Bank Plans ‘Huge Redeployment,’” February 24, 2026. https://www.cnbc.com/2026/02/24/jpm-ceo-jamie-dimon-ai-reshaping-workforce-redeployment.html

11. Elena Alfaro et al., “The Hidden Demand for AI Inside Your Company,” Harvard Business Review, April 14, 2026. https://hbr.org/2026/04/the-hidden-demand-for-ai-inside-your-company

12. BBVA, “Harvard Business Review Recognizes BBVA as a Benchmark for Corporate AI Adoption,” April 2026. https://www.bbva.com/en/innovation/harvard-business-review-recognizes-bbva-as-a-benchmark-for-corporate-ai-adoption/

I call this cohort the Translation Layer. It is the human work in the middle of an org that turns AI tools into changed workflow, changed capability, and changed output. Without it, the tools sit on top of the same routines that ran the company before the tools arrived. With it, the tools compound.

Most companies are funding that layer at zero — and reporting flat AI ROI to their boards.

AI without the Translation Layer is a Ferrari engine bolted to a 1990s chassis. The horsepower is real. The platform can’t carry it.

Where the premium actually sits

Start with Bain. Their 2025 read on AI-in-production puts the productivity gain at 10–15% when companies deploy the tools alone. The figure jumps to 25–30% when companies pair the tools with end-to-end workflow redesign.^¹ Same tools. Same vendors. Same models. Double the output.

The differentiator is what most boards are not funding: the human work that sits between the tool and the team. Redesigning who does what. How decisions get made. How the team’s day actually changes when AI shows up inside the workflow. That work happens at the team unit — which means it has to be done by the person who runs -that unit;In other words: the middle manager.

Now layer Prosci on top. Their 2025 best-practice benchmark for change-management investment is 10–15% of the total program budget.^² Most companies fund that line below 5%. Some fund it at zero. The premium Bain measures is exactly the work the Prosci benchmark is asking you to fund — manager development, workflow redesign, and team-unit experimentation – the line that converts a tool license into actual adoption. If you cut that line, you cut the premium with it.

Here is how a CFO should read that pairing: every dollar redirected from tooling to translation captures roughly three dollars of incremental productivity. Not as a precise multiplier — as a directional one. The math is consistent across the dataset. The line item is consistent across the companies that miss it.

The cohort that captures the premium

Then Gallup. Their 2025 global engagement study landed two findings worth sitting with.^³

First: manager engagement is at a multi-year low. The cohort responsible for translating strategy into team routine is the one most disengaged with their own work. Female managers fell another seven points further still. Whatever the AI rollout asks of this cohort, the cohort is showing up to the ask running a deficit.

Second — and this is the one to mark: where the manager actively supports the AI rollout, employees are 8.7 times more likely to report that AI changed how much work gets done. Not 8.7 percent. 8.7 times. Whatever the AI premium looks like in the data, this cohort is the gate.

Sit with that arithmetic for a moment. Bain says the workflow redesign is where the premium lives. Prosci says the redesign requires a real change-management line item. Gallup says the manager is the cohort that converts the line item into actual adoption. The three sources point at one person. That person, in most companies right now, is under-resourced, under-engaged, and on the list of cuts.

The manager is not a soft variable. The manager is the multiplier.

Develop before you delayer

The flat-org thesis is real. A serious chunk of middle management is administrative — meeting forwarder, status compiler, budget approver — and AI agents will absorb that work fast. That part of the layer should compress.

But the same layer also contains the people who can do the workflow translation. The two functions are sitting in the same headcount line, often inside the same job description. If you cut the headcount before you separate the two functions, you keep the middle-management workflow translator inside the cuts instead of engineering this critical layer out of the org chart by accident.

Bezos used to describe his job as keeping the company two sizes smaller than it should be. The Translation Layer sits inside the size you keep. It is not the bloat. It is the load-bearing wall.

The maxim: develop before you delayer. Identify the managers who can run the translation work. Move them off the administrative load. Give them the budget and the authority to redesign workflow at the team unit. Let the AI agents automate the rest. That is how you absorb the AI transition shock without losing the premium.

The Translation Layer is not the bloat. It is the load-bearing wall.

The Translation Layer premium

Buffett once described the difference between a good business and a great one as the spread between what the business earns on its capital and what its cost of capital actually is. The Translation Layer is the same spread, applied to AI.

Two companies buy the same enterprise license. Same vendor. Same seats. Same training program. Company A drops the tools into the existing workflow and reports a 12% productivity bump to its board. Company B identifies the managers already running the translation work, gives them a budget line for workflow redesign, and reports 28% productivity.

The spread between 12 and 28 is the Translation Layer premium. It is not a model upgrade. It is not a vendor swap. It is the same dollar, redirected to the cohort that knows how the work actually moves through the team.

That premium has a cohort attached to every percentage point. Once you can name the cohort, you can fund it.

📅 The leaders treating this as their Q2 reset are booking a 45-minute Strategy Session to map the redirect. → Book here.

The Playbook

Three Questions

1. What is our current AI-program budget split between tooling, training, and workflow translation? And who owns middle-management workflow redesign?

