AI fatigue and the work nobody asked for

There used to be a certain reassurance in receiving a substantial proposal from a coworker. Whatever its merits, you could reasonably assume somebody had invested some time putting it together. Corporate PowerPoint has enjoyed a long career making questionable decisions look respectable, but producing a sufficiently elaborate presentation still required a fair amount of effort.

That effort wasn’t evidence that the idea was good. It was, however, evidence that somebody had considered it worth the trouble.

This has become a considerably less reliable signal.

A proposal can now arrive with a market analysis, implementation roadmap, financial projections, and an impressively confident executive summary, all assembled between meetings. It might be excellent. It might also contain very little original thinking. Either way, the people receiving it still have to determine whether the assumptions are sound, whether the recommendations make sense, and who would be responsible for carrying them out.

The document has become much cheaper to produce than it is to evaluate.

I spend a frankly unreasonable amount of time working with AI. I use it for research, software development, technical troubleshooting, and exploring ideas that I probably wouldn’t have the time to pursue otherwise. I’ve watched these systems become capable of doing work that would have seemed implausible a few years ago, and I continue to find that fascinating.

So I have little patience for the argument that AI is incapable of producing meaningful work. I’ve seen too much evidence to the contrary.

Yet I understand why people are getting tired of hearing about it.

There’s an exhausting quality to the corporate conversation surrounding AI. Every software release seems to arrive with another announcement about intelligence. Every industry has its collection of transformative use cases, and employees are encouraged to demonstrate how they’re adopting AI, sometimes before anyone has established what successful adoption would accomplish.

Even customers seem increasingly interested in getting past the label. Describing a feature as AI-powered doesn’t necessarily explain what it does, why it matters, or how it improves anything. A customer who asks those questions isn’t necessarily skeptical of technological progress. They may simply have become less impressed by the vocabulary used to advertise it.

Inside organizations, something similar happens when AI adoption becomes a measure of progress in its own right.

But beyond the marketing fatigue and the pressure to keep up, there’s another problem that interests me more, partly because I think those of us who are enthusiastic about AI are especially susceptible to creating it.

We’ve made it extraordinarily easy to produce work without giving nearly enough consideration to the obligation that work creates for everyone else.

The cost of someone else’s good idea

Consider what happens when someone uses AI to generate an elaborate proposal for a new initiative. The proposal might be accurate, thoughtful, and well researched. It might represent a genuinely good use of AI.

It can also be completely unnecessary.

That’s a category of waste that gets surprisingly little attention in conversations about AI adoption. We spend a great deal of time discussing hallucinations, inaccurate information, and the low-quality material people have started calling “workslop.” Those are legitimate problems. Receiving a polished document that turns out to be nonsense is irritating, particularly when correcting it becomes your responsibility.

I’m increasingly interested in what happens when the output is good.

An excellent analysis of a market your company has no intention of entering is still an analysis someone has to read. A technically sound proposal for an unnecessary feature can consume just as much organizational attention as a poorly conceived one. In some ways, the better proposal is more difficult to deal with. It’s harder to dismiss, easier to circulate, and more likely to acquire a meeting.

And the trouble with a reasonable proposal is that rejecting it often requires work.

Someone may need to investigate its assumptions, consult the relevant teams, estimate the costs, or explain why other priorities deserve the resources. The person generating the proposal might have saved several hours. The people responsible for evaluating it can easily spend those hours again.

It’s an odd form of delegation. You can create work for other people without ever formally assigning it to them.

Of course, that isn’t true of everything AI produces, and I think the distinction matters.

If I spend twenty minutes generating several approaches to a technical problem, discover that most of them are useless, and discard them, I haven’t necessarily created work for anyone else. I may have saved myself an afternoon. If I use AI to summarize a lengthy document accurately, I might spare several colleagues the effort of reading it. Both are perfectly reasonable examples of productivity improvement.

The problem changes when the output begins making demands on other people’s attention.

