The short answer
To use AI well at work, start with one repetitive piece of thinking work that is low-risk if the tool gets it wrong. Give the tool the context a new colleague would need and a clearly defined job. Check the output against something real before you use it. Then tighten the instruction, save what worked, and write down what improved.
That sequence matters more than any prompt. Controlled studies show large gains when AI is applied to tasks inside its competence, such as drafting and summarising, and real losses when it is applied to tasks just outside it: in one experiment, consultants using AI on a task designed to trip it up were 23 percent worse than those without it.6 Choosing the task well is most of the skill.
This guide gives you an eight-step way of working, a method for briefing the tool, and worked examples from ordinary jobs, not only technical ones.
Where most people are starting from
You are not behind. In Gallup's May 2026 survey of over 22,000 United States employees, 52 percent said they used AI in their role at least a few times a year, 30 percent a few times a week or more, and 15 percent daily.1 Use is concentrated in roles that can be done remotely and among leaders; among people in roles that cannot be done remotely, only about a third used AI at all in late 2025.2 Pew Research found in September 2025 that 21 percent of United States workers said at least some of their work was done with AI, up from 16 percent a year earlier, while 65 percent said they did not use it much or at all.3 Among the minority who used AI chatbots for work in Pew's earlier survey, the most common uses were finding information, editing written content and drafting documents.4
What people use it for is narrow. Among Gallup's AI users, writing and editing (51 percent) and search and research (49 percent) dominated; data analysis, automation and coding were each under 20 percent.1 The pattern is the same in the Federal Reserve Bank of St. Louis survey, where the most AI-assisted tasks were reading documents to gather technical information, preparing reports and analysing data for trends.5
What the evidence says AI is good and bad at
Before the method, it helps to know what the research actually shows, because it explains why task selection matters so much.
Inside its competence, the gains are large
- In a preregistered experiment with 453 college-educated professionals doing realistic writing tasks, ChatGPT cut average time by 40 percent and raised output quality by 18 percent, with the biggest gains for the weakest writers.8
- When a software company gave an AI assistant to around 5,000 customer support agents, issues resolved per hour rose by about 14 percent on average, and by 34 percent for the newest and least skilled agents, with little change for the most experienced.9
- In a field experiment with 758 consultants at Boston Consulting Group, those given GPT-4 on tasks inside the tool's capabilities completed 12.2 percent more tasks, 25.1 percent faster, at quality rated around 40 percent higher.67
Outside its competence, it makes people worse
- The same consulting experiment included a business problem designed to sit just outside the tool's abilities. On that task, the consultants using GPT-4 performed 23 percent worse than the control group. The researchers called the boundary a "jagged frontier": it is uneven, invisible, and moves.67 That was one task with one 2023 model, so the number will not transfer to your work; the shape of the finding does.
- A randomised study of 16 experienced open-source developers working on their own mature codebases found they took 19 percent longer with AI tools in early 2025, while believing they had been sped up by 20 percent. The researchers stress this is one setting at one moment, and their 2026 follow-up suggests tools have improved, but the gap between felt and measured benefit is the lasting lesson.10
Across a whole job, the gains so far are modest
Population-level self-reported time savings are small, because most people use AI on a small slice of their week. Danish workers in exposed occupations reported saving about 2.8 percent of their hours,11 and United States workers reported savings equal to about 1.6 percent of all work hours in 2025.5 The gap between "40 percent faster on this task" and "2 percent of my week" is the gap between using AI on a task and redesigning work around it. Closing it is what the rest of this guide is about.
The Stay Useful workflow: eight steps
This is a Stay Useful framework, built from the evidence above and from the guidance New Zealand's privacy regulator and business ministry give on using these tools responsibly.1415 It is a way of thinking, not a list of prompts. Once it is habit, each step takes seconds.
Stay Useful framework
Task, AI role, human check, proof
Find repetitive cognitive work
Look for thinking tasks you do more than once a week that follow a pattern: summarising, drafting, reformatting, explaining, comparing, classifying. Not the hardest part of your job. The most repeated part.
