The Trap Is Also the Business Model

Why seeing the AI productivity race clearly does not make opting out rational

June 28, 2026Read on Substack

The most interesting reaction to my last essay on AI was not disagreement. It was recognition followed by resignation.

People already understood the basic problem. Faster tools do not automatically create freer days. A task that takes twenty minutes instead of an hour can become three tasks rather than forty minutes returned to the person who completed it. More drafts create more drafts to review. More options create more choices. A faster answer can quietly become the expected answer.

None of that seemed particularly controversial. What interested me was how often the recognition ended there. The client expects the response. The founder has a runway. The employee does not want to look slower than the colleague using the tools more aggressively. The consultant knows that if someone else can turn something around overnight, “three days” begins to sound different.

The trap is not difficult to see. It is difficult to leave when everyone else’s productivity helps set the price of your restraint.


The Gain Does Not Stay Private

A productivity gain begins as an advantage and can end as a baseline.

If one analyst can produce the first draft of a report before lunch, that may be genuine leverage. If everyone can, the market does not continue paying for yesterday’s production time out of gratitude. Estimates change. Staffing assumptions change. Client expectations change. What was unusually fast becomes ordinary.

This is not unique to AI. Email, spreadsheets, search engines and mobile connectivity all changed the amount of work people could reasonably complete. Generative AI is unusual because it can lower the cost of many different cognitive tasks at once.

The result is a ratchet. Saved time may initially belong to the person who created it, but once the gain becomes visible it starts influencing what other people consider possible. The manager sees higher throughput. The client sees faster turnaround. The competitor offers more revisions. The employee takes on another account. The founder adds another experiment because the experiment now looks cheap enough to run.

None of those decisions is automatically wrong. The problem is that the reference point moves.

A person can decide that a slower process would produce better work and still be evaluated against people using the faster one. A firm can prefer more considered delivery and still compete against firms promising tomorrow.


The Productivity Effect Is Not One Thing

The evidence already makes simple claims about AI productivity difficult to defend.

In a large field study of customer-support agents, Erik Brynjolfsson, Danielle Li and Lindsey Raymond found that access to a generative AI assistant increased issues resolved per hour by 14 percent on average, with much larger gains among less experienced and lower-skilled workers. The effect was small for the most experienced workers.

A very different experiment produced the opposite result. METR studied experienced open-source developers working in repositories they knew well and found that the early-2025 AI tools available to them increased completion time by 19 percent. The developers expected AI to make them faster before the study and still believed it had made them faster afterward. METR is explicit that this result should not be generalized to software development as a whole.

Neither result cancels the other. They show why “AI makes people more productive” is too coarse a sentence. Productive at what, for whom, measured where, with which tool, and including which downstream work?

That last part matters because production is only one stage of knowledge work. A system that makes the first stage dramatically cheaper can increase the amount arriving at every stage that follows.


Production Gets Cheap Before Judgment Does

AI can draft the document, but someone still has to decide whether it is true, useful, appropriate, legally safe, strategically aligned and worth sending. It can summarize the meeting, but someone still has to know what mattered. It can generate ten approaches, but someone still has to choose. It can produce code quickly, but someone remains responsible for understanding what changed and what might break.

Human-factors researchers have described this as one of the recurring ironies of automation. As production becomes easier, the human role can shift toward evaluation. Generative AI adds its own versions of that problem: workflow disruption, interruptions, and a tendency to make some easy tasks easier while leaving the hardest judgment concentrated with the person.

That is not necessarily a bad trade. In many settings, evaluation is exactly where human effort should move. I would rather automate repetitive production than preserve it merely because humans used to do it. But evaluation is still work. It requires context, standards, attention and responsibility, and a model can often generate possibilities much faster than a person can responsibly evaluate them.

A request for one option becomes five because five are cheap. A report gains three extra analyses because they are easy to produce. A founder asks for another version because another version no longer requires another afternoon. Each addition looks small from the production side.

The review burden accumulates on the other side.


