Ten minutes came off a single writing task. That's the finding from Shakked Noy and Whitney Zhang's MIT working paper of March 2023, which gave 444 college-educated professionals an occupation-specific writing assignment and let half of them use ChatGPT: the control group averaged 27 minutes, the treatment group finished about 10 minutes faster, and graders scored the assisted work higher.
That result has a shape, and the shape is the useful part. It was a defined task, with a brief, producing a document of a known type. Most of a working week doesn't look like that, which is why the same tool that saves a colleague an hour a week does nothing at all for you.
People who try this and quietly stop rarely complain about the quality of what came back. Their complaint is that getting to something usable took as long as doing the work themselves. That's almost always a task-selection problem rather than a prompting problem.
Which work tasks are worth a prompt?
A work task is worth a saved prompt when it repeats, produces the same shape of output every time, and takes longer to start than to finish. All three conditions matter, and dropping any one of them is what turns a promising idea into a tool nobody opens twice.
Repetition is what pays for setup. Writing a good prompt for your weekly status report costs maybe fifteen minutes once. Produce that report 45 times a year and the fifteen minutes has paid for itself before the end of the month. Produce it twice a year and you've spent fifteen minutes to save eight.
Format stability is the condition people skip. When the output has a predictable structure, you can specify that structure once and stop re-explaining it. Meeting notes have owners and dates. Status reports have progress, risks and asks. A prompt that names those slots gets you something you can skim rather than something you have to reorganize.
The third condition is about friction, and it's the one that decides how much you actually feel the difference. Some tasks are hard to start and easy to finish: you know what the email needs to say, you just can't get the first line out. A prompt is worth most there, because it's dissolving the blank page rather than doing the thinking.
Against those three sits one cost that quietly cancels the saving. Output you have to verify claim by claim costs review time, and review time is slower than writing time when the stakes are real. Anything built on facts you can't check at a glance belongs in a different workflow, which is the subject of our guide on using AI for research without made-up facts.
Who saves the most time?
The gains skew hard toward whoever is newest to a task. In Generative AI at Work, the NBER study by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, customer support agents with access to an AI assistant resolved 14% more issues per hour on average, and novice and low-skilled agents resolved 34% more. Experienced agents improved barely at all, because what the tool surfaced was mostly what they already knew.
That split holds up away from support queues. A manager writing their first performance review gains a structure they didn't have. A manager who has written two hundred of them gains very little on the writing itself, and picks up whatever is left in the admin around it: the summarizing, the chasing, the reformatting for the person above them.
Which suggests a rule for anyone senior enough to feel unimpressed. Point these prompts at the packaging rather than the substance, because packaging is the part where you have no accumulated speed advantage.
Both studies come with a caveat worth stating plainly. They measured particular tasks, in particular workplaces, in 2023, on models that have since been replaced twice over. They're good evidence that the effect is real and unevenly distributed, and they are not a forecast for your week.
The email you rewrite every week
Correspondence is the highest-frequency writing most people do, and almost all of it falls into a handful of recurring types.
Chasing something you're owed is the clearest case. You've asked for a quote, a signature or a decision, nothing has come back, and the third nudge is the one where tone gets difficult. The follow-up email prompt takes the thread and the ask and gets you a version that stays polite without going limp. Give it the number of days since your last message, because the right tone for four days differs from the right tone for three weeks.
Deadline reminders to people who don't report to you have the same problem in a smaller package, and the reminder email prompt exists for exactly that. So does the out-of-office message, which is a once-a-year task everyone rewrites badly at 6pm the night before a holiday.
The one worth saving even if you save nothing else is the tone adjustment prompt. You've written the email. It's accurate and it's slightly too blunt, or slightly too apologetic, and fixing that yourself means rereading it four times. Paste it, name the tone you want, and read one alternative. That's a sixty-second loop replacing a ten-minute one.
Subject lines are the smallest item here and the one with the best ratio, since the email subject line prompt gives you five options for a message you've already written and you pick in seconds.
Meetings, before and after
Meeting admin is the purest example of stable-format work, which is why it's where most people's first real saving shows up.
Say a 47-minute call produces a transcript of around 6,800 words. Reading it back to extract what was decided takes fifteen minutes and you'll still miss something. The meeting summary prompt turns that into decisions, owners and dates, and your job shrinks to checking the owner names and dates against the transcript, which is a two-minute pass when there are five of them. Verifying five specific facts is a genuinely different activity from reading 6,800 words.
Before the meeting, the meeting agenda prompt is worth more than it looks. Its real output is the timebox rather than the agenda: forced to allocate minutes per item, you find out the meeting needs 25 of them rather than 60.
For daily updates, the standup update prompt turns yesterday's scattered notes into three lines. This is a small saving repeated 200 times a year, which is the profile of most genuine productivity gains and the reason they're easy to dismiss.
Documents you'd rather not start
Long-document work splits into reading and producing, and AI helps with both in different ways.