Separate low-risk from high-risk tasks
Two questions: how bad is a confident mistake, and how sensitive is the information involved? Start where both answers are "not very". The matrix below shows where each combination lands.
Decide what context the AI needs
The tool knows nothing about your organisation, your audience or your standards. Decide what a capable new colleague would need to do this task: the purpose, the audience, an example of good output, the constraints. Then decide what it must not be given, such as personal or confidential information.
Give the AI a clearly defined job
One task, stated plainly: what to produce, for whom, in what form, at what length, with what to avoid. Vague requests get generic output. A defined job gets a usable draft.
Review the output
Every fact, figure, name, date and quotation gets checked against a source. Every omission gets noticed. Every recommendation gets tested against what you know. You are the accountable one; the tool is not.
Improve the workflow
Where did the output fall short? Usually it was missing context or an unstated standard. Add it to the instruction and run again. Two or three iterations turn a mediocre first attempt into a reliable one.
Save and reuse what works
Keep the instruction that worked, with the context it needs, somewhere you can find it. A personal library of five reliable instructions is worth more than a hundred clever prompts you cannot remember.
Capture proof of the improvement
Note what the task took before and after, what quality changed, and what you checked. This is what turns "I use AI" into evidence you can show a manager or a future employer.
Step two in detail: the risk matrix
| Low-sensitivity information | Sensitive or confidential information | |
|---|---|---|
| Low cost of error (a person reviews it before use) | Start here. Drafts, summaries of public material, reformatting, explanations, practice questions, first-pass structure. | Only on an approved enterprise tool, or after removing the sensitive parts. Redact first; the drafting is still low-risk. |
| High cost of error (used directly, affects money, safety, or a person) | AI can help you think, but a qualified person owns the result. Use it to prepare questions, check your reasoning, or produce a first version that gets full review. | Not in a general-purpose chat tool. This is the territory of approved, assessed systems with human accountability, if at all. |
The bottom-right cell is where the cautionary cases live: source code pasted into a public tool, fabricated legal citations filed in court, personal information entered into a service that retains it. Our guides on whether you can use ChatGPT at work and on what happens to the data you enter cover that territory in detail. The short version from New Zealand's Privacy Commissioner: do not put personal or confidential information into a generative AI tool unless the provider has confirmed it is not retained or disclosed.14
Step four in detail: how to brief an AI
A good brief looks like a good brief to a person. It has five parts, and leaving any of them out is the usual reason output disappoints.
The anatomy of a clear instruction
- Role and purpose
- "You are helping an operations coordinator prepare the minutes of a weekly team meeting for people who did not attend."
- The task
- "Turn the notes below into minutes. List decisions made, actions with an owner and a date, and open questions."
- The inputs
- "Here are my raw notes: [notes, with names of external people replaced by their role]."
- The output
- "Use three headings: Decisions, Actions, Open questions. Keep each item to one line. Plain language. Under 250 words."
- Constraints and checks
- "Do not add anything that is not in the notes. If something is unclear, list it under Open questions rather than guessing. Mark anything you inferred."
The last line is the one beginners leave out and the one that matters most. Telling the tool to flag uncertainty rather than fill gaps is what New Zealand's business guidance calls for when it recommends asking for sources and step-by-step reasoning so you can see where errors creep in.15
Step five in detail: what to check
The most common failure in workplace AI use is not a bad prompt. It is an unchecked output. In the KPMG and University of Melbourne study, 66 percent of employees globally said they had relied on AI output without evaluating its accuracy, and 56 percent had made mistakes in their work because of AI.12 Regulators describe the failure mode precisely: generative tools "produce confident errors of fact and logic", and having a human nominally in the loop "may not be enough" because people stop looking.14
- Facts and figures. Every number, date, name and quotation gets traced to a source you can open. If the tool cannot point to one, treat the claim as unverified.
- Omissions. What did the tool leave out? Summaries drop the awkward item. Compare against the original for anything that matters.
- Invented specifics. Product features, prices, policies, legal provisions, citations. These are where fabrication does the most damage.
- Fit. Does it say what you would say, to this audience, in this organisation? Generic is a defect.
- Policy. Nothing personal or confidential went in, and the output is not being used for a decision about a person without real human review.
