The middle is the part being removed

Executive summary

• The comfortable answer is wrong. “AI does the work, humans orchestrate and review” puts people in the middle — and the middle, coordinating and checking, is the part being removed.
• Review is not a control. People catch far fewer errors once an automated aid is present (97% → 59%), fewer still once its reliability is steady (82% → 33%), and AI-assisted answers read well even when they are wrong.
• The productivity mostly isn’t showing up. 89% of executives report no impact; time the tools save comes back as time spent checking them.
• PM hiring is growing, but the org is delayering. Postings are up 27%; headcount cuts are deepest at VP (−38%) and Director (−35%), shallowest for senior ICs (−21%). Only 2% of AI PM postings are entry-level.
• What to do: write down your bets and score them, replace “review it” with checks decided in advance, learn where your domain’s frontier is, and protect the junior pipeline.

The value has moved to the two ends: deciding what is worth building, and being answerable for how it turns out.

Every product person I know has reached roughly the same comfortable conclusion. The machines will do the chores. The machines will do the heavy lifting too — the research, the synthesis, the first draft of the strategy. And we will sit in the middle, directing traffic and checking the work. Orchestrators. Conductors. The brain of the product, freed at last from the tedium.

It’s a lovely idea. It flatters us, it doesn’t require us to change much, and it has the shape of a compromise, which makes it feel reasonable.

I’ve spent a year building tools to test it, and I’ve come to think it’s wrong in a specific and expensive way. Not because AI will replace product managers — the data says clearly that it isn’t. But because “the middle” is the part of the job being removed, and “a human reviews it” is not a safety mechanism. It’s where errors get laundered into confidence.

The experiment nobody noticed

In January 2026, at a leadership retreat in Hawaii, Meta began designing something called Project OT — Organization Transformation. Reuters reported the details in August.

The plan was the orchestrator thesis, executed at scale. Teams of ten to twenty would be replaced by “AI-native pods” of three to five. Engineers, designers and product managers would fold into a generic “builder” role. Middle management would shrink sharply. The total cuts were meant to match or exceed the 2022–23 round — roughly a quarter of the company, or more. (Meta told Reuters this was scenario planning, not a finalized plan.)

On 19 May 2026, hours before the first wave of layoffs — about 10% of the company — went ahead, Zuckerberg canceled the second wave planned for November.

Three reasons are on record, and the first is the one to sit with. The productivity data didn’t support it. In Meta’s internal software platforms and infrastructure, code changes had risen 220% year over year. New or improved features reaching users had risen 36%. Major technical and security incidents had risen 40%.

Code changes+220%

User-facing features+36%

Technical incidents+40%

Source: Reuters via Engadget, 26 August 2026.

The numbers

Measure	Change, year over year
Code changes	+220%
User-facing features	+36%
Technical incidents	+40%

Read that again slowly, because it is the whole argument in three numbers. Output of work more than tripled. Output of product rose by a third. Output of defects rose faster than product did.

Be precise about where those numbers come from. Not from Project OT’s pods — the full restructuring never ran. They come from Meta’s existing organization, already pointing enormous generative capacity at its codebase. That is what AI output looks like before anyone restructures around it. You don’t get more product. You get vastly more motion, slightly more product, and meaningfully more breakage. Project OT would have bet the shape of the company on the first number.

By July, Zuckerberg was conceding at an internal town hall that AI agent development “hasn’t really accelerated in the way that we expected.” Reuters later heard the remarks. By September, Fortune reported Meta was quietly bringing managers back, including a voluntary offer for Applied AI engineers to return to management roles.

The largest planned bet on the human-as-orchestrator model was largely called off, before most of it ran, by the people with the most data about whether it would work.

That cuts both ways. Meta didn’t discover that the middle is indispensable. It discovered that you can’t remove the coordination layer by decree while the tools still produce more motion than product — somebody has to absorb the breakage, and in September that somebody was a manager again. The middle doesn’t go in one reorganization. It goes as the work it existed to do stops needing a person, which is slower, and much less announced.