A private experiment has a different cost from a proposal requiring formal review. A draft that can be discarded tomorrow has different consequences from a software feature that needs to be supported for the next five years.

And something can be inexpensive to produce while becoming extraordinarily expensive to maintain.

In my work across websites and technical systems, this distinction is particularly apparent. A landing page can look finished in a design file long before it’s ready to serve customers. There are questions about accessibility, localization, performance, analytics, integration with existing systems, and what happens when someone needs to update it six months later.

AI can help with much of that work, too. I’ve found that enormously useful. But making it easier to generate another page concept doesn’t establish whether the website needs another page.

The same applies to software features, marketing campaigns, internal processes, and practically any other kind of collaborative work.

What interests me is the point at which a possibility becomes an expectation. When does a sketch become a proposal? When does a proposal become something a team is expected to review? At what point does someone’s inexpensive experiment become another person’s responsibility?

These transitions have always existed in organizations. AI is making them easier to cross, often without anyone noticing the costs accumulating on the other side.

When productivity gets stuck in the pipeline

There’s some interesting research beginning to illustrate the problem.

In an August 2026 working paper, Harvard researchers Fiona Chen and James Stratton examined approximately 300 million engineering work events across 718 software firms. They found that adopting AI coding agents was associated with substantial increases in coding activity, including roughly 30 percent more lines of code and 23 percent more pull requests per worker.

Yet those increases didn’t translate into statistically detectable gains in completed Jira issues or larger projects. The average interval required to review and merge a pull request also increased by 49 percent, while requests for changes and review comments became more common.

Those findings deserve careful interpretation. The study is observational, completed Jira issues are an imperfect measure of software value, and the increased review interval includes waiting and coordination rather than simply time spent reading code. An absence of statistically significant improvement doesn’t establish that no improvement occurred.

Still, the research illustrates a familiar organizational problem. Accelerating one part of a process doesn’t guarantee that the entire process becomes faster.

There’s also good evidence that AI can produce genuine time savings. A randomized field experiment involving 7,137 knowledge workers across 66 firms found that employees who used an AI assistant spent about two fewer hours per week on email and reduced their work outside regular hours. The researchers didn’t detect corresponding changes in the overall quantity or composition of work.

I find both results entirely plausible.

Writing emails is something an individual can often improve without requiring the organization to redesign how it operates. Delivering software, on the other hand, requires several interdependent activities to work together. Faster code generation is useful, but the code still has to make its way through everything required to turn it into reliable software.

The International Labour Organization’s 2026 review of generative AI research reaches a similarly nuanced conclusion: productivity gains are real in some settings but uneven, and reported time savings have not consistently appeared as improvements in measured output or employment outcomes.

None of these studies proves that AI inevitably creates more work or fails to improve productivity. They examine different technologies, people, and environments.

What they do suggest is that individual efficiency and organizational efficiency are related, but they’re not interchangeable.

I think that distinction deserves considerably more attention than it receives.

A company might calculate its productivity gains by measuring how much faster employees can create reports, proposals, or code. Those measurements can be useful. But the calculation becomes misleading if it stops when the employee presses Send.

The people receiving that work have schedules, priorities, and responsibilities of their own. They may be perfectly capable of evaluating the additional material without having the time to do so.

And when everyone becomes more productive at generating things, the demand for everyone else’s attention can increase at the same time.

The bottleneck moves.

The friction we should be glad to lose

There’s an important complication to this argument.

Historically, the effort required to produce something acted as a kind of accidental filter. Preparing an elaborate proposal took time. Building a prototype required resources. Exploring an idea meant accepting an opportunity cost.

It would be easy to conclude that removing this friction was a mistake.

I think that would be a terrible conclusion.

Those same barriers prevented countless worthwhile ideas from being explored. People with limited technical ability, fewer resources, or less organizational influence were often unable to demonstrate the value of their ideas in the first place. An excellent proposal that never gets written isn’t much use to anyone.

One of the things I find most exciting about AI is that it makes serious experimentation accessible to far more people.