Worked examples from ordinary jobs
These describe how the workflow applies in roles most people actually have. They are illustrations of method, not measured case studies, and the review step is the point of each one.
Administrator: meeting notes into minutes
- Task
- Produce minutes and an action list from a weekly meeting.
- Old way
- Forty minutes after the meeting turning scribbled notes into a formatted document, usually the next day.
- AI-assisted way
- Paste the raw notes, with external people's names replaced by roles, using the brief above. Get a structured first draft in under a minute.
- Human review
- Check every decision and action against the notes. Confirm owners and dates. Add anything the tool dropped. Remove anything it inferred that was not said.
- Result
- Minutes go out the same day, in a consistent format, and the coordinator's time goes into chasing the actions rather than typing them.
Salesperson: preparing for a discovery call
- Task
- Walk into a first meeting with a prospect knowing what to ask.
- Old way
- Ten minutes skimming the company website, then improvising.
- AI-assisted way
- Give the tool the prospect's public website text and your product's one-paragraph description. Ask for the ten questions most likely to reveal whether they have the problem you solve, grouped by theme, and the three objections you should expect.
- Human review
- Cut the generic questions. Check anything the tool asserted about the prospect against the source. Never let it invent claims about your own product or prices.
- Result
- A sharper conversation. The customer relationship management system stays out of the tool entirely.
Accountant: a plain-English client note
- Task
- Explain to a small-business client what a change in a tax rule means for them.
- Old way
- An hour writing, rewriting, and worrying about the tone.
- AI-assisted way
- Paste the relevant public guidance and your bullet-point conclusions. Ask for a 200-word client note at the reading level of a busy owner, with no client details in the prompt.
- Human review
- Check every figure and threshold against the primary source. Recompute anything numeric yourself. Confirm the note says what you concluded, not what the tool concluded. You sign it.
- Result
- Clearer notes in less time, with the professional judgment exactly where it was.
Teacher: differentiating a worksheet
- Task
- Adapt one reading comprehension exercise for three reading levels.
- Old way
- An evening writing three versions, or using one version for everyone.
- AI-assisted way
- Paste the passage and questions. Ask for two additional versions at specified reading ages, keeping the same learning objective, plus an answer key for each.
- Human review
- Check facts and curriculum fit. Read each version as a student would. No student names, grades or assessment data go anywhere near the tool.16
- Result
- Three versions in the time it took to write one, and marking that is consistent across them.
Recruiter: from hiring brief to job advertisement
- Task
- Turn a manager's rough brief into a clear advertisement.
- Old way
- Adapting an old advertisement for a different role and hoping it fits.
- AI-assisted way
- Paste the brief. Ask for a draft with responsibilities, must-have and nice-to-have requirements separated, inclusive language, and a plain description of how the role is assessed.
- Human review
- Check for requirements the manager did not ask for. Check for language that excludes people. Check legal requirements for your jurisdiction. Candidate information never enters the tool, and AI is not used to rank or screen applicants without human review.14
- Result
- Better advertisements faster. Screening decisions stay with people.
Project manager: a long thread into a status update
- Task
- Extract decisions, risks and open items from a 60-message email thread.
- Old way
- Reading the whole thread twice and still missing something.
- AI-assisted way
- Paste the thread with client names replaced. Ask for decisions, risks with severity, open questions with owners, and anything that contradicts something said earlier.
- Human review
- Cross-check every date and owner against the plan. Read the messages the tool flagged as contradictory. Confirm the tone of the summary matches what was actually agreed.
- Result
- A status update that reflects the thread rather than the last message in it.
Operations supervisor: a voice note into a shift handover
- Task
- Produce a written handover at the end of a shift on a site or floor.
- Old way
- A rushed paragraph, or nothing written at all.
- AI-assisted way
- Record a two-minute voice note walking the floor. Transcribe it and ask the tool to structure it as a handover: what happened, what is outstanding, what the next shift must check first, any safety items.
- Human review
- Facts about incidents come from you, not the tool. Confirm safety items against the actual procedure. Anything you are unsure about is flagged as such.
