In February 2025, I wrote a blog post telling you I’d saved fifty-two hours a month with AI. I gave you the stack, the cost ($105), the breakdown by task. I called it Unlock Your Time. The math was correct. The framing was wrong.
Fifteen months later, I can tell you exactly what happened to those fifty-two hours. I have no idea. They never appeared on my calendar as free time. I didn’t take Fridays off. I didn’t even take afternoons off. The hours were real — they came off the original tasks. They just didn’t show up anywhere as a reduction in how much I worked, because the moment they appeared as slack, something else moved in to fill them.
This is not a discipline problem. It’s not a “you’re using AI wrong” problem. It’s a structural pattern that has a name in economics, and I want to explain why measuring AI in hours saved is a measurement error no productivity coach is going to fix for you.
Six hours down to two. And nobody went home early.
The clearest evidence I can show you is a Fortune article from March 2026 with a title that does half my work for me: “AI just gave you six extra hours back. Your boss already took them.”
The Dun & Bradstreet CTO, Mike Manos, said it on the record: “I got the eight hours to two hours, but now I can get 20 hours of work, because the work came down.” The KPMG U.S. chair, Tim Walsh, said the same thing in different language: “That means I can put more volume through my business. My business should be growing and will grow.” Google’s senior developer-experience lead Ryan Salva confirmed the pattern at Google itself — developers using AI got redeployed to additional projects, not freed up. AES turned a 14-day auditing process into one hour. Dun & Bradstreet shrank number-crunching from hours to minutes. Nobody went home.
The cleanest description of what’s actually happening came from McKinsey partner Eric Buesing, in the same Fortune piece. Once AI use becomes the expected baseline, he said, workers “find themselves on a wagon wheel of having to build more agents to try to keep up with the expectations of production.”
The worker-level data tells the same story. A Workday survey of 3,200 employees, reported by Axios in January: 85% say AI saves them 1–7 hours a week — and 37% of those savings disappear into “rework,” correcting errors, rewriting content, verifying output. Only 14% report consistently positive outcomes.
Translated out of consultant: the time savings are real. They go directly into more output, more rework, or more concurrent work. Almost none of it shows up in a place a person could find it.
If you’ve ever gotten a raise, you already know
When your income goes up, you don’t actually feel richer the next month. The math says you should. The bank statement says you should. You don’t.
Within a few weeks, something has moved into the space. A slightly nicer apartment. A slightly better grocery store. A takeout habit that didn’t exist before. Economists call it lifestyle inflation. By six months in, the raise has been absorbed, and you’re back to feeling exactly as squeezed as you did before — just at a slightly higher standard of living.
This isn’t a moral failing. Some of those upgrades are reasonable. Some are invisible until they’re baked into the budget. The point is that raises don’t sit there as cash you can choose to spend. They expand the surface area of your life, and the surface area happily accepts the expansion.
AI is a raise paid in hours. It works exactly the same way.
There’s a 160-year-old name for this
Economists actually do have a name for what’s happening with AI hours, and it predates AI by about 160 years. In 1865, William Stanley Jevons noticed that more efficient steam engines didn’t reduce England’s coal consumption. They increased it. Cheap coal made new uses for coal viable. The efficiency gain went into expanded use, not banked savings. This is the Jevons paradox.
The paradox isn’t really a paradox — it’s a description of how demand responds to cheaper resources. When something becomes cheaper to use, you don’t consume less of it. You consume more. Cheap transportation didn’t reduce time spent traveling; it increased it, because more trips became viable. Cheap lighting didn’t reduce energy spent on illumination; it increased it, because dark spaces became lit.
This is what’s happening with AI. Cheap drafting doesn’t mean less time spent drafting — it means more drafts, on more projects. Cheap analysis doesn’t mean less time spent analyzing — it means more parallel work on more clients. The hours you “save” don’t get banked. They get spent on something the friction used to make uneconomical.
This is also why the perception-reality gap shows up so reliably in controlled studies. Developers using AI tools in a METR study took 19% longer to finish their work but believed they were 20% faster. A 40-point gap. Not because they were lying. Because the felt experience of “I used AI on this” is “this should be faster,” and the felt experience overrides the clock.
“Use AI better” is the productivity-coach exit, and it doesn’t work
Notice the asymmetry. The KPMG CEO will tell Fortune, on the record, that AI lets him put more volume through his business. The Dun & Bradstreet CTO will tell Fortune that two-hour days are actually twenty-hour workloads. The executive class is candid about the Jevons mechanic because it serves them — it’s a growth story, and growth stories are good for the stock.
The advice marketed to workers and operators still says the opposite. AI will give you your time back. If you’re not feeling the savings, you’re using AI wrong. Set better boundaries. Block your calendar. Push back on requests. Work smarter, not longer.
