How much time does AI save employees, and where does it go?

Employees who use AI save roughly 1 to 2 hours a week: 5.4% of work hours in a US survey of more than 10,000 people, and 2.8% in the April 2025 version of a 25,000-worker Danish study. Across all workers, users or not, the figure is 1.4%. Most of the time saved goes to other work, not free time.

Sources last checked: 23 September 2026. Every external figure on this page links to the publisher that produced it, and was re-read at that source before publication.

• Users, US: 5.4% of work hours saved; 1.4% across all workers (St. Louis Fed / NBER, November 2024 survey).
• Users, Denmark: 2.8% of work hours, from 0.6% to 6.8% depending on occupation and employer support (Humlum and Vestergaard, April 2025 BFI working paper, 11 occupations).
• Measured, not self-reported: Microsoft 365 Copilot users spent 2 fewer hours a week on email, with no change in meeting time (7,137 workers, 66 firms, randomised).
• Where it goes: 85% of Danish chatbot users say they moved saved time to other job tasks. Fewer than 10% took it as breaks or leisure (Humlum and Vestergaard, March 2026 NBER revision).
• The trap: in a randomised trial, experienced developers took 19% longer with AI, yet still believed it had made them 20% faster.

What counts as time AI saves an employee, and what does not

Time saved by AI is the reduction in hours an employee needs to produce the same output, measured against a baseline without the tool. Both halves matter: the same output, and a baseline.

Three things are often reported as time saved and are not. Hours assisted is time spent working with AI, not time removed. The St. Louis Fed estimated that 1.3% to 5.4% of all US work hours were assisted by generative AI, which is a separate measure from its time-saving figure. Vendor minutes-per-prompt counts the task in isolation and ignores checking and rework. Cost removed needs a changed invoice or roster, which is a separate question covered on our page about which AI cost reductions actually reach the P&L. A time saving can be real and still remove no cost.

How much time does AI save employees? Five studies, with their samples

Only a handful of studies measure time saved at scale with a stated sample. They use different denominators, so the last column converts each one to a 40-hour week, the conversion the St. Louis Fed itself uses.

The self-reported surveys and the telemetry study land in the same band, 1 to 2.2 hours a week. The honest single answer to “how much time does AI save employees” is between one and two hours a week for a regular user, and about half an hour averaged across a whole workforce.

When the answer changes: the conditions that move the number

The April 2025 version of the Danish study shows how far the figure moves with circumstances. Time saved ranged from 0.6% of hours for teachers in schools that did not encourage chatbot use to 6.8% for marketing professionals whose employers did. In the paper’s March 2026 NBER revision, 19% of workers report saving more than an hour a day in workplaces that combined encouraged use, an enterprise chatbot and training.

Four conditions move the answer:

• Frequency of use. Bick, Blandin and Deming report that time savings are highly correlated with how intensively people use the tool.
• Employer support. In the April 2025 version of their paper, Humlum and Vestergaard found time savings and other benefits 10% to 40% greater where employers encouraged use.
• Solitary versus coordinated work. In the Microsoft trial, email time fell because each worker controls their own inbox. Meeting and document time did not change, which the authors suggest is because changing them requires colleagues to agree new norms.
• Expertise on familiar work. METR’s slowdown came from experienced developers working in codebases they knew well. The researchers say it may not apply to less experienced developers or unfamiliar code.

AI time savings are largest on solitary, high-frequency drafting work and smallest on anything that needs a second person to change how they work.

Where does the time AI saves go instead?

This is the finding most summaries leave out. The largest of these studies, about 25,000 Danish workers, asked the question directly. The figures below are from the March 2026 NBER revision of that paper. Saved time is overwhelmingly reabsorbed into work, with no measurable effect on pay or recorded hours.

The percentages overlap because one worker can report more than one destination. Depending on occupation, 53% to 91% of the new tasks are directly linked to AI use: drafting and ideation, reviewing AI output for quality and compliance, and building the tools into workflows. In the Microsoft trial, treated and control workers replied to the same number of email threads, attended the same number of meetings and completed the same number of documents. The time was saved, but the measurable output stayed the same.

Time AI saves an employee is reallocated, not released: in the Danish data, fewer than one user in ten took any of it back as breaks or leisure.

Is the time saved real, or just felt?

Three of the five studies rely on employees estimating their own savings, and the one with a stopwatch shows why that matters. METR’s developers expected AI to speed them up by 24% before the trial. They were measured at 19% slower, and afterwards still believed they had been 20% faster. That is a 39-point gap between belief and measurement on the same tasks.