- Result
- Handovers that exist and are consistent, from people whose job is not typing.
Graduate or new starter: learning the job faster
- Task
- Understand the jargon, acronyms and unwritten rules of a new workplace.
- Old way
- Nodding in meetings and searching later.
- AI-assisted way
- Keep a running list of terms you did not understand. Ask the tool to explain each one in the context of your industry, and to suggest how an experienced person would approach a task you have been given.
- Human review
- Confirm with a colleague before acting on anything organisation-specific; the tool does not know your organisation. Do the task yourself first, then compare.
- Result
- Faster learning without outsourcing the learning. The evidence that AI helps novices most is also a warning: the point is to build the skill, not to skip it.9
The mistakes beginners make
- Pasting first, thinking later. Personal and confidential information goes into a consumer tool because it was convenient. See the risk matrix above and our guide on what happens to the data you enter.
- Asking the tool to check itself. "Verify your answer" is not verification. The tool will confidently confirm its own error. Verification means a source you open.
- Mistaking fluency for accuracy. Polished prose reads as correct. Two-thirds of employees admit relying on output without checking it.12
- Producing "workslop". Research affiliated with BetterUp and Stanford found 40 percent of United States desk workers had received AI-generated work that looked fine but lacked substance in the previous month, costing the recipient nearly two hours each time, and 42 percent saw the sender as less trustworthy.17 Passing along unreviewed output damages you, not the tool.
- Using it on the wrong task. The consulting experiment is the lesson: the same tool that lifted quality by around 40 percent on tasks inside its competence left people performing 23 percent worse on a deliberately difficult one.6 Start where a wrong answer is cheap and visible.
- Not saving what worked. Every good instruction reinvented from scratch is time wasted. Keep a library.
- Hiding it. In the KPMG study, 57 percent of employees said they hid their AI use and presented AI-generated work as their own.12 Openness costs less than people fear, and it is the only way to build the evidence in step eight.
Safety, in one paragraph
Use the tool your employer has approved, on the account they gave you. Do not enter personal information about other people, or confidential material, into a consumer tool.14 Ask the tool for its sources and check them.15 Treat every output as a draft. Never use AI output about a person to make a decision with a legal or material effect on them without real human review. If you are unsure whether something is allowed, ask before pasting; our guide on using ChatGPT at work has a checklist for exactly that moment.
Related questions
Do I need to know how to code to use AI at work?
No. Among Gallup's AI users, only 16 percent used it for coding; writing, editing and research dominated.1 The skills that matter are the ones in this guide: choosing a suitable task, briefing clearly, and checking carefully.
Which AI tool should I use?
The one your employer has approved, if there is one, because the data protections attach to that account. If there is not, ask. The general-purpose assistants from the major providers are broadly similar for the tasks in this guide; the method transfers between them. Our 25 ways to use AI to make your job easier gives specific instructions you can adapt to any of them.
How do I write a good prompt?
Write it as a brief to a capable new colleague: purpose, task, inputs, output format, and what to do when unsure. The example in step four above is a template. The word "prompt" makes people think of a magic phrase; a clear brief is the whole trick.
What if the AI makes mistakes?
It will. Regulators describe "confident errors of fact and logic" as a known property of these tools.14 Design for it: use AI where a mistake is cheap and visible, check every specific against a source, and never let output reach anyone without a person having read it.
How do I know whether I am doing this well?
Check the behaviour rather than the frequency. Our AI skills checklist of seven things to check at work turns the steps above into seven questions you answer about one real piece of work you did this week, including the one people skip: whether you counted the review and correction time.
My manager has told me to use AI. Where do I start?
With one recurring task and a written question about what is approved, not with a tour of the tools. Where to start when your boss wants you to use AI has a message you can adapt and a two-week experiment you can report on.
Will using AI make me look lazy or replaceable?
Nearly half of desk workers in a 2024 Slack survey worried about exactly that.18 The evidence points the other way when AI use comes with checking and a record of results: the risk to jobs concentrates where AI substitutes for people, and the people who understand the tools well enough to design the checks are the ones deciding how it is used. See our guide on how to stay valuable at work as AI gets better.