Does AI Actually Save You Time? An Honest Accounting

You have probably had the experience of finishing something in eleven minutes that used to take an hour, feeling briefly delighted about it, and then arriving at Friday to find the week just as full as every other week. Does AI save time? That week says no, and you would still say yes.

That gap is the whole problem. The tools are genuinely, obviously fast. Whether they are saving you time is a separate question, and almost nobody is measuring it in a way that could produce an answer.

The position, stated up front: most people cannot tell whether AI saves them time, because the feeling of speed is a terrible instrument and because the part that got faster is the part you sit and watch. Generation collapses from an hour to thirty seconds, in front of you. Verification, correction and re-prompting either get slower or appear from nowhere, and those you do not count, because they do not feel like the task. They feel like tidying up afterwards.

Net time can go up while the work feels faster. That is the most dangerous combination available, because it quietly removes the signal that would otherwise tell you to stop.

The Ledger Almost Nobody Fills In

Time saved is a subtraction, and most of us only do half of it.

The full sum has four terms: generation time, plus verification time, plus rework time, minus what the task would have cost you by hand. People track the first and the last, find an enormous difference, and get on with their day.

Verification is the line that decides the answer, and it is the hardest one to see, because checking does not announce itself as work. You read something, you feel a small flicker of doubt, you read it again, you open a second window to confirm a figure, you fix a sentence that is technically true and subtly wrong. None of that registers as effort in the way that staring at an empty document does. It still takes twenty minutes.

It also takes something other than minutes. Reviewing output is judgement work, and judgement is the resource that runs out first: by late afternoon the checking gets shallower, which is exactly when a tired brain stops generating alternatives and reaches for the default, and the default when you are tired is to accept what is in front of you. The cost of verification is therefore highest at the moment you are least able to pay it.

The best evidence on all of this comes from software, because software is where the effect is easiest to measure. In 2025 the research group METR ran a randomised controlled trial with sixteen experienced open-source developers working through 246 real tasks on repositories they had maintained for years. Half the tasks allowed AI tools, half did not. Before they began, the developers expected AI to cut their completion time by 24 per cent. Measured against the clock, they were 19 per cent slower when they had it.

The finding that matters more is the one after that. Having done the work, having been timed, the same developers estimated that AI had sped them up by 20 per cent. The stopwatch said one thing and the experience said the reverse, and the experience won by a distance of roughly forty points.

Before anyone feels bad about their own estimating, the outside experts asked to forecast the result did worse than the participants. Economists predicted a 39 per cent speedup. Machine learning experts predicted 38. Everybody was wrong in the same direction, which tells you the illusion is structural.

Does AI save time for experts? One study, one domain, sixteen people. Treat it as a single data point. It is also the only data point we have from a properly controlled trial of expert practitioners doing real work, which makes it considerably better than the vibe.

Does AI Save Time? Where The Gain Is Real

None of this means the gains are imaginary. They are large, well documented, and concentrated in places you can describe precisely.

Shakked Noy and Whitney Zhang ran a controlled writing experiment with 444 college-educated professionals on realistic mid-level office tasks. Time fell by around 40 per cent and graded quality rose by 18 per cent. Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied 5,179 customer support agents given a conversational assistant and found a 14 per cent rise in issues resolved per hour on average, 34 per cent for the newest and least skilled agents, and close to nothing for the most experienced ones.

Look at what those tasks have in common. A cover letter, a press release, a support reply: you can tell in about four seconds whether the output is acceptable. Verification is nearly free. And if something slips through, the cost is a slightly awkward email rather than a corrupted database or a wrong number in front of a board.

That is the rule, and it is unglamorous. The gain is largest where you can judge the output instantly and the cost of being wrong is low. It shrinks as either of those conditions weakens, and it goes negative where checking the answer takes about as long as producing it would have.

