The short answer
The most useful ways to use AI at work are not "write my emails". They are specific workflows where the tool does the repetitive thinking, a person checks the result against something real, and the improvement is visible. Below are 25 of them, organised by what you are trying to do: think, write, analyse, communicate, organise, learn and review.
Every one follows the same pattern: task, AI role, human check, proof. The check is not optional. Two-thirds of employees admit relying on AI output without evaluating it, and more than half have made mistakes at work because of it.3 The use cases below tell you exactly what to check in each case.
Two rules apply to all 25. Use the tool your employer has approved, on the account they gave you. And never put personal information about other people, or confidential material, into a consumer tool; New Zealand's Privacy Commissioner is explicit about this.4
How to use this list: find your two or three
What most people use AI for at work is narrow. Among United States employees who use it, writing and editing (51 percent) and search and research (49 percent) dominate; analysis, automation and coding are each under 20 percent.1 The tasks workers say AI assists most are reading documents to gather technical information, preparing reports, and analysing data for trends.2 The list below deliberately goes wider than that, because the value sits where fewer people are looking.
| If most of your day involves | Start with | Then try |
|---|---|---|
| Writing documents, proposals or reports | 5. Bullet points into a first draft | 7. Cut to length and 23. Critique as the reader |
| Meetings | 13. Agenda from the outcome you need | 14. Same-day follow-up and 3. Questions before the meeting |
| Spreadsheets and data | 9. Write and explain formulas | 10. Feedback into themes and 12. Patterns in a text log |
| Admin and coordination | 17. Extract tasks from a long thread | 19. Checklist for a recurring procedure and 18. Process document from a walkthrough |
| Customer or client communication | 14. Same-day follow-up | 8. Adapt tone for the reader and 15. Stakeholder summary of a change |
| Research and reading | 1. Summarise before you read | 20. Unfamiliar material at your level and 2. Compare options against your criteria |
| Planning and decisions | 16. Notes into a plan | 4. Challenge your own assumptions and 25. Gap analysis before hand-over |
Pick the row that describes your week, read the two or three examples it points to, and ignore the rest until one of those is working. Every example below is written in the same shape, so once you have run one you can run any of them.
Stay Useful framework
Task, AI role, human check, proof
A Stay Useful framework. Each use case below is written in this shape so you can copy the shape as well as the example. Risk levels are Stay Useful's judgment: low means a reviewed draft where errors are cheap and visible; medium means the output could reach someone or affect a decision if the check is skipped.
Task
A specific, repeated piece of thinking work. Not "marketing"; "turn customer feedback into three themes with example quotes".
AI role
What the tool does: draft, structure, compare, classify, explain, critique. Never "decide".
Human check
What a person verifies before the output is used, against a source that is not the tool.
Proof
What you note afterwards: time before and after, what changed, what you caught. This is what makes the improvement yours.
Think
1. Summarise a long document before you decide whether to read it
- Task
- Get the substance of a 40-page report, a long contract, or a policy update.
- Old way
- Read it all, skim it badly, or put it off.
- AI-assisted way
- Paste the document (or the public parts of it) and ask for a structured summary: purpose, main findings, what it asks of you, what is missing, and the three passages worth reading in full.
- What you still check
- Read the passages it pointed to. Summaries drop the awkward item; scan the headings yourself for anything it skipped.
- Risk level
- Low, if the document is not confidential. Medium if it is; use an approved tool.
- Try this instruction
- "Summarise this document for someone who has to decide whether it affects their team. Give: purpose, five main findings with page references, any actions requested, anything notably absent, and the three sections I should read in full and why."
- Worth recording
- Time to a decision on the document, before and after, and one item the summary missed that you caught.
2. Compare options against your own criteria
- Task
- Choose between suppliers, software, approaches or candidates' proposals.
- Old way
- A spreadsheet started with good intentions, or a gut call.
- AI-assisted way
- Give the tool your criteria and the public information on each option. Ask for a comparison table, then for the two questions that would most change the ranking.
- What you still check
- Every fact in the table against the source material. The tool will fill gaps with plausible guesses unless told not to.