Why “a human reviews it” doesn’t mean what you think

The orchestrator model has a checkpoint in it: AI does the big work, humans review. That checkpoint carries the entire integrity of the system, so it’s worth asking how well humans perform it.

We have known the answer since 1999, and it is not encouraging.

Skitka, Mosier and Burdick ran eighty people through a flight-simulation monitoring task. Participants without an automated aid caught 97% of critical events. Participants with an aid caught 59%. The aid didn’t have to be wrong to cause this — its mere presence collapsed vigilance by nearly forty points.

Worse: when the automation recommended the wrong thing, participants followed it about 65% of the time — while reliable contradicting information sat on their own instruments. One participant out of forty made no such errors at all.

You might assume expertise fixes this. It doesn’t. Parasuraman and Manzey’s 2010 review of the literature is blunt: experienced pilots show it, glass-cockpit pilots show it, experienced air traffic controllers show it. Practice doesn’t help either — sixty minutes of training showed no reduction. And explicit instructions to verify the automation, tested by Mosier and colleagues in 2001, had no effect on the strength of automation bias.

Then there’s the finding that should worry product people most. Parasuraman, Molloy and Singh, 1993: when automation was consistently reliable, operators detected only 33% of its failures. When reliability visibly fluctuated, they detected 82%.

Automated aid added
Skitka, Mosier & Burdick, 199997%59%

Reliability made steady
Parasuraman, Molloy & Singh, 199382%33%

0%25%50%75%100%

Sources: Skitka, Mosier & Burdick (1999), n=80; Parasuraman, Molloy & Singh (1993).

The numbers

Condition	Before	After
No aid → automated aid present	97% caught	59% caught
Reliability wobbled → consistently reliable	82% caught	33% caught

Reliability that wobbles keeps you awake. Reliability that’s consistently good puts you to sleep — which means the review function degrades precisely as the AI improves. The better your tools get, the worse you get at catching them.

Nor is this confined to cockpits. In June 2026, PNAS Nexus published a randomized experiment with 1,339 practicing teachers. They reviewed identical student work carrying deliberately wrong scores, labeled at random as human- or AI-generated. When the scores were unfairly harsh, the AI label left the final grades 22% further from the fair mark than identical human errors did. Over half that effect came from teachers crediting the algorithm with more ability and responsibility.

The errors that read best are the ones that are wrong

Now put that alongside the most important AI study in business research.

Dell’Acqua and colleagues at Harvard Business School ran 758 BCG consultants through real tasks with and without GPT-4, published in Organization Science this March. On tasks inside the model’s competence, results were excellent — quality up roughly 30%, and below-average performers gained 43% against 17% for the top half.

On tasks just outside the frontier, correctness fell from 84.5% to between 60 and 70%.

Here is the part that should change how you think about review. Even when AI-assisted consultants got the answer wrong, their recommendations were rated higher in quality than those of consultants working without AI.

Consultants, outside-frontier task
Dell’Acqua et al., n=75884.5%60–70%

0%25%50%75%100%

Source: Dell’Acqua et al., Organization Science, 2026. The gold band is a range, not a point.

The numbers

Measure	Without AI	With AI
Correct on the outside-frontier task	84.5%	60–70%

The failure mode isn’t garbled output you’d catch. It’s output that is more articulate, better structured and more persuasive than the truth — produced by a system whose authority you already over-weight, reviewed by a human whose vigilance that system has already dulled.

Prompt-engineering training made out-of-frontier performance worse. It taught people to trust a frontier they still couldn’t see.

This is not an argument against using AI. It’s an argument that “humans review the output” describes a hope, not a control.

The time doesn’t come back

The orchestrator thesis assumes an economics: AI absorbs the chores and frees humans for higher-order work. Test that.

Humlum and Vestergaard studied roughly 25,000 Danish workers across AI-exposed occupations. On earnings and hours, the result was null, with confidence intervals excluding effects above 2%. That held even though 19% of workers reported saving more than an hour a day. Eighty-five percent of users reallocated the time saved to other work rather than banking it. And 35% of the new tasks AI created were AI oversight.

The chores you automate come back as the chore of checking the automation.

A 2026 NBER working paper by Yotzov, Bloom and colleagues surveyed around 6,000 executives: 69% adoption, and 89% reporting no productivity impact.