Someone who couldn’t previously write software can begin testing an idea. Someone who lacks specialized research experience can investigate a difficult subject. A small team can explore possibilities that would once have required a much larger investment.

I wouldn’t want to restore the old barriers merely because they happened to prevent some unnecessary work.

But those barriers did serve a filtering function, however imperfectly. As they disappear, organizations need better ways to distinguish interesting possibilities from things worth committing resources to.

I don’t think we need to make experimentation more difficult. We need to become more deliberate about what happens afterward.

A twenty-minute AI-assisted investigation might be all that’s needed to determine whether an idea deserves further consideration. If the answer is no, that’s a successful experiment. The exploration did exactly what it was supposed to do, and the organization can move on.

If the answer is maybe, that doesn’t automatically justify a thirty-page proposal and a meeting with seven people.

Sometimes the next step should be another small experiment. Sometimes the relevant person just needs a brief explanation of what was discovered.

And sometimes the best outcome is realizing that the original problem doesn’t require a new solution.

There’s a tendency in professional environments to treat an idea becoming more elaborate as evidence that it’s making progress. AI makes elaboration particularly easy. A rough concept can acquire polished mockups, implementation timelines, and cost estimates before anyone has established that the concept deserves further investment.

The presentation may become more convincing without the underlying idea becoming any more valuable.

An organization that understands this can encourage experimentation while maintaining a much higher standard for turning experiments into commitments.

That’s a distinction I’d like to see become normal.

A better way to spend the time we’re saving

I don’t think the solution requires a complicated new management framework. In fact, turning every AI-generated idea into another intake form, approval process, and collection of mandatory fields would be a fairly impressive way to recreate the problem we’re trying to solve.

But organizations could benefit from being more deliberate about when work becomes shared.

An exploratory draft should be allowed to remain an exploratory draft. Employees ought to have room to investigate ideas, discover that they aren’t useful, and abandon them without having to justify the time spent experimenting.

When an idea does require other people’s involvement, I think a few basic expectations are reasonable.

What problem are we actually trying to solve? What decision needs to be made? Why does it deserve attention now? What evidence would convince us that the idea isn’t worth pursuing?

And perhaps most importantly, who will have to do the work if we decide to proceed?

These are ordinary questions, but they can get lost when the cost of producing a compelling proposal becomes so low.

I’d also like to see organizations acknowledge review effort as an actual cost. If someone can generate a prototype in an hour, but validating it requires several days of engineering and security review, the cost of that prototype isn’t one hour.

That doesn’t mean the prototype wasn’t worth creating. It means the organization should understand what it’s committing to before celebrating the time saved.

The same logic applies to measuring results. Faster document production may be valuable, but it’s only one part of the picture. How quickly did the team reach a useful decision? How much rework was required? Did the change improve the customer experience? What new maintenance obligations were introduced?

Not every improvement needs a dashboard, and not every benefit is easily quantified. But if we’re going to make claims about productivity, we should at least attempt to measure the part of the process that produces the intended result.

There is another measure I’d like to see taken more seriously: work that no longer needs to happen.

An outdated process that can be eliminated instead of automated. A feature that doesn’t need to be built. A recurring report nobody uses. A proposed initiative that turns out to be unnecessary after a short investigation.

These outcomes tend to be less visible than completed projects, which makes them easy to overlook. Nobody gets to demonstrate the feature that was never developed, and it’s difficult to put an avoided meeting in a quarterly presentation.

Yet eliminating unnecessary work can create value without requiring anybody to become faster at doing it.

The irony is that AI may be particularly good at helping us identify those opportunities. It can examine existing processes, explore alternatives, surface overlooked assumptions, and help investigate whether a proposed solution is justified.

We just have to be willing to accept the possibility that the best answer is to stop.

When possibility becomes obligation

There’s also a more personal dimension to all of this.

The promise of increased productivity has a peculiar habit of becoming an expectation of increased productivity.

When something that once took a day can be completed in an hour, the remaining time doesn’t necessarily become an opportunity to think, learn, or recover. It may simply make room for additional assignments.