- Risk level
- Low for the comparison. The decision stays with you.
- Try this instruction
- "Compare these three options against these criteria: [list]. Produce a table with one row per criterion, cite where each fact comes from in the material I gave you, mark anything you could not find as unknown rather than guessing, and finish with the two questions whose answers would most change the ranking."
- Worth recording
- The comparison table with sources, and the two questions it surfaced that changed the decision.
3. Prepare questions before a meeting, interview or supplier call
- Task
- Walk in knowing what to ask.
- Old way
- A few questions jotted down in the lift.
- AI-assisted way
- Describe the meeting, the people, and what you need to leave knowing. Ask for questions grouped by theme, the answers you expect, and what would surprise you.
- What you still check
- Cut the generic ones. Confirm any assertion about the other party against what you actually know.
- Risk level
- Low.
- Try this instruction
- "I am meeting [role] about [topic]. I need to leave knowing [three things]. Give me ten questions grouped by theme, ordered from easy to hard, with a note on what a concerning answer would sound like."
- Worth recording
- The question list, and which question produced the answer that mattered in the meeting.
4. Challenge your own assumptions before someone else does
- Task
- Stress-test a plan, a forecast or a recommendation.
- Old way
- Hope the meeting goes well.
- AI-assisted way
- Paste your plan and ask the tool to argue against it: the strongest objections, the assumptions that carry the most weight, and what would have to be true for it to fail.
- What you still check
- Objections are hypotheses, not facts. Investigate the ones that land; discard the ones that misread the plan.
- Risk level
- Low, and one of the highest-value uses on the list.
- Try this instruction
- "Here is a plan. Act as a sceptical senior colleague. List the five strongest objections, identify the three assumptions the plan depends on most, and describe the most likely way it fails in practice. Do not soften anything."
- Worth recording
- The objection you acted on, what you changed because of it, and what would have happened if you had not.
Write
5. Turn bullet points into a first draft
- Task
- Produce a document, proposal or briefing from notes.
- Old way
- Staring at a blank page, then writing and rewriting the opening paragraph.
- AI-assisted way
- Write the bullets yourself: what you want to say, to whom, and why. Ask for a draft in a specified structure and length. The thinking is yours; the tool does the sentences.
- What you still check
- Every claim traces back to your bullets. Anything the tool added that you did not say gets cut or confirmed.
- Risk level
- Low. Controlled experiments show this is where AI helps most, with the biggest gains for weaker writers.5
- Try this instruction
- "Draft a [document type] for [audience] from these points: [bullets]. Structure: [headings]. Length: about [n] words. Plain language, no jargon, no claims beyond what is in my points. Mark anything you inferred in square brackets."
- Worth recording
- Drafting time before and after, and a before-and-after pair a reviewer could compare.
6. Restructure a messy draft
- Task
- Fix a document that has the right content in the wrong order.
- Old way
- Cut and paste, lose track, start again.
- AI-assisted way
- Paste the draft and ask for a proposed structure with each existing paragraph assigned to a section, and a note on what is missing or repeated.
- What you still check
- That nothing was dropped in the move. Compare paragraph counts.
- Risk level
- Low.
- Try this instruction
- "Here is a draft. Do not rewrite it. Propose a clearer structure with headings, assign each existing paragraph to a heading by its first few words, flag repeated points, and list anything a reader would expect that is missing."
- Worth recording
- The old and new outlines side by side, and the reader feedback on the restructured version.
7. Cut a document to a required length
- Task
- Get 900 words down to 300 without losing the point.
- Old way
- Deleting adjectives and hoping.
- AI-assisted way
- State the audience and what they must take away. Ask for the cut, and for a list of what was removed so you can restore anything essential.
- What you still check
- The list of removals. Meaning drifts when text is compressed; re-read for changed emphasis.
- Risk level
- Low.
- Try this instruction
- "Cut this to no more than 300 words for [audience]. They must come away knowing [two things]. Keep my wording where possible. Then list what you removed, in order of importance."
- Worth recording
- Word counts before and after, and confirmation from the reader that nothing essential was lost.
