AI Can Make Every Task Easier. So Why Does Work Feel Harder?

Mark Zuckerberg once explained why he wore essentially the same grey T-shirt every day. He didn't want to spend mental energy deciding what to wear. Barack Obama took a similar approach to suits, usually sticking to blue or grey, and he was direct about why. "I'm trying to pare down decisions," he told Michael Lewis in 2012. "I don't want to make decisions about what I'm eating or wearing. Because I have too many other decisions to make."

I've always found those examples interesting because these weren't people optimizing a morning routine for social media. They were people making an unusual number of consequential decisions, and they deliberately removed some of the meaningless ones to protect the ones that counted.

That makes me wonder whether the rest of us are now moving in the opposite direction.

AI reduces the cognitive cost of an individual task while increasing the number of tasks a person can be expected to handle, and the speed they are expected to handle them at. We spend most of our time talking about the first half of that sentence. The second half matters more.

Every Revolution Raised the Number

Productivity revolutions tend to arrive with some version of the same promise: machines will do more of the work, some jobs will disappear, and people will eventually work less.

The mechanized loom displaced skilled hand weaving. The automobile gutted the horse-and-carriage economy while creating entirely new industries around manufacturing, roads, fuel, dealerships and repair. Computers automated large categories of routine clerical and cognitive work.

But economists David Autor, Frank Levy and Richard Murnane found that computerization also reduced demand for routine tasks while increasing demand for non-routine problem solving and complex communication. We automated some of the simpler cognitive work and pushed more people toward work requiring judgment and coordination.

Work didn't disappear. It changed.

The paperless office is the case I keep coming back to. Offices consumed more paper after computerization, not less, because generating a document became trivial and everybody printed what they were sent.

Economist Stephen Roach put numbers to the wider pattern in a Morgan Stanley newsletter in April 1987, finding that output per production worker climbed while output per information worker actually fell over a period stretching from the mid-1970s into the 1980s. A few months later Robert Solow made the observation that stuck: we saw computers everywhere except in the productivity statistics.

One way to read that pattern is that machines took over more execution while humans inherited a larger queue of judgment calls.

Over the very long run, people in wealthy countries really do work fewer hours than workers did in the late 19th century, so it would be wrong to claim technology has never bought back time. But the dramatically shorter workweek never arrived. Keynes imagined productivity growth delivering a fifteen-hour week by around 2030, and he was much closer on the productivity than he was on the leisure.

One reason is not complicated. When a company discovers that a task which took four hours can now be done in two, there are two possible responses. One is to give the employee two hours back. The other is to ask for twice the output.

Most organizations are very good at seeing the second option.

What Is Different About This One

Generative AI does something earlier technologies did not do at anything close to this scale. It creates choices.

Ask for one strategy and you get one. Ask for five more and you get those. Ask it to rewrite the email three ways, generate alternate pricing models, produce several forecasts, name the objections you missed, or build another version of the deck.

The marginal cost of producing one more option is approaching zero, while the cost of deciding among them hasn't fallen at anything like the same rate.

Cassie Kozyrkov frames generative AI as automation for tasks that have endless right answers, and her sharper point is what that does to the test. The question stops being "was it correct" and becomes "was it useful," which is a question only a human holding the context can answer.

Somebody still has to decide whether an output is accurate, useful, defensible and better than whatever else was on offer, and what "good enough" means this time. The consequence of choosing the wrong answer still belongs to a person.

Research is starting to show the shift. Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers and collected 936 real examples of generative AI use. What they found was a change in where the cognitive work sits rather than a simple reduction of it. People reported moving from information gathering toward verification, and from producing responses toward integrating them and standing behind them. Effort disappears at the front of the process and reappears at the back as judgment and accountability.

They also found a trust relationship worth flagging: the more a worker trusted the tool, the less critically they reported engaging with what it produced.

The Arithmetic Nobody Runs

Imagine someone who used to prepare three substantial analyses in a week. AI cuts the effort on each one by 40 percent. It would be tempting to say their workload just fell by 40 percent.

That isn't what happens.