Cheap Production Can Move Work Downstream

There is another possibility that the language of personal productivity tends to miss. The time does not disappear or return to the sender. It moves.

Someone generates the document in fifteen minutes. Someone else spends an hour working out what it means. Someone produces five strategic options. A manager now has five options to evaluate. Someone forwards a polished AI summary without checking the source. The recipient inherits the verification work. The sender may genuinely have saved time while the organization lost it.

Recent survey research from BetterUp Labs and Stanford Social Media Lab has given one version of this problem a useful name: “workslop,” AI-generated work that looks finished but transfers unfinished thinking to the recipient. In their September 2025 U.S. survey, 40 percent of desk workers reported receiving workslop in the previous month, and respondents estimated that resolving an incident took about two hours on average.

That is survey evidence, not a universal measure of workplace AI. But it illustrates a broader problem: production cost and organizational cost are not the same thing. If one team cuts drafting time while another quietly spends more time correcting, clarifying or recreating what arrives, a local productivity metric can improve while the broader workflow gets worse.


Marketing Taught Me This Before AI Did

I have seen a version of this problem for years in B2B marketing. Most good marketers already know that customer understanding, positioning, trust, brand memory and timing matter. The work that creates durable demand is often slower, more contextual and harder to defend inside a short reporting window.

At the same time, marketing teams have to justify budgets. Pipeline, attribution, conversion rates, MQLs and quarterly performance are easier to place in a dashboard than the quality of a market understanding that prevented a bad campaign from being launched. The reporting system influences which work can be defended, even when everyone in the room knows that the reporting system is incomplete.

Once a proxy becomes the language through which value is defended, work begins organizing around what the proxy can see.

AI adoption can inherit the same problem. Generated output is easy to count. Response time is easy to count. Number of experiments, drafts, campaigns, tickets or completed tasks can all move visibly. Judgment that prevented unnecessary work is much harder to display. So is the decision not to generate another version, not to launch another channel, or not to automate a process whose complexity is already unnecessary.

If visible activity becomes the primary proof of AI value, the technology can look most successful precisely where it creates the most activity.


Founders Have a Selection Problem

Founders face an especially sharp version because possibility is part of the job. They are supposed to see openings before other people do, move around constraints and keep finding ways to create leverage when resources are limited.

AI makes many possible actions cheaper. Another landing page. Another outbound sequence. Another campaign. Another market analysis. Another prototype. Another content channel. Another product variation. Another automation. Another dataset to examine.

Each can be rational in isolation. The portfolio can still become irrational.

This changes the founder bottleneck in an important way. The question is no longer only whether the company has enough people or time to produce something. Increasingly, the harder question is whether the possibility deserves to enter the company at all. Production becomes cheaper while selection becomes more important.

That is harder than it sounds because tools make possibilities visible. Once something can be done quickly, choosing not to do it can feel like leaving leverage unused. The discipline required is not mainly the discipline to work less. It is the judgment to distinguish between a cheap experiment and an experiment worth creating organizational surface area around.

More capability therefore does not remove scarcity. It can relocate scarcity into attention, selection, review and coordination.


The Professional Cost Of Restraint

This is why “just slow down” is not a serious answer. A person can prefer a slower equilibrium and still be unable to create one individually.

A consultant may genuinely believe that a client would benefit from more thinking time while knowing another firm promises tomorrow. An employee may dislike the expectation of instant replies while knowing that silence is interpreted relative to colleagues who answer immediately. A creator can believe the publishing cadence is absurd and still watch distribution fall when they stop feeding it. A manager may want fewer drafts while also fearing that refusing available output will make the team look less productive.

The expensive behavior is not necessarily irrational. It can be locally rational and collectively costly.

This is the part that recognition does not solve. The individual is being asked to absorb the private cost of restraint while much of the benefit would be shared by other people. The consultant takes longer and risks the client. The employee answers less and risks appearing less available. The company promises more realistic turnaround and risks losing to the company that promises speed.