8. Adapt tone for a different audience
- Task
- Turn an internal technical note into something a customer, a board or a new starter can read.
- Old way
- Rewrite from scratch, or send the technical version and field the questions.
- AI-assisted way
- Provide the original and describe the reader. Ask for a version at their level that keeps every fact and changes nothing substantive.
- What you still check
- That simplification did not become inaccuracy. Technical readers should confirm the simplified version is still true.
- Risk level
- Low to medium, depending on whether the result goes to a customer.
- Try this instruction
- "Rewrite this for [reader description]. Keep every fact and figure unchanged. Replace jargon with plain terms and define anything that cannot be replaced. Do not add reassurance or claims that are not in the original."
- Worth recording
- The two versions, and one comment from the new audience showing they understood it.
Analyse
9. Write and explain spreadsheet formulas
- Task
- Build a lookup, a conditional sum, or a date calculation, or understand one someone else wrote.
- Old way
- Searching forums, or asking the one colleague who knows.
- AI-assisted way
- Describe the columns and the result you want, or paste the formula and ask for a plain-English explanation and its failure cases.
- What you still check
- Test the formula on rows where you already know the right answer, including edge cases. The tool cannot see your data.
- Risk level
- Low if tested. Medium if the result feeds a decision untested.
- Try this instruction
- "Column A has [description], column B has [description]. I want a formula in column C that [result]. Give me the formula, explain each part, and tell me what happens with blanks, duplicates and errors."
- Worth recording
- The formula, the test rows you checked it against, and the time saved against building it by hand.
10. Turn qualitative feedback into themes
- Task
- Make sense of 200 survey comments, support tickets or interview notes.
- Old way
- Reading them all and remembering the loud ones.
- AI-assisted way
- Remove names and identifiers, then paste the comments. Ask for themes, a count per theme, three representative quotes each, and anything that appeared once but matters.
- What you still check
- Read the quotes in context. Check the counts on one theme yourself. Themes are the tool's interpretation; the conclusions are yours.
- Risk level
- Medium. Feedback often contains personal information; redact before pasting, or use an approved tool.
- Try this instruction
- "Here are anonymised comments. Group them into no more than eight themes. For each: a name, how many comments, three verbatim quotes, and what people seem to want. Then list any single comment that raises a serious issue on its own."
- Worth recording
- The theme summary, your spot-check of one theme against the raw comments, and the decision it informed.
11. Classify or tag a list
- Task
- Categorise 300 expense lines, product enquiries or job applications by type.
- Old way
- An afternoon of manual tagging.
- AI-assisted way
- Define the categories with examples. Paste the list in batches. Ask for the category, a confidence rating and a reason for each item.
- What you still check
- Every low-confidence item, and a random sample of the high-confidence ones. Track the error rate; if it is high, refine the category definitions.
- Risk level
- Low for expenses and enquiries. Medium to high for anything about people; do not use AI to screen or rank job applicants without human review.4
- Try this instruction
- "Categories: [name: definition and two examples, for each]. Classify each line below. Output a table: item, category, confidence (high, medium, low), one-line reason. If an item fits none, say so."
- Worth recording
- The error rate on your sample, and how it changed after you tightened the category definitions.
12. Find patterns and anomalies in a text log
- Task
- Spot what changed in a week of incident notes, complaints or shift reports.
- Old way
- Nobody reads the logs until something goes wrong.
- AI-assisted way
- Paste the anonymised log. Ask what recurs, what is new this week compared with the previous one, and what a careful supervisor would want to look at first.
- What you still check
- Go and look at the items it flagged. Pattern-finding produces false positives; the value is in the shortlist, not the verdict.
- Risk level
- Medium if logs contain personal details. Redact first.
- Try this instruction
- "Here are two weeks of shift reports, anonymised. Tell me: what issues recur, what appears in week two that did not appear in week one, and the three entries a supervisor should read first. Quote the entries."
- Worth recording
- The shortlist, which flagged items turned out to matter, and the false positives you discarded.
