How to Use AI to Save Time at Work

Two years of measured research now exist on whether AI tools actually save time at work. The findings are consistent enough to act on and narrow enough to disappoint. On well chosen tasks, controlled trials show real and sometimes large reductions in time spent. On whole jobs, the effect is modest. The difference between the two outcomes is not the sophistication of the model. It is the choice of task, and the discipline applied to checking the output.

This piece is a working guide to that task level reality. It draws on the randomised trials and field experiments published since 2023, because those are the studies that measure time rather than sentiment. The practical advice that follows sits downstream of their results.

What the measured evidence actually says

The cleanest single result comes from a preregistered experiment published in Science in 2023. Noy and Zhang assigned midlevel professional writing tasks to 453 college educated workers and gave half of them ChatGPT. Average time on task fell by 40 percent while output quality rose by 18 percent, and participants who had used the tool during the experiment were twice as likely to use it in their real jobs two weeks later. That is the ceiling case: a self contained, text shaped task with a clear deliverable and no integration with anything else.

Field experiments inside real companies show smaller and stranger numbers. Microsoft researchers ran a six month randomised trial with 7,137 knowledge workers across industries, giving half of them an AI assistant built into the applications they already used. Workers who used the tool in most weeks spent 3.6 fewer hours on email per week, a 31 percent reduction, and completed documents moderately faster. They did not spend less time in meetings. The tool reshaped the work people could change on their own and left untouched the work that requires other people to agree.

Who benefits matters as much as how much. Brynjolfsson, Li and Raymond studied 5,172 customer support agents given a generative assistant. Average productivity rose about 15 percent, but the gain for newer and less experienced agents was around 34 percent while the most experienced saw little change. The assistant compressed the knowledge gap between novices and veterans. If you already perform a task quickly, the tool has less to hand you.

Software development fits the same shape. A randomised controlled trial at Google with 96 full time engineers estimated that AI assistance shortened time on an enterprise grade task by about 21 percent, with a confidence interval wide enough to signal honest uncertainty. And the counterexample carries as much weight as the positives: a preregistered trial with police officers found that AI assisted report writing produced no significant reduction in report duration at all, a null result that held across a full year of reports.

Read together, the trials establish four boundaries. Text shaped tasks with clear deliverables save the most time. Email and first drafts respond well. Meetings and coordination do not respond. The most experienced practitioners in a domain benefit least. Anyone planning to save time with AI is planning inside those boundaries, whether they have read the research or not.

Start from a time audit, not from the tool

The common approach runs backwards. Someone hears about a new model, tries it on a random Tuesday afternoon, gets uneven results, and concludes the technology is overhyped. The trials suggest the opposite sequence. Begin with where the hours actually go, and let the tasks choose the tools.

A working audit takes one week. For five days, record the work in half hour blocks: drafting, reading, summarising, searching for information, attending meetings, coordinating across teams, reviewing other people's work, waiting on decisions. Most knowledge workers find the distribution surprising. Email and document production typically dominate the clock, while the judgement work that feels most important is a minority of the week.

Then classify each block against the evidence. Tasks that are self contained, text shaped, and measured by a deliverable are the candidates: first drafts, summaries, meeting notes, status reports, boilerplate code, reformatting, restructuring existing material. Tasks that require other people to change behaviour, or where a wrong output costs more than the time saved, are not candidates, whatever the marketing claims.

Where the time is genuinely saved

First drafts

Drafting is the best documented win in the literature. The 40 percent result from the Science experiment was a drafting task, and the practical pattern behind it is simple. Treat the model as a fast, mediocre first drafter rather than a slow, reliable author. Supply the raw material, the outline, the constraints and the tone, and demand a complete first draft in one pass. Then edit. The time is saved because you never start from a blank page, not because the model's output is publishable unedited.

Email and short communication

The 3.6 hour weekly reduction in the Microsoft trial came substantially from email. The mechanics are unglamorous: drafting replies, tightening tone, summarising a long thread before responding, and writing the routine status updates that fill a working week. None of these is difficult, which is exactly the point. Difficulty is not the criterion for automation. Repetition and text shape are.

Summaries and catch up

Reading long documents and long threads to extract what matters is a second reliable category. A model can produce a serviceable summary of a fifty page report in seconds, and the reader's job becomes checking the summary against the sections that matter rather than reading everything at equal depth. The saving survives only when the reader actually knows which sections matter. Skimming a good summary is faster than reading a full document and, for experienced readers, frequently more accurate than trusting memory of it.