DORA’s 2024 research found that a 25% increase in AI adoption correlated with a 7.2% decrease in delivery stability. Meta’s incident numbers again, at industry scale.

89%

of executives report no productivity impact at 69% adoption

NBER w34836, ~6,000 executives

−7.2%

delivery stability for every +25% of AI adoption

DORA 2024

85%

of users reallocated the time AI saved, not banked it

NBER w33777, ~25,000 workers

35%

of the new tasks AI created were checking the AI

NBER w33777

Sources: Yotzov, Bloom et al. (NBER 34836); DORA (2024); Humlum & Vestergaard (NBER 33777).

The numbers

Finding	Value	Source
Executives reporting no productivity impact	89%	NBER w34836
Change in delivery stability per +25% AI adoption	−7.2%	DORA 2024
Share of users reallocating time savings	85%	NBER w33777
Share of new tasks that were AI oversight	35%	NBER w33777

Even METR’s famous study — experienced developers 19% slower with AI while believing they were 20% faster — has been partly walked back. In February 2026 METR conceded that developers are more sped up now than its earlier estimate suggested. But the finding that survives is the one that matters here: a thirty-nine-point gap between perceived and measured productivity.

What developers believed20% faster

What was measured19% slower

−40%−20%0+20%+40%

Source: METR, arXiv 2507.09089, n=16. METR revised the estimate in February 2026; the 39-point gap is the part that holds.

The numbers

	Change in speed with AI
Believed	20% faster
Measured	19% slower

We are systematically unable to tell whether these tools help us. That is an unnerving trait in a checkpoint you rely on.

What AI actually does to a profession

Three studies, one pattern.

Dell’Acqua: bottom half +43%, top half +17%. Brynjolfsson, Li and Raymond in the Quarterly Journal of Economics, 5,172 support agents: +15% on average, with the highest-skilled workers seeing small gains in speed and small declines in quality. Noy and Zhang in Science, 453 professionals: 40% faster, quality up 0.45 standard deviations — and the correlation between a person’s performance on the first task and on the second fell from 0.49 to 0.25.

Bottom half of performers+43%

Top half of performers+17%

Task-one score predicting task-two
Noy & Zhang, Science, n=4530.490.25

00.250.500.751.00

Sources: Dell’Acqua et al. (2026); Noy & Zhang (Science).

The numbers

Measure	Value
Quality gain, bottom half	+43%
Quality gain, top half	+17%
Correlation, task one to task two, without AI	0.49
Correlation, task one to task two, with AI	0.25

That last number is the one nobody quotes and everybody should. AI decouples output quality from individual ability. After you add the tool, how well someone did last time stops predicting how well they’ll do next time.

If you are a good product manager, your execution has always shown your competence — your documents were sharper, your analysis tighter, your specs clearer. That signal is being compressed away. Not because you got worse, but because the floor came up underneath you and the ceiling didn’t move.

Which means the market will stop paying for execution quality, because it no longer distinguishes anyone. It will pay for whatever still varies.

So where do humans actually sit?

Not in the middle. At the ends.

A note on that word, because I’ve been using it for two things. There’s the middle of the workflow — the step where a human checks what the machine produced. And there’s the middle of the org chart — the people who route work, coordinate handoffs and translate between functions. They are the same role. Reviewing and coordinating both mean standing between the people who decide and the work that gets done. Both are thinning for the same reason: the tools make the handoff cheap, and the check, as we’ve seen, doesn’t hold.

Upstream

Deciding what is worth building

Choosing between forty options the model will happily generate.

The middle

Routing, coordinating, reviewing

Handoffs and checks — what the tools make cheap and delayering takes out.

Downstream

Being answerable for it

Accountability cannot be delegated to a system that cannot be punished.

Upstream: deciding what is worth building. When generating options is free, having options stops being valuable and choosing between them becomes the whole game. AI will give you forty features, twelve strategies and a hundred user segments. It has no opinion about which one your company should stake a year on. That judgment — taste, if you want the honest word — is not a coordination task and doesn’t become one.