Communicate
13. Build a meeting agenda from the outcome you need
- Task
- Run a meeting that ends with decisions.
- Old way
- An agenda copied from last time.
- AI-assisted way
- State the decisions the meeting must produce and who is attending. Ask for an agenda with time allocations, the pre-reading each item needs, and the question each item must answer.
- What you still check
- That the agenda fits the time and that each item has a decision, not a discussion.
- Risk level
- Low.
- Try this instruction
- "A 45-minute meeting with [roles] must produce these decisions: [list]. Draft an agenda with minutes per item, the question each item answers, and what attendees should read beforehand."
- Worth recording
- The agenda, and whether the meeting produced the decisions it was built to produce.
14. Draft a follow-up from your own notes
- Task
- Send the summary and next steps after a call, the same day.
- Old way
- Tomorrow, from memory, with something forgotten.
- AI-assisted way
- Type or dictate your notes straight after the call. Ask for a short follow-up: what was agreed, who does what by when, open questions.
- What you still check
- Every commitment against your notes. Do not let the tool add promises you did not make. Names and dates especially.
- Risk level
- Low to medium: it goes to a client or colleague, so read it as they will.
- Try this instruction
- "Turn these call notes into a follow-up email to [name and role]. Sections: what we agreed, actions with owner and date, open questions. Under 150 words. Do not add anything not in the notes."
- Worth recording
- The time between the call ending and the follow-up going out, and any correction the recipient sent back.
15. Write a stakeholder summary of a technical change
- Task
- Tell non-technical people what changed and what it means for them.
- Old way
- Forward the technical email and answer questions for a week.
- AI-assisted way
- Give the technical description and describe the readers. Ask for what changed, what it means for them, what they need to do, and when.
- What you still check
- That the technical owner agrees the simplified version is accurate. Simplification is where errors hide.
- Risk level
- Low to medium.
- Try this instruction
- "Explain this change to [audience] who do not know the technical background. Four short sections: what changed, what it means for you, what you need to do, when. No jargon. Flag anything you are unsure how to simplify."
- Worth recording
- The summary the technical owner signed off, and the questions it saved you from answering afterwards.
Organise
16. Turn notes into a plan
- Task
- Convert a brainstorm, a workshop, or a page of scribbles into a sequenced plan.
- Old way
- The notes stay notes.
- AI-assisted way
- Paste the notes and ask for a plan: phases, tasks with dependencies, what is unclear, and what decisions are needed before it can start.
- What you still check
- The dependencies and the omissions. The tool does not know your constraints; you do.
- Risk level
- Low.
- Try this instruction
- "Turn these notes into a plan. Group tasks into phases, note which tasks depend on which, list the decisions that must be made before phase one starts, and list anything in the notes that is too vague to plan."
- Worth recording
- The plan, the constraints you added that the tool could not know, and whether the phases held.
17. Extract tasks from a long thread
- Task
- Find everything anyone committed to across 60 messages.
- Old way
- Scroll, search, miss one.
- AI-assisted way
- Paste the thread with external names replaced by roles. Ask for actions with owner, date and the message they came from, plus decisions and contradictions.
- What you still check
- Read the messages it cites for each action. Confirm owners actually agreed.
- Risk level
- Medium if the thread contains client or personal information. Redact, or use an approved tool.
- Try this instruction
- "From this thread, list every action with owner, due date and the message it came from. Then list decisions made, and any place where a later message contradicts an earlier one."
- Worth recording
- The action list against the thread, with any commitment the tool invented or missed.
18. Write a process document from a walkthrough
- Task
- Document how a task is actually done so someone else can do it.
- Old way
- It lives in one person's head until they leave.
- AI-assisted way
- Dictate or type a walkthrough as you do the task, including the small checks and the exceptions. Ask for a numbered procedure with prerequisites, steps, checks and what to do when things go wrong.
- What you still check
- Have someone else follow the document without your help. Every step they stumble on is a gap.
- Risk level
- Low.
- Try this instruction
- "Turn this walkthrough into a procedure a new colleague could follow alone. Sections: purpose, prerequisites, numbered steps with a check after each risky step, common problems and what to do. Ask me about anything ambiguous rather than guessing."
