How Much Time Does AI Actually Save? The Numbers, Reconciled

Somebody asks how many hours a week AI saves. You look it up. You get 2.2, and one to seven, and plenty of others, all from sources that sound credible.

They are not really disagreeing. They are measuring different things, on different populations, with methods that are not comparable. Once you line up what produced each number the range stops being confusing and starts being informative.

The numbers, with their methods attached

Federal Reserve Bank of St. Louis: about 2.2 hours a week. They found time savings equal to 5.4 percent of work hours. The method was a survey question asking generative AI users how many additional hours they would have needed to finish their work without it. That is a self-reported counterfactual, which people are famously poor at estimating, and the direction of the error is not obvious. Users may overstate because the tool felt helpful, or understate because they have forgotten how the old way went.

Computer and mathematics occupations: about 2.5 percent of work time. This one, from the same St. Louis Fed survey, is worth reading alongside its other half. The same workers used generative AI in nearly 12 percent of their work hours. Twelve percent of hours producing two and a half percent of savings is a poor exchange rate, and it is the number the sceptics should be citing instead of the MIT one.

Workday, surveying 3,200 employees and business leaders: 85 percent report saving between one and seven hours a week. Note who was surveyed. Every respondent worked full time at a company with more than $100 million in revenue and was already an active AI user, which is a narrower and more enthusiastic group than the workforce as a whole.

Why they diverge

Four things, and once you see them the spread makes sense.

Who is being asked. Self-report by the user, report by their manager, and instrumented measurement produce three different numbers, reliably in that order from lowest credibility to highest. Almost nothing in this literature is instrumented.

Which workers. Everybody in a company, versus AI users only, versus knowledge workers only, versus knowledge workers using agents. Each filter raises the number, and headlines rarely carry the filter.

What counts as saved. Time not spent on the task, or time freed to do something else. Those diverge sharply, for the reason below.

Adoption intensity. Somebody using AI for 12 percent of their hours and somebody who has restructured their week around it are not the same measurement.

The finding that eats the gains

Here is the one that should change how you read all of the above.

Nearly 40 percent of the time saved is lost to rework, according to the same Workday research: correcting errors, rewriting content and verifying output.

The mechanism is not surprising once said out loud. A model drafts the thing in four minutes instead of forty. Then somebody checks it, finds the parts that are wrong, and redoes them. The drafting time was genuinely saved. A large part of it came straight back as verification and correction.

This does two things to the numbers above. It means the honest figure is somewhere below whichever one you were citing. It also means the gains are real but arrive net of a tax that scales with how much you trust the output, which is why the same tool produces wildly different results in two firms doing the same work.

Firms that treat AI output as a first draft to be reviewed pay the tax and keep a smaller, real gain. Firms that treat it as finished work do not pay it up front and pay it later, with interest, in the form of something wrong reaching a client.

What I would actually tell you

If you want a working number for planning, start from 2.2 hours a week, the St. Louis Fed survey's figure for someone who actually uses it, take rework off the top, and treat anything above that as a claim requiring evidence about that specific workflow.

That is unexciting. It is also enough to matter. A couple of hours a week per person across a forty-person firm adds up without anyone needing to believe anything dramatic.

Measure your own, because none of this is about you

The reason I keep pushing firms to time their own baseline is that every number above is an average across populations that look nothing like your practice.

It takes two weeks and it is dull. Pick the task. Time it as it is currently done, for ten instances, before touching anything. Then build, then time ten more. You now have a number that is true about your firm rather than about 3,200 survey respondents at other companies.

Do that and you stop needing anybody's statistic, including mine. Skip it and you are choosing to run on someone else's average, which is how firms end up with an AI budget nobody can defend and a saving nobody can find.