Boilerplate and transformation

The last broad category is mechanical transformation: code boilerplate, data reformatting, converting between formats, adapting one document into seven variations for different audiences. The Google trial measured this territory at roughly a fifth of task time. These tasks share a property with drafting: the output is easy to check at a glance, and a silent error is caught before it propagates.

Why email responded and meetings did not

The Microsoft result deserves its own note, because it is the clearest natural experiment in the data. Email is an individual task wearing a collaborative costume. One person can draft faster, triage quicker, and summarise a thread alone, and the calendar records the difference. A meeting requires every participant to change behaviour simultaneously, and no tool on one person's desk can shorten a meeting that eight people must attend. The rule the trial demonstrates generalises: AI saves time on the parts of work you control alone, and cannot touch the parts that require coordination. Budget expectations accordingly.

Where it does not

The police report study is the instructive failure. Report writing looks exactly like the kind of task that should benefit: text shaped, repetitive, well defined. Yet the trial found no meaningful time reduction. The authors point to why the task is harder than it appears: reports are grounded in physical events, must meet legal standards of precision, and the officer's own recollection is the input the tool cannot shortcut. When the bottleneck is thinking and observation rather than typing, a drafting assistant types faster into the same bottleneck.

The same logic applies anywhere the cost of an error is high relative to the minutes saved: legal filings, clinical documentation, financial figures, safety critical instructions. The null result is a reminder that task category is not a sufficient test. The question is not whether a task is text shaped, but whether its true bottleneck is producing words or deciding what is true.

Verification is the hidden cost line

Every time saving from a generative tool carries a verification bill. The honest formula is time saved minus time spent checking, and the second term scales with consequence. Models produce fluent errors, and fluency is exactly what makes the errors expensive to catch: a confident wrong number in a well written paragraph hides better than an awkward one.

A practical discipline is to classify every output by consequence. Internal and low stakes material can be skimmed. Work shared externally but correctable needs its claims and numbers checked. Consequential work, anything with legal, financial, medical or reputational weight, requires verifying every factual statement against a primary source, or not using the tool at all. Much of the reported disappointment with AI time savings comes from applying the strictest rigour to trivial drafts, or trivial rigour to consequential ones.

Build the habit: small loops, measured honestly

The individual version of this advice is deliberately small. Pick two or three recurring tasks from your audit. Build a reusable prompt for each, with the context, constraints and format written in once so you never retype them. Run the loop for a month and log the minutes before and after. The St. Louis Fed's nationally representative survey work found that among US workers using generative AI, roughly a third spend an hour or more a day with it, and self reported savings average around an hour a day among users. Self reported savings reliably overstate measured ones. Your own log is the only number that describes your work.

The organisational ceiling

Individual technique eventually hits a wall that the trials have now measured. Microsoft's 2026 Work Trend Index, drawing on a survey of 20,000 workers across ten countries, reports that organisational factors, culture, manager support and talent practices, account for roughly twice the reported AI impact of individual effort alone, a 67 to 32 percent split. MIT's Project NANDA reached a compatible conclusion from the other direction: its review of 300 enterprise deployments found 95 percent of generative AI pilots delivered no measurable profit and loss impact, and attributed the failure to integration and workflow design rather than model quality.

The individual playbook above works inside an organisation that has not changed. Its ceiling is the set of hours you personally control. Beyond that ceiling, time is saved only when the workflow is redesigned so that the verification, the handoffs and the approvals change too, which is a management project rather than a personal habit.

Limitations

The trials measure months, not years, and behaviour may shift as the tools and the norms around them mature. Most measure one tool on one task family, so the headline percentages should not be read as universal constants. Self reported savings overstate measured savings. Publication favours positive results, which the null studies partially but not fully correct. The numbers above are averages over populations, and your own tasks may sit far from any of the means. Nothing here is a forecast about your job. It is a description of where measured time savings currently live.

The defensible version of the promise is modest. AI saves real time on self contained text work, most reliably for people earlier in their expertise, and loses money on tasks whose errors cost more than the minutes they save. Choose tasks accordingly, verify in proportion to consequence, and measure your own numbers before trusting anyone else's.