The self-reported figures also sit well below controlled experiments. In the April 2025 version of their paper, Humlum and Vestergaard note that randomised trials in their study occupations often found productivity gains above 15%, against 2.8% of hours in everyday use. The UK figure has a built-in ceiling: its top answer band was scored at 60 minutes however much time a respondent actually saved.

A self-reported AI time saving is a ceiling to test, not a number to budget against.

The Reallocation Ledger: a worked calculation for one team

The Reallocation Ledger turns a study figure into a number a manager can check. It has four lines, and it rests on one rule: time saved is not a result until it has a destination with a counter on it.

Worked example: a 50-person team on 40-hour weeks. Substitute your own inputs at each step.

1. Users. In the Microsoft trial, 80% of workers given access used the tool. 50 × 0.80 = 40 users.
2. Gross hours. Use the published range. At 5.4%: 40 × 40 × 0.054 = 86.4 hours a week. At 2.8%: 40 × 40 × 0.028 = 44.8 hours a week. Apply the user rate to users only. Applying 5.4% to all 50 staff overstates the total by 25%.
3. Net hours. Subtract the time spent checking AI output and the new AI tasks from the table above. No study publishes a general figure for either, so measure both in your own team. METR is the warning that for experts on familiar work, checking can use up the whole saving.
4. Destination. Name the task that receives the net hours and the output count that should rise: proposals sent, tickets closed, leads called back. If no count rises, the hours were absorbed.

The arithmetic shows why a saving can look large and still be hard to find. 86.4 hours a week is 2.16 full-time equivalents on paper. In practice it is 2 hours 10 minutes on each of 40 calendars, or 0.054 of a role per person. No one is doing 2.16 roles’ less work. At 5.4%, AI saves a 40-user team two full-time roles of hours and not one removable role. That is why the destination line matters more than the gross line.

How to measure how much time AI saves your own team, and what it costs to run

The Microsoft design is the one to copy, because it measured time instead of asking about it:

• Randomise access, or at least keep a comparable group without it, so you can separate the effect of AI from seasonal workload.
• Take a pre-period baseline from activity logs: email, meeting and document time per person per week.
• Read the result late. The trial measured months 4 to 6 of a 6-month run, after early experimentation had settled.
• Count output alongside time, using the operational efficiency metrics that fit your process, so a time saving that produced nothing extra is visible.

What this costs to run yourself is mostly calendar time and access. You need a six-month window, administrator access to productivity-suite telemetry (which usually needs privacy sign-off), someone who can run a difference-in-means comparison, and a manager willing to leave a control group without the tool. For drafting-heavy marketing work, the controlled trials behind generative AI in marketing tell you what effect size to expect before you start.

The same logic applies wherever the output can be counted directly. On revenue work the cleanest destination is a booked meeting, which is why AI appointment setting for corporate sales teams is measured in appointments rather than hours. LeadsNow reports its own record the same way: 50,769+ AI-booked sales appointments since 2017. The wider AI for business overview covers where else a countable output exists.

Frequently asked questions

How much time will AI save me each week?

If you use it regularly, expect about 1 to 2 hours a week. The St. Louis Fed’s nationally representative survey found users saved 5.4% of work hours, which it converts to 2.2 hours of a 40-hour week. Heavy, employer-supported users in drafting roles report more; occasional users report less.

Do employees use the time AI saves for breaks?

Rarely. In Humlum and Vestergaard’s Danish study, published as NBER working paper w33777, fewer than 10% of chatbot users reported taking extra breaks or leisure. 85% reallocated the time to other job tasks, and about 30% spent more time on the same tasks.

Is 26 minutes a day a realistic figure for a business case?

Treat it as a ceiling. The UK government’s Copilot experiment derived 26 minutes from self-reported ranges scored at their midpoints, and its own report states it was not possible to identify how the saved time was spent. Budget against measured output, not reported minutes.

Can AI make experienced staff slower?

Yes, on familiar, complex work. In METR’s randomised study of 16 experienced open-source developers, tasks took 19% longer with AI tools allowed, while the developers believed AI had sped them up by 20%.

Why doesn’t the time AI saves show up in pay or headcount?

Because it is spread thinly and reabsorbed. In the March 2026 NBER revision of their study, linking survey answers to Danish administrative records, Humlum and Vestergaard found no measurable effect on earnings or recorded hours, ruling out effects larger than 2% two years after ChatGPT launched. Most saved time went into reorganised and new tasks.