The Bain + Prosci pair predicts what the productivity number will be once you have the answer. If translation is below 10% of the program and no one owns it, the gap to the 25–30% premium is structural, not tactical. What do we change first?

2. Which of our managers have been trained on AI tools — not just given access? Is the AI productivity metric higher for those teams?

Deloitte’s 2026 enterprise read^⁴ shows access is not the bottleneck. Daily-use is. The Gallup 8.7× multiplier sits inside the manager-trained cohort. If our number is below the benchmark, the deficit is in capability, not in licenses. Who runs the capability fix?

3. If we redirected 20% of next quarter’s AI tooling budget to deploy AI tooling and workflow translation at the team-unit level, where would the first commitment go? Which team is closest to a measurable workflow win?

Pick the team where the line manager is engaged, the workflow is well-mapped, and the metric is already on the board. That is where the premium will show first. What does the first month look like?

Three Plays

Play 1 — Name the Translation Layer. Identify the managers in your company who already do the workflow-translation work. Selection criteria: they are running the teams where AI rollouts have produced something other than a flat productivity number. They are the Translation Layer. They hold the key to the productivity opportunity for every AI deployment that follows.

Play 2 — Move the line item. In your next AI program review, propose a redirect of 5–10% of the tooling budget to a Translation Layer line. Earmark it for workflow redesign, manager development, and team-unit experimentation. Name the owner.

Play 3 — Run a Translation Layer pilot. Pick one team. Give the manager the budget, the authority, and the workflow-redesign brief. Measure the productivity delta against a comparable team running the same tools without the redesign. The delta is the conversation you bring to the board.

More strategic plays in the weeks ahead. For now, road test these with your teams and tell us how it goes in the comments.

• • •

📬 Free preview ending soon. Subscribe to get decision-grade AI intelligence that prepares you to move before your competitors do. First 100 subscribers receive bonus content.

Subscribe → www.cognivalab.blog

• • •

Sources

1. Bain & Company, Technology Report 2025: AI Leaders Are Extending Their Edge — productivity gains of 10–15% with AI tools alone vs. 25–30% with AI tools paired with end-to-end workflow redesign. https://www.bain.com/insights/topics/technology-report/

2. Prosci, Best Practices in Change Management — 12th Edition — change-management investment benchmark for premium-grade adoption; organizations executing excellent change management see an 88% project-objective success rate vs. 13% for those with poor practice. https://www.prosci.com/blog/change-management-best-practices

3. Gallup, State of the Global Workplace 2025 — manager engagement at a multi-year low; female managers fell another seven points; employees are 8.7× more likely to report AI changed how much work gets done where the manager actively supports the rollout. https://www.gallup.com/workplace/349484/state-of-the-global-workplace.aspx

4. Deloitte, The State of AI in the Enterprise 2026 — workforce access to AI tools has expanded to ~60%; among workers with access, fewer than 60% use AI in their daily workflow. https://www.deloitte.com/us/en/what-we-do/capabilities/applied-artificial-intelligence/content/state-of-ai-in-the-enterprise.html

• • •

Paola Sanmiguel is the founder of CognivaLab, an AI Transformation advisory practice for executives leading AI integration without losing the human capability that compounds it. The AI Playbook lands every Tuesday at www.cognivalab.blog.

They are not wrong. PwC’s 2026 AI Performance Study surveyed 1,217 executives across 25 sectors and found that 74% of AI’s economic value is being captured by just 20% of companies. The other 80% are deploying the same tools, spending the same budgets, and getting almost nothing back. The performance gap between AI leaders and laggards is now 7.2x.¹

The instinct in most American boardrooms is to read that number and accelerate. Move faster. Deploy wider. But the data says the opposite. The 20% who are winning did not move faster. They moved differently. PwC found that 80% of any AI initiative’s value comes not from the technology itself, but from redesigning how people work with it.

Eighty percent of AI’s value lives in how people use it—not in the technology itself. Most companies are fighting over the other twenty.

What the Rest of the World Already Knows

Outside the U.S., the conversation sounds fundamentally different. Not slower—more structural.

A European Investment Bank study of more than 12,000 firms across the EU and the United States found that AI adoption increases labor productivity by roughly 4%—driven by capital deepening, not job cuts—but only when organizations make complementary investments in software, data infrastructure, and workforce training.² The productivity gain is real, but it is conditional. Skip the human investment and the gain evaporates.

Germany’s codetermination system—where works councils have legal authority to shape how AI is deployed in the workplace—is producing measurably better outcomes. A 2026 study in Work and Occupations found that German firms where workers were consulted on AI deployment reported stronger adoption and better working conditions than firms where technology was imposed from the top.³ OECD evidence confirms the pattern: workers consulted about new technology are significantly more positive about AI’s impact on their work—and positive workers adopt faster.

Singapore made workforce transformation a government mandate. Its 2026 budget created a single statutory board—one agency with legal authority over both workforce training and career services—so that AI readiness moves at the speed of policy, not bureaucracy. Every citizen gets AI readiness diagnostics; workers over forty get up to $4,000 in retraining credits. South Korea committed $960 million to a lifetime AI talent development plan.⁴ These governments are not waiting. They are building capability first.