On the reading side, the document summary prompt handles the 40-page vendor PDF you need one answer from. The instruction that matters is to answer only from the supplied document, which is standard guidance in Anthropic's prompt engineering docs and in OpenAI's as well. Without it you get a plausible summary of what such a document usually says.
On the producing side, three prompts cover most of what lands on a desk. The project status report prompt takes your scrappy notes and returns progress, risks and asks in the same order every week, which is what makes a report skimmable to the person reading eleven of them. The executive summary prompt compresses a finished document for a reader who won't open it. The internal memo prompt handles announcements where the hard part is deciding what to leave out.
Process documentation deserves a mention because it's the task people postpone for years. The SOP prompt turns "how I do this" said out loud into numbered steps someone else can follow, and it's most valuable when written by the person leaving the role.
Planning your own week
Personal planning is where prompts help least at the level of output and most at the level of getting started, which is a distinction worth being honest about.
Prioritizing is not a task AI can do for you, because it needs judgment about consequences only you can see. What the task prioritization prompt does is make you list everything and state each deadline, and the ordering that falls out is often obvious once the list exists. The decision matrix prompt works the same way for choices between options: naming the criteria is the work, and the tool makes you name them.
Friday afternoon has the weekly review prompt, which asks the questions you'd skip if left alone. Ten minutes, and it's the cheapest planning habit on this page.
What each prompt still needs from you
Every prompt here has an input cost, and the ones that fail in practice are the ones where that cost was never counted.
| Recurring task | Prompt | What you supply | Skip it when |
|---|---|---|---|
| Chasing an unanswered ask | Follow-up email | The thread, the ask, days elapsed | Your reply is one line |
| Softening a written message | Tone adjustment | Your draft, the tone you want | You already like the draft |
| Writing up a call | Meeting summary | Transcript or detailed notes | Nobody reads the notes |
| Extracting one answer from a long PDF | Document summary | The document itself | The answer is in the contents page |
| The weekly report | Project status report | Raw notes, blockers, dates | The project has one reader who sits next to you |
| Documenting a process | SOP | Every step, including the obvious ones | You're the only person who'll ever run it |
| Ordering a messy workload | Task prioritization | The full list plus deadlines | The list has four items |
Reading down the third column is the honest test. If supplying the input takes longer than producing the output by hand, the prompt is costing you time no matter how good the response looks.
Where AI won't save you anything
A two-line Slack reply doesn't need a tool. Reaching for one anyway is a habit worth catching early, since the overhead is invisible and it accumulates across a day.
Work that rests on facts you can't verify is the expensive case. Output that reads well and contains one wrong figure costs more than the blank page did, because you now have to check something that looks finished.
Then there's the context trap, where the input is larger than the deliverable. Explaining six months of project history to get a three-sentence update is a bad trade, and no amount of prompt craft fixes it.
Judgment work with real consequences is the last of them: a personnel decision, a legal question, a call about someone's pay. AI can structure your thinking around the edges of those, and it cannot make them for you. Where information is confidential, keep it out of reusable prompt templates entirely and use placeholders instead, a discipline covered in our guide to building a prompt library for teams.
Making a saving repeat
A prompt used once is a novelty. The gains compound only when reuse becomes automatic, and that depends on where you keep the thing, not on how well you wrote it.
Save anything you've run three times. Retyping from memory produces a slightly different prompt each run, and a slightly different prompt produces a differently shaped output that you then re-edit, which is where the saving quietly goes. Keep the saved version wherever you can find it in ten seconds, whether that's a notes app, a pinned document or your tool's own saved-prompt feature. Google's prompting guidance makes the same point about iterating on a prompt and keeping what worked rather than starting fresh each time.
If your prompts are worth sharing with colleagues, the ownership and review questions that come with that are covered separately in the team prompt library guide. If your prompts aren't producing what you want in the first place, the fix is upstream of this page, in prompt engineering for beginners.
Role-specific collections exist for the two jobs where these tasks cluster hardest: prompts for managers, which is heavy on reports and one-to-ones, and prompts for customer service reps, where the response templates carry the volume. The full set sits in the prompt library if your week has a shape neither of those matches.
Sources
- Shakked Noy and Whitney Zhang, Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence, MIT working paper, 2 March 2023: the 27-minute control average and the 10-minute reduction cited above. The paper notes it was not peer reviewed at the time of that draft.
- Erik Brynjolfsson, Danielle Li and Lindsey R. Raymond, Generative AI at Work, NBER Working Paper 31161: the 14% average and 34% novice figures for customer support productivity.
- OpenAI, Prompt engineering: official guidance on specificity and supplying reference text.
- Anthropic, Be clear and direct: guidance on stating the task and the output format explicitly.
- Google AI, Prompting strategies: documentation on iterating prompts and constraining output format.
The tell that this is working isn't a dramatic week. It's noticing, on a Thursday, that the status report you used to leave until Friday evening is already sitting in drafts.