Research in Nature Human Behaviour last year found that AI raises the quality of individual ideas while reducing the diversity of the idea pool. Everyone converges. If your competitors are all orchestrating the same models toward the same obvious answers, the differentiated bet becomes more valuable, not less.

Downstream: being accountable for the consequences. This can’t be automated, for structural rather than technical reasons. Someone has to be answerable when it goes wrong, and answerability cannot be delegated to a system that cannot be punished.

Note what the automation-bias literature found did help, when almost nothing else did: Skitka and colleagues found that making people accountable for decision accuracy reduced both omission and commission errors. Not instructions to check. Not training. Ownership of the outcome.

The middle — routing work, coordinating handoffs, translating between functions, running the process — is the part that both AI and organizational delayering are eating at once. A second Dell’Acqua study, this one with 776 people at Procter & Gamble, makes it vivid: an individual with AI matched a two-person cross-functional team without it. And the split between R&D staff proposing technical solutions and commercial staff proposing commercial ones largely disappeared.

AI dissolves functional silos, which is wonderful. It also dissolves the job of bridging functional silos, which is what much of product management has quietly been.

What the market is actually doing

Now the part that contradicts almost everything you’ve read.

PM hiring is growing. TrueUp’s tracking across 9,000+ tech companies had product management openings up 27.2% year over year as of 31 August 2026. That is faster than design at 17.4%, faster than data and analytics at 14.8%, and above the all-category average. By Mind the Product’s count, PM openings are up about 75% from their late-2022/early-2023 low. Anyone telling you the profession is being automated away is not looking at the data.

Product management+27.2%

Design+17.4%

Data & analytics+14.8%

Source: TrueUp job category trends, 9,000+ tech companies, 31 August 2026.

The numbers

Category	Openings, year over year
Product management	+27.2%
Design	+17.4%
Data & analytics	+14.8%

PM-to-engineer ratios haven’t moved. By the estimates I could find, roughly 1:9.5 — squarely inside the 1:6 to 1:10 range Marty Cagan described in 2007. Every claim of compression I could find was extrapolation with no dataset behind it.

036912

1 : 9.5 today — inside the 2007 band of 1 : 6 to 1 : 10.

Source: Cagan (2007); current ratio as cited in the essay.

The numbers

	Engineers per PM
Cagan range, 2007	6 to 10
Today	~9.5

But headcount is down 28% from peak, and the gradient is the story. VP −38%. Director −35%. Manager −31%. Senior IC −21%.

• Coordinating levels
• Individual contributors

VP−38%

Director−35%

Manager−31%

Senior IC−21%

Source: Live Data Technologies via Mind the Product. Overall: −28% from peak.

The numbers

Level	Change from peak
VP	−38%
Director	−35%
Manager	−31%
Senior IC	−21%

Those two headlines don’t conflict. Postings measure the flow of new openings, rebounding from a 2023 trough. Headcount measures the stock of people, against a 2022 peak. Hiring is recovering; the profession is still smaller than it was, and it is being rebuilt with fewer layers.

Look at where the cuts landed. They are steepest at VP and Director — the levels whose work is mostly coordinating other product people — and shallowest for senior ICs, who do the product thinking themselves. If AI were substituting for product work, the people doing that work would be going first. That is not the shape. The shape is delayering: the coordination layer is being removed, while the people who do the product thinking are the most resilient group in the profession. (Other ICs fell 30%, nearly as far as managers — but this counts people already inside. Who gets let in is its own story.)

This is the strongest evidence against standing in the middle. The middle is being taken out — not by AI, but by organizations that no longer need as many humans passing work between other humans.

The real crisis, which is not the one you’ve been worrying about

Stanford’s Digital Economy Lab, using ADP payroll data, found employment for 22-to-25-year-olds in AI-exposed occupations running 19% below trend as of June 2026. Revelio’s own tracker independently puts the gap at 19% too, and Revelio found entry-level postings in highly AI-exposed jobs down more than 40%. In software, Indeed’s data has postings at 69.3% senior and 4.5% entry-level.

For product specifically: of 12,397 AI product management postings, 2% were entry-level.

