← Back to blog

AI Doesn't Hand You Time: The Jevons Paradox and the Toll of Hyper-Productivity

We were promised less work with AI and we're working faster instead, which is not the same thing. Jevons, a Berkeley study, and why the time you gain has to be defended.

By · Language: English

In this article

We were sold a very specific promise: with artificial intelligence we would work less. The machine would handle the tedious part and we would keep the interesting part, plus the afternoon off. Two years in, I don’t know anyone who uses AI seriously and works fewer hours. I know plenty of people who do more things, faster, and end the day just as drained.

It isn’t just my impression, and it isn’t a personal discipline problem. It has a name, it’s a century and a half old, and it now has a field study behind it. And understanding it matters, because the answer is not to stop using the tool: it’s to stop giving away what the tool hands back.

1. From three days to fifteen minutes

The example that explains it best I heard from Marjori Haddad. A commercial proposal used to take her team about three days: researching the client, analyzing what they need, structuring, pricing, reviewing everything. Now an agent takes the meeting notes and she briefs it in two sentences, something like “make it premium but not scary, and make it obvious I understood their problem even though they couldn’t explain it well.” Fifteen minutes. Sent.

So far, promise kept. Three days turned into a quarter of an hour. Any vendor’s brochure would stop right at this paragraph.

2. The other side has an agent too

The problem is that the client also uses AI. The proposal they used to take two weeks to review comes back that same afternoon. And it doesn’t come back with “yes, let’s go”: it comes back with a full analysis, questions nobody had anticipated, and new requests instead of a start date.

So there is no free time. There are two new emails, three versions of the proposal and an open negotiation before eleven in the morning. The work didn’t shrink; the cycle sped up. What used to be correspondence chess is now blitz, and in both you still have to think through every move.

Email works the same way. If everyone replies with a model’s help, the message you expected by end of day arrives in five minutes. And you have to keep pace, because two hours of silence now reads as lack of interest.

3. What they saw at Berkeley

Aruna Ranganathan and Xingqi Maggie Ye, of Berkeley’s Haas School of Business, published a piece in Harvard Business Review in February 2026 with a title that leaves little room: AI doesn’t reduce work, it intensifies it. Behind it is an eight-month ethnographic study at a 200-person U.S. technology company where staff had broad access to generative AI tools: regular presence on site, meetings, everyday conversations and more than 40 interviews.

What they found is that AI didn’t free up time; it expanded what people felt capable of taking on. And that intensification took three forms you’ll recognize right away:

  • Wider scope — people began doing work that would previously have belonged to someone else, or wouldn’t have been attempted at all.
  • Fewer gaps — work seeped into the pauses: a prompt at lunch, another before a meeting, another in the evening when an idea came up.
  • More open threads — AI processes running in the background while reviewing code, drafting a document or sitting in a meeting.

An honest caveat: it is a single company and the research is still in progress. It’s not a universal law; it’s a very sharp photograph of something many of us already feel.

There’s a photo going around of a sign at a San Francisco gym asking members not to code with AI between sets. I don’t know whether it’s real or staged, and it almost doesn’t matter: the fact that it’s believable says enough.

4. Jevons: the coal of 1865

This has a name. In 1865 the English economist William Stanley Jevons noticed something that contradicted the intuition of his time: increasingly efficient steam engines had not made England burn less coal. It burned far more. When something becomes cheaper and faster, we don’t use the same amount at lower cost: we start using it for everything.

Swap coal for intellectual work and you have the full picture. If a proposal costs fifteen minutes, you don’t make one in less time: you make ten, twenty get requested, and all of them get reviewed. Efficiency doesn’t reduce demand; it sets it off. Jevons wasn’t talking about agents, but he described your inbox exactly.

5. My version: the bottleneck is now me

It happens to me with infrastructure. I work with agents that write scripts, review configuration and document changes, and what used to take me an afternoon now comes out in minutes. Do I work less? No. I make more changes per week, open more fronts and, above all, review far more.

That’s the toll nobody prints in the brochure. Generating got cheap; deciding didn’t. Everything an agent produces in five minutes still needs someone to read it, understand it and answer for it. The queue didn’t disappear: it moved from “doing” to “reviewing,” and at the end of that queue it’s me, with the same head and the same hours as before.

And I can’t get off either. Without these tools I’m not competitive, it’s that simple. The point isn’t to give up the time savings, which are real. The point is that the time doesn’t stay with you unless you defend it: by default, the next request takes it.

6. Defending the time

The Berkeley team proposes three thoroughly unglamorous practices, which is why I believe them. Haddad adds one of her own, and I add another.

  • Intentional pauses — brief, structured moments before a major decision, to look for the counterargument or check that it still lines up with the goal. Scheduled, not whatever is left over: we’ve already seen who keeps the leftovers.
  • Sequencing — batch what isn’t urgent, protect focus windows and let work move forward in phases. Running four agents in parallel doesn’t give you four heads to review them.
  • Human grounding — protect time to talk, compare notes and reflect with other people, precisely now that the tool lets you do everything alone.
  • Don’t do it in five minutes just because you can — Haddad’s, and my favorite. Replying instantly trains the other side to expect instant replies.
  • Decide what doesn’t get done — mine. If producing is almost free, the scarce judgment is choosing. A short list of what I won’t touch this week saves me more than any prompt.

Conclusion: the savings are real, the split is undecided

AI does save time. What it doesn’t do is decide where that time goes, and if you don’t decide, your inbox will. Jevons saw it with coal: efficiency alone never reduced the consumption of anything.

If you’d rather play with the paradox than read about it, I built a Jevons simulator into No more excuses, an interactive course of eight short chapters on AI and its second-order consequences: about 40 minutes, no sign-up.

And keep using your agents — the good news is that a proposal takes fifteen minutes. But before you send the next one, ask yourself a single question:

Of the hours AI saved you this week, how many did you keep?


Idea and starting point: Marjori Haddad (@marjorihaddad) — nobody works less using AI. The commercial proposal example and the advice not to rush just because you can are hers. The study is by Aruna Ranganathan and Xingqi Maggie Ye: AI Doesn’t Reduce Work—It Intensifies It (Harvard Business Review, February 9, 2026); details on the method come from the Berkeley Haas interview. The part about agents and infrastructure is my own experience.