What AI Actually Saves You: Proving It With Tracked Hours

The short answer

You can't know what AI saves you until you track it. Most published "hours saved" figures are estimates people made from memory, and the few studies that timed the work found a different picture. The fix is cheap. Track the same kinds of tasks for a week without AI and a week with it, count the review time, and compare minutes per task.

What the big numbers actually measure

The headline figures are real, but they measure perception. A St. Louis Fed survey found that generative AI users saved an average of 5.4% of their work hours, about 2.2 hours a week for someone working 40 hours. Across all workers, including people who never use AI, the figure drops to 1.4%. Respondents were asked how many extra hours they would have needed to finish the same work without AI. That is a guess about a situation that never happened.

Microsoft's Work Trend Index surveyed 20,000 knowledge workers who use AI. It found that 66% say AI lets them spend more time on high-value work. That tells you how people feel their time is being used, not how many hours were saved, and it is also self-reported.

Why self-reported savings mislead

Memory is a poor stopwatch. You remember the dramatic wins, like a draft that appeared in seconds, and forget the twenty minutes spent rewording the prompt and fixing the result. If you have read about time blindness, this will sound familiar. The same gap shows up with AI time-tracking insights: the tool can tell you a lot, but it can't tell you how long the work felt to you or what you did with the time.

What happens when time is measured

METR ran a randomized trial in 2025 with 16 experienced open-source developers working on 246 real issues, each about two hours long. Each issue was randomly assigned to allow or disallow AI. Developers recorded their screens while they worked.

Allowing AI made them take 19% longer. Before the study they expected AI to speed them up by 24%, and afterward they still believed it had sped them up by 20%.

Be careful with what this shows. It's a small group of experienced developers working on codebases they knew well, using early-2025 tools, and METR notes that newer results may differ. It doesn't say AI slows you down. It says that feeling faster and being faster can diverge, which is the reason to measure.

A two-week before-and-after test

You need a time tracker, two weeks, and a little discipline.

Pick two or three repeat tasks

Choose work you do often and that comes in similar sizes: client emails, first drafts of a proposal, invoice descriptions, meeting summaries, bug triage. Avoid one-off projects, since you can't compare them.

Track one week without AI, one week with it

In week one, do those tasks the way you always have and log the time. In week two, use AI where you think it helps and log that too. Use the same project or tag for each task type so the totals line up.

Log as you work

Start the timer when you begin, stop it when the task is done. Logging as you go beats filling in hours at the end of the day, and it matters more here because the whole point is to catch the small stretches memory drops.

Count the review time

The most common mistake is stopping the clock when the AI finishes. The task isn't done until the output is checked and fixed. Count everything in the AI column: writing the prompt, waiting, reading the result, correcting it, rewriting the weak parts. If you ever wonder why project estimates are always off, hidden review time is a likely culprit.

Compare minutes per task, not total hours for the week. A busy week with more tasks will make any total look worse, and a quiet week will flatter it.

Reading your results

Look at each task type separately. Where AI cut the minutes per task even after review, keep it. Where it matched or lost to your usual method, drop it for that task, or try a different way of prompting and run one more week. Tools change fast, so repeat the test every few months on the tasks that matter most to your billing. If you bill by the hour, remember that time AI saves you is time you can spend on another client or on rest, and it affects what you earn per hour. Your effective hourly rate shows that clearly on fixed-price work.

How Pomlo fits in

Pomlo gives you one-tap start and stop, so logging a two-minute prompt session is as easy as logging a two-hour block. Organize hours by project or client, using one for AI-assisted work and one for manual work, and the Reports view shows where the week went, task by task. Focus sessions help you track focused work, not just total elapsed time. It's built for freelancers, indie hackers, and small teams who want honest numbers for their own decisions.

Try the two-week test on iOS or Android, with the web app keeping everything in sync.

Frequently Asked Questions

How long should I track before judging an AI tool?

Two weeks is a practical minimum: one week without the tool and one with it, on repeat tasks of similar size. Shorter windows are dominated by whatever mix of work happened to land that week.

Does the METR result mean AI slows everyone down?

No. It covered 16 experienced developers on their own mature codebases with early-2025 tools, and METR does not claim AI fails to help most people or other kinds of work. What it shows is that felt speed and measured speed can differ.

What counts as AI time?

Everything from writing the prompt to the finished, checked output: prompting, waiting, reviewing, correcting, and rewriting. Time spent fixing the output belongs in the AI column.