Software postings 4.5%

AI product management postings 2%

−19%

employment for 22–25-year-olds in AI-exposed jobs, against trend

Stanford Digital Economy Lab

−19%

the same measure, independently

Revelio Labs

Sources: Indeed Hiring Lab (23 July 2026); Stanford Digital Economy Lab (ADP payroll, August 2026 update); Revelio Labs (August 2026).

The numbers

Measure	Value
Entry-level share, software postings	4.5%
Senior share, software postings	69.3%
Entry-level share, AI PM postings (n=12,397)	2%
Early-career employment vs trend (Stanford)	−19%
Same, Revelio	−19%

Note what this measures that the headcount figures above don’t. Those count who was cut; this counts who gets let in. New juniors aren’t being hired.

The profession isn’t shrinking. Its entry ramp is.

One honest caveat, because the research is unsettled: that same Revelio analysis found low-exposure entry-level postings fell 33% against more than 40% for high-exposure ones. Much of this, then, is macro, not AI alone. And LinkedIn’s economists state flatly that outside clinical healthcare, “hiring patterns look the same for jobs with high AI exposure and those with low exposure.” That contradiction hasn’t been resolved.

But the direction is consistent across three independent datasets, and it should frighten anyone who wants this craft to survive. Senior judgment is built by doing junior work badly for several years under supervision. We are removing the years.

What to do about it, concretely

Move your evidence upstream. If execution quality no longer signals ability, stop trying to signal with it. What’s scarce is a defensible answer to why this and not the forty other things. Write down the bet before the outcome is known. Keep a record of what you predicted and what happened. Prediction accuracy is one of the few product signals AI cannot manufacture for you.

Stop calling review a control, and build actual controls. Telling yourself to check more carefully does not work — that’s the finding, not an opinion. What the evidence supports is accountability for outcomes, not for process. Practically: decide what would falsify an AI’s answer before you read it. Check the load-bearing claim rather than skimming the whole thing, because fluency will fool you on the whole thing. And be most suspicious of the output that reads best, since AI-assisted answers read well even where correctness collapses.

Learn where the frontier is for your domain. The jagged-frontier finding means capability is discontinuous and the boundary is invisible from the outside. The only way to map it is to use these tools on real work and get burned in specific, memorable ways. This is new expertise and it does not transfer between domains — which is why prompt-engineering training made BCG’s consultants worse.

Build something real, and let people use it. Not because building is the job, but because it’s the fastest way to find the frontier. And when everyone’s documents are excellent, a working thing is the only artifact that hasn’t been compressed.

If you manage people: protect the entry ramp, deliberately. It is individually rational to hire only seniors and give the junior work to a model. It is collectively suicidal. Someone has to keep making product people or in ten years there won’t be any, and that someone will have to eat a cost their competitors don’t.

The uncomfortable summary

The optimists are wrong that this changes everything: 89% of executives report no productivity impact, PM-to-engineer ratios are where Cagan left them in 2007, and Meta scaled back the most aggressive restructuring in the industry because the numbers didn’t hold up.

The pessimists are wrong that product management is being automated: postings are up 27% and senior ICs are the most resilient group in the profession.

What’s happening is narrower and stranger. The coordination layer is being removed. The entry ramp is closing. And the tools are compressing the difference between competent and excellent execution, which relocates all the value to the two things that never compress — knowing what’s worth doing, and being answerable for having done it.

Humans aren’t in the middle. The middle is the part that’s going away. We’re at the ends, where we probably should have been all along.

Sources

Every number above, where it came from. Each row opens the paper, report or article itself; the tag says which part of the argument it supports.