The person is now expected to prepare five analyses instead of three, compare twenty AI-generated alternatives, answer more messages because communication got faster, and take part in decisions that previously never reached them.

The task gets easier. The job can still get harder.

There is an obvious objection here. AI doesn't make anyone request another analysis, send another email, schedule another review or ask for six versions of the deck. Organizations choose to do that.

That's true.

But technology changes what those choices cost. When another analysis takes four hours, somebody has to decide whether it is worth four hours. When it takes twenty minutes, the easier answer is usually to ask for it. A scenario that was previously too expensive to model becomes cheap enough to model. A question that once died because nobody had time to investigate it now gets answered and sent to somebody else to review.

Multiply that across an organization and a series of reasonable individual choices starts behaving like a workload system.

There is early evidence that productivity and work intensity can rise together. Aruna Ranganathan and Xingqi Maggie Ye followed roughly 200 employees at a US tech company for eight months and published the results in Harvard Business Review in February 2026. Eighty-three percent said AI had increased their workload. The study describes task expansion and cognitive overload, with the boundary between work and everything else blurring as people took on more across the day. Much of that extra load was taken on voluntarily because AI made it possible to hold more threads open at once.

Holding threads open is the thing that has a limit.

I don't want to overclaim. Overloaded knowledge workers existed long before generative AI arrived, and one study of one company does not establish cause. What AI changes is how quickly the work can accumulate, and how little friction stands in the way.

Hours Are the Wrong Unit

We measure workload in hours, and hours are an increasingly incomplete measurement.

Two people can both work an eight-hour day. One does relatively consistent work with clear standards and few interruptions. The other moves continuously between email, meetings, dashboards, customers, documents, chat and AI-generated recommendations, making judgment calls the whole way.

Those are both eight-hour workdays. They are not the same cognitive day.

We have a rough sense of how the second one is trending. Microsoft's 2025 Work Trend Index put the average interruption rate at once every two minutes during core hours, 275 times a day, against 117 emails and 153 chat messages. Meetings after 8pm were up sixteen percent year over year. Gloria Mark has measured attention on screens for two decades and watched the average time on a single screen before switching fall from about 2.5 minutes in 2004 to roughly 47 seconds in more recent measurements.

Nobody designed that day. It accumulated.

Where the Bottleneck Moves

Most of the conversation for the last several years has been about whether AI replaces human labour. It will replace some of it. Every major shift has eliminated certain jobs while creating others, and there is no reason to think this one is exempt.

But a different constraint may bind first, and it is the one I keep circling back to: human capacity to decide.

I am not claiming there is an established number of decisions a person can make in a day. There isn't, and the popular version of that claim is worse than unsupported. The willpower-as-battery model failed a preregistered replication across 23 labs and more than 2,000 participants, with an effect statistically indistinguishable from zero. The "35,000 decisions a day" figure that gets quoted in every article on this subject has no traceable source at all.

The honest version is a throughput argument. There is a rate at which a person can absorb context and commit to an answer, and when decisions arrive faster than that rate, the queue does not politely wait. It gets serviced badly.

People start answering the most recent thing rather than the most important thing. They defer whatever requires reading. The calendar begins choosing priorities for them, and a confident-looking AI answer becomes easier to approve when five more decisions are waiting behind it.

That doesn't prove a cognitive-capacity limit. But it is exactly the kind of environment in which less critical engagement with a trusted tool becomes consequential.

What I'd Actually Do About It

I spend my working life around cutovers and integrations, which are decision-density events by definition. A go-live compresses a quarter's worth of judgment calls into a weekend.

You learn pretty quickly that the answer isn't to make people better at tolerating chaos. You reduce the number of decisions that have to be made in the moment, make ownership obvious, protect the capacity required for the ones that genuinely cannot be made in advance, and keep the queue from filling with things that never needed a person at all.

Most of what makes a cutover survivable applies here.

Pre-decide the reversible ones. Cheap-to-undo choices still cost full price to make.