Japan frames AI transformation through an entirely different lens. With a working-age population shrinking by 600,000 people per year, AI is not a displacement threat—it is the only plausible way to maintain economic output.⁵ Even under that urgency, Japan’s Society 5.0 framework prioritizes workforce education and collaborative AI design—not replacement.

Four countries, four different urgencies—and the same conclusion: capability first, technology second.

The Pattern Underneath

Three continents. Different regulatory traditions, different labor markets, different urgencies. And the same conclusion: the organizations producing real value from AI are the ones that invested in human capability before they scaled the technology.

McKinsey’s State of Organizations 2026 puts a ratio on it: for every dollar spent on AI technology, five dollars should go to reskilling, workflow redesign, and change management—the organizational infrastructure that makes technology produce returns.⁶ A company increasing AI infrastructure spend by 44% while the enablement budget grows 5% is not underinvesting. It is engineering its own failure.

Stanford’s 2026 AI Index confirms what that failure looks like at scale: organizational adoption has reached 88%, but the Foundation Model Transparency Index dropped 31% in a single year and documented AI safety incidents rose 55%.⁷ Speed without structure does not produce transformation. It produces expensive conformity—organizations adopting the same tools, in the same way, producing the same middling results while the governance infrastructure collapses underneath.

Where the Leverage Actually Lives

The U.S. conversation frames this as a speed problem: who can deploy AI fastest wins. But the global evidence says it is a sequence problem. The winning organizations—in Berlin, in Singapore, in the 20% PwC identified—are not moving slowly. They are moving in the right order: people first, then technology. More on each of these strategies in the weeks ahead.

Psychological safety before automation. Capability investment before tool deployment. Redeployment before replacement.

For today, here’s the crucial nugget: sequence is not a luxury of European labor law or Asian government subsidies. It is a strategic discipline available to any leader willing to resist the pressure to deploy first and figure out the human side later.

The question I keep bringing back to the executives I work with is this: are you building the capability of your people to meet the capability of your tools? Because if the answer is no, the tools are not your competitive advantage. They are your most expensive line item.

AI Leadership Playbook

Essential AI Leadership Questions

1. For every dollar we are spending on AI infrastructure, how much are we investing in the people who will use it? Is the ratio anywhere close to 5:1—enablement to infrastructure?

2. Were our teams consulted on how AI would change their workflows, or were they informed after the decision was made? What’s the adoption gap between those two groups?

3. If we sequenced capability investment before the next technology deployment—reskilling, workflow redesign, change management first—what is the first change we have to make?

AI Leaders Next Plays

1. Pull our AI infrastructure spend and our workforce enablement spend for the last two quarters. Put them side by side. Calculate the ratio and send it to me. If we’re not at 5:1 enablement-to-infrastructure, flag the gap and what it would take to close it? How long would it take?

2. Before we approve the next AI tool purchase, I want a one-page workflow redesign proposal from the requesting team. ROI is driven by changes in how people work—not just adding technology.

3. Ask each of your direct reports this week: were your teams consulted on how AI would change their work, or were they told after the fact? I want the answer—and the adoption numbers for each group.

If you are working through the question of how to sequence your AI transformation—where to invest in your people, how to redesign work so AI and your teams perform at their best together—that is exactly the conversation I help leaders navigate.

📅 Book a complementary 1:1 Strategy Session — 45 minutes to discuss your AI transformation sequence.

📬 Free preview ending soon. Subscribe to get decision-grade AI intelligence that prepares you to move before your competitors do. First 100 subscribers receive bonus content.

#LeverageAI #RedeployBeforeReplace #GoSlowToGoFast #AITransformation #FutureOfWork #EnterpriseAI

Sources:

¹ PwC. “2026 AI Performance Study.” PwC Global, April 2026. pwc.com/gx/en/issues/technology/ai-performance.html

² European Investment Bank. “AI Adoption, Productivity and Employment: Evidence from European Firms.” EIB Working Paper 2026/02, January 2026. eib.org/en/publications/20250383-economics-working-paper-2026-02

³ Doellgast, V., Kämpf, T. & Langes, B. “Building Worker Voice and Power in AI Decisions: Three Cases in the German ICT Industry.” Work and Occupations, 2026. doi.org/10.1177/07308884251412886

⁴ Singapore Budget 2026, Ministry of Manpower / SkillsFuture Singapore, February 2026; Republic of Korea, “AI Talent Development Plan for All,” 2025. mycareersfuture.gov.sg/budget-2026

⁵ Tech for Impact Summit. “The Future of Work: Human Talent, AI Agents, or Post-Work Society?” 2026. tech4impactsummit.com

⁶ McKinsey & Company. “The State of Organizations 2026: Three Tectonic Forces.” McKinsey, March 2026. mckinsey.com/the-state-of-organizations

⁷ Stanford Institute for Human-Centered Artificial Intelligence. “The 2026 AI Index Report.” Stanford HAI, April 2026. hai.stanford.edu/ai-index/2026-ai-index-report

Subscribe now