This is the productivity-coach exit. It’s the same advice anyone gives you about your raise. Just save more of it. Just don’t increase your spending. It treats a structural pattern as a personal discipline problem. The honest version: nobody’s career or business sits in a vacuum. The hours that became available didn’t become available to you — they became available to the system around you. Your clients. Your team. Your competitors. Your own ambition. The pressure that fills them isn’t optional in the way that an “extra” coffee subscription after a raise isn’t really optional. Both are downstream of structural patterns that don’t bend to willpower.
I made this argument before from a different angle: PwC’s 2026 finding that technology delivers only 20% of an AI initiative’s value, with the other 80% coming from redesigning work. The Jevons paradox is the measurement implication of the same insight. If process redesign is where the value is, then individual-task hours-saved is measuring the 20% slice and missing the 80%. The 20% slice will always report success. The 80% slice will always report something else, because the work itself has changed.
Three metrics that survive Jevons
The framework I should have used in 2025, and the one I’m using now:
Capacity expansion. What can you now take on that you couldn’t before? Not “what’s faster” — what’s newly possible. In my own work: I run more concurrent client engagements than I could have in 2024. The hours per engagement haven’t dropped. The number of engagements I can hold in my head has.
Scope change. What’s now in scope for a single engagement that used to require a second engagement, a hire, or a different person on the team? When I deliver a marketing attribution project, the documentation, the data dictionaries, the onboarding materials, the executive summaries — those used to be a separate workstream. They’re part of the same engagement now, because AI made it economical to include them.
Velocity at quality. Not “how fast did we ship,” but “how fast did we ship without dropping the things that distinguish our work.” If AI made you twice as fast and your output is half as good, you didn’t capture productivity. You traded standards for speed. The hours-saved metric will report this as a win. It isn’t. In my own work: AI made first-draft blog posts hour-fast. I still spend three or four on each — because the archive overlap check, the editing pass that catches AI tells, the source verification, none of those got faster. The hour version would still ship. It would also dissolve into the writing I just spent another post mocking. Speed without those guardrails isn’t productivity. It’s a trade.
These three are harder to measure than hours. They’re also the only metrics that survive the Jevons effect, because they describe what changed about the system, not what changed about your calendar.
Stop selling AI as hours saved
Stop selling AI initiatives — to your team, your investors, or yourself — on the hours-saved pitch. The pitch is going to fail. Not because you implemented it wrong. Because the hours don’t come back. They never come back.
Sell it on what’s now in scope. Sell it on the capacity to take on a client you couldn’t have served last year. Sell it on the standard of work you can hold while moving faster. These are real, defensible, and they survive the actual mechanics of AI adoption. The hours pitch survives only until your team is exhausted and asks you, accurately, where those hours went.
What I should have written in 2025
The 2025 post was correct on the arithmetic and wrong on the meaning. The fifty-two hours did come off the original tasks. They did not return as time off. The business is materially different in ways that didn’t exist before — more concurrent engagements, more in-scope per engagement, more output at higher quality — but none of that fits in “hours saved.” If I’d measured it that way then, I’d have been measuring my own raise as a windfall. It wasn’t. It was a structural change to what I could carry. Like every raise.
Run the audit on yourself. What did you tell your team — or your investors, or yourself — six months ago about what AI would save you? What actually happened? If the hours never appeared, you’re not the exception. You’re the rule. The fix is to start measuring the right thing, and to stop expecting a number that was never going to come true.
Pitching AI to your team or your board on hours-saved metrics that aren’t materializing? That’s a measurement problem with a structural cause — and it’s the work I do at Pallas Advisory. I help clients move AI ROI off labor-hour reduction and onto metrics that survive the Jevons effect: capacity expansion, scope change, velocity at quality. If you ran the audit and the answers came back the way I think they did, let’s talk.
Sources and Further Reading
- Nick Lichtenberg, “AI just gave you six extra hours back. Your boss already took them,” Fortune, March 10, 2026.
- Workday, 2026 AI Productivity Survey (3,200 employees), reported by Emily Peck in “Sure, AI is speeding up work — it’s also creating more,” Axios, January 14, 2026.
- William Stanley Jevons, The Coal Question (London: Macmillan, 1865). Reference.
Related Reading
- Unlock Your Time: 5 Marketing Tasks I Automated (52 Hours Saved) — the post this one revises.
- AI Can Complete Your Tasks. It Can’t Do Your Job. — the perception-reality gap that makes the Jevons effect feel like success.
- PwC Says Technology Is Only 20% of AI’s Value. I Could Have Told Them That for Free. — the process-redesign argument this one extends into measurement.
- The Tell Isn’t AI. The Tell Is Unedited AI. — on the editing discipline AI doesn’t get faster.