1. Meta reportedly abandoned an AI-focused restructuring planReuters, via EngadgetThe Meta caseNewsAug 2026↗︎Project OT, and the three numbers
2. The Pulse: Meta wanted to reduce teamsThe Pragmatic EngineerThe Meta caseNewsletterAug 2026↗︎Pods, the “builder” role
3. Meta’s managers come backFortune, 12 September 2026The Meta caseNewsSep 2026↗︎The management rebuild
4. Does automation bias decision-making?Skitka, Mosier & Burdick · Int. J. Human-Computer Studies 51(5)Human reviewJournal1999n=80↗︎97% → 59% caught; 65% followed wrong advice
5. Accountability and automation biasSkitka, Mosier & Burdick · Int. J. Human-Computer Studies 52(4)Human reviewJournal2000↗︎Accountability for accuracy reduced both kinds of error
6. Complacency and bias in human use of automationParasuraman & Manzey · Human Factors 52(3)Human reviewJournal2010review↗︎Experts and training don’t fix it
7. Performance consequences of automation-induced “complacency”Parasuraman, Molloy & Singh · Int. J. Aviation Psychology 3(1)Human reviewJournal1993n=40↗︎82% vs 33% of failures caught
8. Aircrews and automation bias: the advantages of teamwork?Mosier, Skitka, Dunbar & McDonnell · Int. J. Aviation Psychology 11(1)Human reviewJournal2001↗︎Prompts to verify had no effect
9. Why do experts miss AI’s errors?Goulas, Megalokonomou & Sotirakopoulos · PNAS Nexus 5(6)Human reviewJournalJun 2026n=1,339↗︎Harsh AI grades left a 22% larger fairness gap
10. Navigating the jagged technological frontierDell’Acqua et al. · Organization ScienceAI and the workJournalMar 2026n=758↗︎The frontier, coherence, +43% / +17%
11. The cybernetic teammate: a field experiment on generative AI and teamworkDell’Acqua et al. · Organization Science · Procter & GambleAI and the workJournal2025n=776↗︎One person with AI matched a two-person team
12. Generative AI at workBrynjolfsson, Li & Raymond · Quarterly Journal of Economics 140(2)AI and the workJournal2025n=5,172↗︎+15%, and small quality losses at the top
13. Experimental evidence on the productivity effects of generative AINoy & Zhang · ScienceAI and the workJournal2023n=453↗︎Ability stops predicting output: 0.49 → 0.25
14. ChatGPT decreases idea diversity in brainstormingMeincke, Nave & Terwiesch · Nature Human Behaviour 9AI and the workJournal2025↗︎Better ideas, fewer different ones
15. Still waters, rapid currents: early labor market transformation under generative AIHumlum & Vestergaard · NBER w33777ProductivityWorking paper2025~25,000 workers↗︎A null on earnings and hours
16. Firm data on AIYotzov, Bloom et al. · NBER w34836ProductivityWorking paper2026~6,000↗︎89% report no productivity impact
17. Accelerate State of DevOpsDORAProductivityReport2024↗︎−7.2% delivery stability
18. Measuring the impact of early-2025 AI on experienced developersMETR · arXiv 2507.09089ProductivityPreprint2025n=16↗︎Believed +20%, measured −19%
19. Job category trendsTrueUpLabor marketDataAug 20269,000+ companies↗︎PM openings +27.2%
20. What the data tells us about tech and PM lay-offs in 2026Live Data Technologies via Mind the ProductLabor marketArticle2026↗︎Headcount by level
21. Canaries in the coal mineStanford Digital Economy Lab · ADP payrollLabor marketWorking paperJun 2026↗︎−19% early-career employment
22. Is AI responsible for the rise in entry-level unemployment?Revelio LabsLabor marketReportAug 2025↗︎Entry-level postings −40% vs −33%
23. The labor market is tilting toward seniorityIndeed Hiring LabLabor marketReportJul 2026↗︎69.3% senior, 4.5% entry
24. AI has already added 1.3 million new jobsLinkedIn via World Economic ForumLabor marketNewsJan 2026counterpoint↗︎Hiring looks the same at high and low exposure
25. Revelio Labs Reports 36.5K US Jobs Added in August, Employment in AI-Exposed Jobs 19% Lower for Workers Under 25Revelio Labs · PR NewswireLabor marketNewsSep 202622–25 employment gap in AI-exposed jobs: 19%↗︎Revelio's independent 19% early-career gap
26. AI Product Management Jobs in 2026: What 12,400 Postings RevealAxial SearchLabor marketDocs202612,397 US AI PM postings since Jan 2026; ~2% entry-level↗︎2% of AI PM postings are entry-level