- Worth recording
- The document, and how far the second person got through the task on their own before asking.
19. Build a checklist for a recurring procedure
- Task
- Stop forgetting the same step in a monthly close, an onboarding, or a site inspection.
- Old way
- Remembering, mostly.
- AI-assisted way
- Describe the procedure and its failure modes. Ask for a checklist ordered by sequence, with the check that confirms each step is complete.
- What you still check
- Run it once against the real procedure. Compare with the official procedure where one exists; the checklist supports it, it does not replace it.
- Risk level
- Low for administrative procedures. Medium for safety-related ones: the source of truth is the official procedure.
- Try this instruction
- "Create a checklist for [procedure]. Order it by when each step happens. For each step, add how to confirm it is done. Include the steps people most often forget: [list]."
- Worth recording
- The checklist, the step it caught in its first month, and what that would have cost.
Learn
20. Understand unfamiliar material at your level
- Task
- Get to grips with a regulation, a technical concept, or another team's domain.
- Old way
- Reading the same paragraph three times.
- AI-assisted way
- Paste the material and describe what you already know. Ask for an explanation at your level, the terms defined, and the questions you should be able to answer afterwards.
- What you still check
- Test yourself on those questions. Confirm anything you will rely on against the primary source; the explanation is a scaffold, not an authority.
- Risk level
- Low, with one caveat: research finds the biggest gains from AI go to novices, and also that the youngest workers are losing ground where AI replaces the learning tasks.6 Use it to learn faster, not to skip learning.
- Try this instruction
- "Explain this to someone who knows [what you know] but not [what you do not]. Define each technical term when it first appears. Finish with five questions I should be able to answer if I understood it."
- Worth recording
- The questions you could answer afterwards that you could not before, and where the scaffold was wrong.
21. Generate practice questions for a certification or a new role
- Task
- Prepare for an exam, an interview or the first week of a new job.
- Old way
- Re-reading notes.
- AI-assisted way
- Give the syllabus or role description. Ask for practice questions of increasing difficulty, and after you answer, for feedback on what you missed.
- What you still check
- Verify any answer key against the official material. The tool can be confidently wrong about specifics.
- Risk level
- Low.
- Try this instruction
- "Here is the syllabus. Write ten practice questions from easy to hard, covering different sections. Wait for my answers, then tell me what I got wrong and which part of the syllabus to revisit."
- Worth recording
- Practice results over time, and any answer key error you found and corrected.
22. Learn new software or a new process step by step
- Task
- Do something for the first time in a system you do not know.
- Old way
- Clicking around, or a 40-minute video for a two-minute task.
- AI-assisted way
- Describe the software, the version, and what you are trying to do. Ask for steps, what you should see after each, and what commonly goes wrong.
- What you still check
- That the steps match your screen. Interfaces change; if a step does not match, stop and check the official help rather than improvising.
- Risk level
- Low for reading and reporting tasks. Medium for anything that changes or deletes data: read the step before you do it.
- Try this instruction
- "I am using [software and version]. I want to [goal]. Give numbered steps, what I should see after each one, and what to do if I do not see it. Tell me before any step that changes or deletes data."
- Worth recording
- The steps that matched your screen and the one that did not, with what the official help said instead.
Review
23. Have your draft critiqued as the intended reader
- Task
- Find the weaknesses in your own work before the reader does.
- Old way
- Send it and wait.
- AI-assisted way
- Paste the draft and describe the reader precisely: their role, what they care about, what they are sceptical of. Ask for the critique that reader would give, and the three changes that would matter most.
- What you still check
- The critique is a simulation, not a verdict. Act on what rings true; ignore what misreads the reader.
- Risk level
- Low.
- Try this instruction
- "Read this as [specific reader: role, priorities, likely objections]. Tell me where it loses your attention, what you do not believe, what question you are left with, and the three changes that would most improve it for you."
- Worth recording
- The critique, the three changes you made, and the reader reaction to the revised draft.
24. Check a document against a set of requirements
- Task
- Confirm a proposal answers every question in the brief, or a report includes every required section.