During a high-pressure event, you don't want six people debating a decision that can be reversed ten minutes later. You set a default, define the conditions that justify overriding it, and move on. The grey T-shirt is this rule applied to a closet.

AI makes this more important because it can generate endless plausible alternatives. More options are useful until comparing the options costs more than the decision is worth.

Name the decision owner before the decision arrives. A lot of what looks like overload is actually ambiguity about who decides.

In a cutover, an issue without a clear owner doesn't sit still. It gets passed around. Technical teams discuss whether it is a business decision. The business waits for a technical recommendation. More people enter the conversation because nobody is sure who has the authority to end it.

One decision has now consumed the attention of five people.

At higher AI-enabled volume, that gets expensive fast. The decision owner should be part of the design of the workflow, not something the team discovers after the question appears.

Budget verification as work. If AI writes the draft, somebody checks the draft, and the check is the job now.

This is one of the easiest costs to hide because the visible production step got faster. The analysis appears in twenty minutes instead of four hours, so the process looks dramatically more efficient. But someone still has to understand the assumptions, check the source material, resolve inconsistencies and decide whether the answer is safe to act on.

If the operating model treats generation as work and verification as overhead, the verification eventually gets done at 10pm or not done at all.

Protect a window where nothing arrives. Two minutes between interruptions is not a working condition. It is a queue with no service window.

A cutover has escalation paths for a reason. Not every issue should reach every person. The people making the highest-consequence decisions need enough uninterrupted time to understand what is actually happening before another decision lands.

The same should be true on an ordinary Tuesday.

If AI increases the volume of work a team can create, then somebody has to govern what is allowed to enter the human decision queue. Otherwise every efficiency upstream becomes another interruption downstream.

It is ordinary governance pointed at a calendar instead of a cutover plan.

The Part We Haven't Priced

We built something that produces work faster than we can judge it, and then handed it to everybody at once. The technology will keep producing.

Zuckerberg dropped the shirt decision because some decisions aren't worth the mental energy. We may now be building a working world that generates consequential decisions faster than people can process them well, and we are finding out where the limit sits by running into it.

So the question I'd put to anyone running a team right now is how many judgment calls a day your operating model is actually asking of one person, and whether anybody has ever counted.

---

About the author

Derrick Greenwood is called in when big programs stall. Cutovers, governance, operating models, and getting a team from debate to done.

Endnotes

1. Michael Lewis, “Obama’s Way,” Vanity Fair, October 2012.

2. Mark Zuckerberg public Q&A at Facebook, November 2014.

3. David H. Autor, Frank Levy, and Richard J. Murnane, “The Skill Content of Recent Technological Change: An Empirical Exploration,” Quarterly Journal of Economics 118, no. 4 (2003): 1279–1333.

4. Stephen S. Roach, “America’s Technology Dilemma: A Profile of the Information Economy,” Morgan Stanley Special Economic Study, 1987.

5. Robert M. Solow, “We’d Better Watch Out,” New York Times Book Review, July 12, 1987.[free +2]

6. Abigail J. Sellen and Richard H. R. Harper, The Myth of the Paperless Office (Cambridge, MA: MIT Press, 2001).

7. John Maynard Keynes, “Economic Possibilities for our Grandchildren” (1930).

8. Cassie Kozyrkov, “Endless Right Answers: Explaining the Generative AI Value Gap,” Medium, January 8, 2025.

9. Hao-Ping (Hank) Lee, Advait Sarkar, Lev Tankelevitch, Ian Drosos, Sean Rintel, Richard Banks, and Nicholas Wilson, “The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers,” CHI 2025.

10. Aruna Ranganathan and Xingqi Maggie Ye, “AI Doesn’t Reduce Work—It Intensifies It,” Harvard Business Review, February 2026.

11. Microsoft Work Trend Index, “Breaking Down the Infinite Workday,” June 16, 2025.

12. Gloria Mark, Attention Span, and related UC Irvine research on attention, distraction, and task switching.

13. Martin S. Hagger et al., “A Multilab Preregistered Replication of the Ego-Depletion Effect,” Perspectives on Psychological Science 11, no. 4 (2016).