AI multiplies output, not human capacity

AI does not remove human effort. It redistributes and concentrates it.

There is a tempting equation behind the way we think about AI productivity. If one person works for eight hours and can run four agents simultaneously, then four agents multiplied by eight hours should produce thirty-two hours of work.

The arithmetic may describe production capacity. It does not describe completed work.

Generated code, documentation, analysis and tests are candidate work until someone has decided that they are correct, coherent and worth keeping. In accountable work, AI shifts effort from production towards specification, verification, integration and maintenance. The amount of scrutiny should vary with the consequences of being wrong, but it does not disappear merely because production became easier.

AI can therefore increase the rate of production much faster than it increases the rate at which responsible humans can absorb and govern the result.

The productivity gain is real

This is not an argument that AI productivity is imaginary. In a preregistered experiment involving 453 professionals completing writing tasks, access to ChatGPT reduced average completion time by 40% while increasing assessed quality by 18% (Noy and Zhang, 2023). The study was bounded to mid-level professional writing, so it should not be treated as a universal multiplier. It does establish that substantial gains are possible.

I experience those gains directly. An agent can produce a proposed change much faster than I could create it myself. But the faster the agents produce, the more context I must retain, decisions I must make and consequences I must understand.

I am not necessarily doing less work. I am compressing a day's judgement into a few hours.

Before working heavily with AI, I could often begin around six in the morning and continue into the evening. Now, on the days when I use AI most effectively, I can feel mentally finished by two in the afternoon. The days on which I produce the most are often the days on which I become exhausted fastest.

The production is visible. The cost is less visible because it often appears later, outside the measurement window: in review, rework, recovery or the next change that requires a mental model I failed to form.

Parallel production creates serial consumption

Several agents can produce changes concurrently. Human understanding is less divisible. Each output belongs to a context that must either remain available in memory or be reconstructed later, and consequential decisions often depend on a shared view of the whole system.

Research on task switching helps explain part of the cost. Sophie Leroy found that switching away from unfinished work can leave attention attached to the previous task and reduce performance on the next one—an effect described as attention residue (Leroy, 2009). AI did not invent that cost, but it can generate occasions to pay it faster than we previously could.

I tried to solve this operationally. I allowed agentic work to run overnight, intending to spend the morning reviewing it and use the rest of the day for planning and deeper work. It failed because the morning review consumed the capacity needed for everything that followed.

I tried batching and separated planning from execution. At one point, I was operating sixteen agents across two providers. I have reduced that to three or four.

The bottleneck was not access to production. It was my ability to consume it responsibly. The agents were technically asynchronous, but my understanding of their work had become serial.

Parallel production becomes a serial review queue wherever responsibility still converges on a person.

Work shame in a compressed day

There is a strange emotional consequence to compressed productivity. By two in the afternoon, I may have produced more valuable work than I previously completed in an entire day. Rationally, that should feel like success.

It often feels like I have stopped working early.

Many of us still use visible time and continued activity as evidence that enough work has been done. When AI compresses a day's output into a few hours, the work can be complete before the feeling of having worked enough arrives.

I think of this as work shame: the guilt that comes from being unable to continue producing, even when the work already completed exceeds what would previously have been considered a productive day. It is my label for the experience, not a clinical category.

There is an adjacent finding in research on leisure. Across four studies involving 1,310 participants, people who regarded leisure as wasteful enjoyed it less; the correlational studies also linked that belief with poorer reported wellbeing (Tonietto and colleagues, 2021). That does not specifically explain AI-assisted work, but it supports the broader idea that treating non-production as waste changes our ability to benefit from it.

Work shame encourages me to continue after my useful capacity has been exhausted. I remain at the desk because stopping feels undeserved, borrow from the energy needed for recovery and begin the next day with less capacity.

When output is compressed but expectations remain time-based, productivity gains become permission to demand more—not permission to stop.

Acceleration fatigue and comprehension debt

There is an apparent solution: trust the AI.

Stop reading every change. Ask for tests, validators and summaries, then accept the result when they pass. This can feel like maturity. Sometimes it is simply exhaustion.

I think of this as acceleration fatigue: the point at which sustained AI-generated velocity reduces our willingness or ability to scrutinise what that velocity is producing.

This risk predates generative AI. Research on automation finds that complacency can arise when multiple tasks compete for an operator's attention, and that automation bias can affect experts as well as novices (Parasuraman and Manzey, 2010). More recent research involving 319 knowledge workers suggests that generative AI shifts critical thinking towards verification, integration and task stewardship; greater confidence in AI was associated with less reported critical-thinking effort (Lee and colleagues, 2025). Because that study relies on self-reported examples, it is a useful warning rather than proof of cognitive decline.

Acceleration fatigue creates comprehension debt—the growing difference between what a system contains and what its human owner genuinely understands.

I have completely reset four projects that reached the point where continuing was less credible than starting again. The code existed. Tests existed. Documentation and validators existed. But I could no longer make relevant, informed decisions because I did not sufficiently understand what those artefacts were proving or which assumptions they shared.

The question was no longer merely whether the code worked. It was whether I still understood what “working” meant.

Trust should be calibrated to consequence and supported by evidence. Where accountability remains human, delegation cannot remove the need to understand what was decided, why it was decided and how failure will be detected.

What gets lost when the process disappears

Before AI accelerated my work, I had more time to inhabit a problem. I would reconsider decisions away from the screen and gradually form a detailed mental model of the system.

That knowledge was not created only by reading the final code. It developed through struggling with the problem, considering alternatives, making mistakes and living with the consequences of decisions.

AI can compress or remove much of that journey.

That is not automatically a loss. Good tools have always removed unnecessary labour and made room for higher-level thought. But there is a risk when the removed activity was also the activity through which a person formed the mental model needed to make later decisions.

A person can understand each individual change during review without integrating the whole system into durable knowledge. Documentation can preserve decisions and reduce reconstruction, but it cannot by itself guarantee that the people responsible for a system possess a coherent model of it.

A system can therefore become increasingly documented while becoming less understood by the people accountable for it.

Measure completed work, not generated work

Sustainable AI productivity should measure the complete delivery loop: specification, generation, review, integration, rework, maintenance and recovery. Agent count, generated output and hours spent in a seat are incomplete proxies.

The resulting work should be:

valuable rather than merely voluminous
understood by the people accountable for it
validated proportionately to its risk
integrated into durable knowledge
maintainable without permanent dependence on hidden context
produced without borrowing unsustainably from future attention

This leads to a few general operating rules. Measure end-to-end cycle time rather than generation time. Let consequence determine the depth of review. Cap concurrency at the rate at which consequential outputs can be absorbed. Preserve hands-on work where learning and mental-model formation matter. Stop when decision quality falls, not only when the clock says the working day is over.

These are not universal limits on automation. Low-risk, reversible work can often be delegated heavily. High-consequence or hard-to-reverse work needs stronger evidence and more human understanding. The governing constraint is not available compute, but the review and comprehension demanded by the risk.

A sustainable pace

Reducing my active agents from sixteen to three or four was not a retreat from AI. It was an attempt to match production with the pace at which I could still make credible decisions.

Rest, reflection and integration are not empty intervals between productive periods. They are part of the mechanism that turns output into knowledge and activity into valuable progress.

AI is a genuine accelerant and an extraordinary source of leverage. That is precisely why generation alone is the wrong thing to optimise.

AI multiplies output, but it also accelerates the consumption of the human attention required to govern that output. When production exceeds comprehension, leverage becomes debt.