Sometimes organizations establish those expectations directly. Sometimes employees discover what they’re suddenly capable of accomplishing and begin raising the expectations themselves.

I understand that temptation.

When you discover that you can accomplish something that previously required far more effort, it’s difficult to resist finding out what else you can do. There’s genuine excitement in that experience. I think it’s one of the reasons AI has attracted such intense interest from developers, researchers, and people who simply enjoy making things.

But an expanding sense of possibility can become an expanding sense of obligation.

An eight-month workplace study described in Harvard Business Review observed employees at a technology company taking on a broader range of tasks, working faster, and extending work into more hours of the day after adopting AI, often without being explicitly asked to do so.

The researchers weren’t describing a universal consequence of AI adoption. They were observing how employees adapted their behavior in one organizational environment.

But the pattern is worth considering, especially for people who enjoy using these tools.

There’s a point where increased capability can make leaving something undone feel like a missed opportunity. When exploration becomes easier, there’s always another question to investigate, another improvement to attempt, another idea that seems just a few prompts away from becoming real.

I don’t think this is unique to AI. Ambitious people have always found ways to overextend themselves.

What’s different is how quickly the tools can turn that ambition into something tangible.

And the consequences don’t necessarily stop with the person doing the experimenting. Every promising idea has the potential to become another conversation, another request, or another project competing for time.

The people downstream of our enthusiasm may be considerably less enthusiastic about the additional work we’ve made possible.

This is one reason I think organizations need to consider the distribution of productivity gains, not just their existence.

When AI saves an employee several hours, what happens to those hours? Does the person gain room to focus on more meaningful work? Does the team deliver something better? Does the organization eliminate an unnecessary process?

Or does the employee simply become responsible for producing more?

Those are different outcomes, and they have different implications for the people doing the work.

The value of doing less

None of this changes my assessment of AI’s potential. I expect these systems to become increasingly capable, and I think organizations that learn to use them intelligently will have real advantages.

Some of the most interesting applications may involve entirely new ways of working rather than improvements to existing tasks.

But I think we’re going to have to become more disciplined about distinguishing what technology makes possible from what an organization should actually pursue.

For decades, we’ve been accustomed to demonstrating productivity through visible output. Reports completed, projects initiated, features delivered, meetings held. Generative AI makes it easier to produce many of those signals, which also makes them less reliable as evidence of progress.

The more capable these systems become, the easier it will be to fill an organization’s available attention with work that appears useful.

And I suspect that will make good judgment, particularly the judgment to decline something plausible, increasingly valuable.

I’d love to see more companies talk about the projects they decided against after using AI to investigate them. The unnecessary process they eliminated. The elaborate proposal that never made it to a meeting because somebody discovered a simpler answer.

I’d be interested in hearing about the time a team used AI to investigate a problem and discovered that the problem had been misunderstood all along.

Or the feature that seemed like a good idea until someone considered the maintenance cost.

Or the initiative that was quietly abandoned because a little additional research made it clear that the expected benefit wasn’t there.

These aren’t failures to capitalize on AI. They’re examples of using increased capability to make better decisions about where effort belongs.

And they’re precisely the kinds of outcomes that can get lost when organizations concentrate too heavily on demonstrating how much more work AI allows them to produce.

The original promise of workplace automation was that technology would take some of the work off our hands. We’ve become so fascinated with everything AI enables us to do that we risk overlooking that rather modest ambition.

I think it’s worth recovering.

Because if we end up using AI to create more reports nobody reads, more projects nobody needs, and more meetings nobody has time for, we will have accomplished something rather strange.

We’ll have developed extraordinary technology for making ourselves unnecessarily busy.

There probably won’t be much to show for the work we successfully avoid. No presentation announcing the meeting that never happened. No product launch celebrating the feature nobody needed.

But if you’ve ever spent an afternoon in a meeting about a project that didn’t need to exist, you can probably appreciate the value of preventing one.

And I can’t think of a much better use of a technology designed to save us time.