Which is, awkwardly, an excellent description of the expert work people most want to speed up. Legal analysis, financial modelling, medical reasoning, architecture decisions in a large codebase. The reason those tasks are slow is that they require someone to be sure, and being sure is the expensive part. Generating a plausible answer was never the bottleneck.

Experience Inverts The Benefit

Read the support-agent numbers again and the shape of the thing appears. The newest agents gained 34 per cent. The veterans gained almost nothing.

This is the uncomfortable symmetry at the centre of the technology. The person who is weakest at a task gains the most from the output and is the least equipped to evaluate it. The person who is excellent gains the least and is the only one in the room who can tell whether what came back is any good.

So the benefit is real and it lands mostly on people who cannot verify it, while the verification capacity sits with people for whom the speedup is marginal. That is a strange machine to have installed in an organisation, and it explains a lot of the arguments currently happening inside them.

It is also why handing work to a model goes wrong in the same way that handing work to a junior goes wrong. Delegation usually fails because the task gets handed over while the judgement stays behind, and the person receiving it can execute but cannot tell when the thing has gone off the rails. A model is that same arrangement, compressed to seconds and stripped of the one useful feature a junior has, which is the ability to come back and say they are stuck. Your model will never do that. It will produce something confident and complete and hand it to you at a speed that makes checking feel unnecessary.

Repetition Beats Difficulty

The second question, and in practice the more useful one, has nothing to do with how hard the task is. It is how many times you are going to do it.

Setup costs are real. Working out how to describe the task, assembling the context, discovering the two or three ways the output goes wrong and learning to head them off. That investment can run to an hour or an afternoon. For something you will do a hundred times, it is trivially worth it, and the hundred-and-first run is where the genuine compounding lives. For a one-off, you have just spent an afternoon shaving fifteen minutes off a job you could have done by hand before lunch.

Difficulty is a poor guide here and repetition is a good one. A hard thing you do weekly is a better candidate than an easy thing you will do once. I use these tools for several hours a day and the pattern holds without exception in my own work: everything that has genuinely paid off was something I had already done many times and could therefore describe precisely and check quickly. The disasters were all novel one-offs, where I could not specify what I wanted well enough to get it, and could not tell whether I had got it.

The Work You Were Avoiding

There is one more line in the ledger, and it is the one most likely to be misfiled.

A great deal of what these tools give us is work we used to skip. The competitor analysis nobody had time for. The three alternative versions of the plan. The long document actually read rather than skimmed. The messy dataset finally interrogated properly instead of eyeballed.

That is a real gain, often the largest one available, and the honest reason to be enthusiastic about the technology. It is also not a saving on the clock. Your hours did not decrease. You converted work you were avoiding into work you now do, which raises the quality of your output and raises your workload at the same time.

Call it what it is and the accounting stays clean. Quality improvements go in the quality column. Capability expansions go in the capability column. Only actual clock time goes in the time column, and it is the emptiest column of the three for most people, most of the time.

Three Tasks, Three Columns

Does AI save time in your work? The measurement is hard, so here is something small enough to actually do this week.

Pick three tasks you would normally hand to a model. For each one, write down three things.

• Before you start: your honest estimate of how long it would take you by hand.
• Afterwards: the real clock time, from first prompt to finished artefact, including every re-prompt, every correction, every moment spent checking a claim you half-believed.
• Then the question that does the work: would you have shipped the unedited output?

The first two columns are useful and will surprise you less than you expect. The third column is where people find the answer. If you would have shipped it untouched, the saving is real and probably enormous. If you rewrote a third of it, you did not save the generation time, you moved it into editing, which is slower work and worse work and harder to stop doing. If you could not tell whether it was right, you have found a task where this tool costs you time and disguises the fact while doing it.

Do that for three tasks and you will have more information about your own use of AI than almost anyone you work with has about theirs. The default instruments everybody is handed are broken: the vendors report generation speed, and your own nervous system reports how fast the work felt.

Time is what the stopwatch measures. And the stopwatch has to keep running while you fix what came back.