Once enough people face that structure, the problem starts to look like coordination. Almost everyone may prefer fewer useless messages, fewer unnecessary drafts, more considered decisions and less pressure to signal constant activity. But the first person to reduce visible output can look less productive before anyone has agreed that less output is the goal.

No conspiracy is required. The incentives are enough.


The Trap Is Also The Business Model

This is what I mean by the title. I do not mean that every company literally earns revenue from overwork or that every AI vendor depends on making people compulsive. I mean that the economic logic surrounding the work can reward the same acceleration people privately recognize as costly.

Clients reward responsiveness. Markets reward lower prices and shorter turnaround. Managers reward throughput. Investors reward growth. Employees protect relevance. Founders protect runway. Creators protect distribution. Each pressure can be reasonable on its own.

Put them together and the system can keep accelerating without anyone explicitly choosing an accelerated system.

AI intensifies this because it changes what can be produced with the same visible resources. Once a gain becomes repeatable, it becomes available to planning. Once it becomes available to planning, it can become expected. The temporary advantage then disappears into the baseline and the search for the next advantage begins.

This is why the same productivity gain can feel liberating at first and compulsory later. The technology did not reverse direction. The competitive environment absorbed the gain.


What Productivity Should Count

If AI makes output cheaper, organizations need to become more careful about what they count as productivity. More documents are not necessarily more work completed. More experiments are not necessarily more learning. Faster responses are not necessarily better service. More generated options are not necessarily better decisions.

The accounting has to include what happens after generation. How much review was required? How much rework? How many options were produced but never useful? How often did someone else have to reconstruct the context? Did the faster workflow reduce the time to a good decision, or merely increase the amount of material surrounding it?

This does not require inventing one perfect metric. It requires refusing a convenient fiction: that the visible saving at the point of production is automatically a saving for the system around it.

The same question should apply to AI adoption programs. If a company reports that employees are producing more, it should also want to know whether people are resolving more valuable work, whether quality changed, whether downstream teams inherited additional verification, and whether the number of generated artifacts grew faster than the number of decisions or outcomes that mattered.

Otherwise AI can improve the metric used to justify AI.

That is not a technology problem alone. It is a measurement and management problem.


The Race Can Be Visible And Still Continue

The AI productivity trap is difficult precisely because it does not require anyone to be fooled.

The employee can see it. The founder can see it. The manager can see it. The client may even see it. Each can still have a rational reason to ask for more.

Faster production lowers the cost of another draft, another experiment, another message, another analysis and another possibility. Once those gains enter competitive systems, they begin changing what other people are expected to deliver. A private efficiency gain becomes a public reference point.

That is why seeing the trap is not enough. The question is not whether people can become disciplined enough to resist useful tools. It is whether the organizations adopting them can distinguish more activity from more value, whether they can see costs transferred elsewhere in the workflow, and whether anyone has enough incentive to protect a slower choice before the market makes the faster one normal.


The trap is already visible. What remains unresolved is who can afford to stop rewarding it.See how the Human OS fits together →If this helped you name something you have been feeling but had not quite put into words, please share it with someone who might need the same language.Share


The Human OS Manual is now . If this work resonates, you can reserve your copy and follow the ideas here as they continue to develop.

Feel free to leave a comment. I read and respond to every thoughtful note.


Notes and sources

Reader signals: LinkedIn and WhatsApp reactions to the previous essay, including comments on AI FOMO, productivity pressure, and the difficulty of resisting the productivity race.

  • Erik Brynjolfsson, Danielle Li, and Lindsey R. Raymond, "Generative AI at Work" (2023; published in The Quarterly Journal of Economics, 2025).
  • Auste Simkute et al., "Ironies of Generative AI: Understanding and Mitigating Productivity Loss in Human-AI Interaction" (2024).
  • Joel Becker, Nate Rush, Beth Barnes, and David Rein, "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity" (2025).
  • BetterUp Labs and Stanford Social Media Lab, "Workslop: The Hidden Cost of AI-Generated Busywork" (2025).