- Old way
- Two documents open side by side, and something missed on page 14.
- AI-assisted way
- Paste the requirements and the document. Ask for a table: each requirement, where it is addressed, and whether the coverage is complete, partial or absent.
- What you still check
- Every "complete" against the text. The tool is good at finding absence and worse at judging adequacy.
- Risk level
- Low to medium. For compliance or legal requirements, this supports a qualified reviewer; it does not replace one.
- Try this instruction
- "Requirements: [list]. Document: [text]. Produce a table with each requirement, the section and quote that addresses it, and a rating: complete, partial, or not addressed. Do not rate anything complete without quoting it."
- Worth recording
- The coverage table, and any "complete" you overturned on reading the text.
25. Run a gap analysis between what was asked and what was delivered
- Task
- Before you hand something over, know where it falls short of the brief.
- Old way
- Finding out from the client.
- AI-assisted way
- Paste the original brief and your deliverable. Ask what was asked for and not delivered, what was delivered but not asked for, and where the deliverable answers a different question from the one posed.
- What you still check
- The list is a prompt for your judgment, not a pass or fail. Some gaps are deliberate; make sure the ones that remain are.
- Risk level
- Low, and one of the cheapest ways to look thorough.
- Try this instruction
- "Brief: [text]. Deliverable: [text]. List: what the brief asked for that the deliverable does not address; what the deliverable includes that was not asked for; and any place the deliverable answers a different question from the brief."
- Worth recording
- The gaps you closed before hand-over, and the ones you left open on purpose, with the reason.
How to pick the one to start with
You do not need 25 workflows. You need one that works, that you have checked, and that you can point to. Three questions narrow the list quickly.
- Which of these do you do most often? Frequency is what turns a saving per task into a saving per week. Population surveys find modest average time savings largely because most people use AI occasionally, which our beginner's guide explains. Daily tasks change that arithmetic.
- Where is a mistake cheap and visible? Start in the low-risk rows. Learn where the tool fails on your material before you rely on it anywhere a mistake costs something.
- What would someone notice? Pick a task whose improvement a colleague or manager would see. The record you keep of it is the beginning of your evidence; our guide on how to stay valuable at work as AI gets better explains why that matters.
Then run it for two weeks, tighten the instruction each time it falls short, and write down what changed. That is the whole method, and our beginner's guide to using AI at work walks through it step by step.
How to avoid making things worse
- Do not send unreviewed output. Research affiliated with BetterUp and Stanford found 40 percent of United States desk workers had received AI-generated work in the previous month that looked fine and lacked substance, costing recipients nearly two hours each time, and 42 percent thought less of the sender.7 Every use case above has a check step for a reason.
- Do not paste what you would not email to a stranger. Personal information about other people and confidential material stay out of consumer tools.4 Our guide on what happens to the data you enter explains the consumer and business distinction.
- Do not use it to skip learning. The evidence that AI helps novices most is real.56 So is the evidence that the tasks novices learned on are the ones being handed to the tool. Use it to learn faster and check yourself, not to avoid understanding.
- Do not let it decide about people. Classification and screening of applicants, customers or staff without real human review is where privacy regulators and provider terms both draw a line.4
Related questions
Do these work with any AI tool?
Yes. The instructions are written for a general-purpose assistant and the pattern, task, AI role, human check, proof, does not depend on the product. Use whichever tool your employer has approved; the data protections attach to that account, not to the brand.
Which of the 25 gives the biggest return?
For most people, the one they do most often. Writing from bullet points (5), summarising before reading (1), and extracting actions from threads (17) are the most broadly applicable. Challenging your own assumptions (4) and the gap analysis (25) are the ones most people never think of and that make the most visible difference to the quality of what they hand over.
Is it cheating to use AI for these?
Using a tool to draft, structure and check your work is not cheating. Presenting unreviewed output as your considered work is where the problem starts, both ethically and practically, because it is where the errors get through. Do the check, own the result, and say so if asked. Our guide on whether you can use ChatGPT at work covers permission and disclosure.