The short answer
Using AI often does not tell you whether it is improving your work. Frequency is the easiest thing to see and the easiest thing to report, and it answers none of the questions that matter: whether the output was good enough, whether you spent longer checking it than you saved, and whether anybody else could repeat what you did.
So check the behaviour instead of the habit. Seven things are observable in ordinary work: choosing an appropriate task, checking permission and information sensitivity, giving a clear brief, checking the output, saving a workflow worth reusing, measuring the full effort including review, and explaining the improvement with evidence.
This is a reflection checklist, not a test. It has no score and certifies nothing. Use it against one real piece of work you did this week.
The seven-point checklist
Pick one task you used AI for recently. Answer these about that task, not in general.
Seven checks, on one real piece of work
- Was this a sensible task to hand to AI at all?
- Did I check whether the tool and the information were permitted?
- Did I give it the context and the constraints a new colleague would have needed?
- Did I check the output against something real before using it?
- Did I keep the instruction or the workflow so I can reuse it?
- Did I count the review and correction time, not just the drafting time?
- Could I show someone what improved, and how I know?
What each check looks like in ordinary work
Stay Useful framework
The seven checks, with an ordinary example of each
A Stay Useful reflection checklist, not a scored test and not a validated instrument. There is no threshold to reach and nothing here certifies competence or permission to use any tool.
Choosing an appropriate task
Good candidates are repetitive, information-heavy and cheap to get wrong first time. Bad candidates sit just outside what the tool can do, which is where measured performance goes backwards: in a field experiment with 758 consultants, those using AI inside its abilities completed 12.2 percent more work and 25.1 percent faster, while on a task built to sit just outside them they scored 23 percent worse than colleagues without it.1 Example: turning your rough meeting notes into a structured summary is a reasonable first task. Deciding which supplier to drop is not.
Checking permission and information sensitivity
Two questions, not one: is this tool approved where you work, and is this information suitable for it. New Zealand's Privacy Commissioner advises against putting personal or confidential information into a generative AI tool unless the provider has explicitly confirmed it is not retained or disclosed.2 Removing names does not settle it, because the surrounding detail often identifies a person anyway. Example: summarising a public policy document is not the same as summarising a colleague's performance notes, even on the same tool.
Giving a clear brief
A clear brief states the audience, the purpose, the constraints and the format, and supplies the source material instead of assuming the tool knows your situation. Example: "Rewrite this for a client who has not read the earlier email, under 150 words, keep the delivery date as written, add nothing that is not in my draft" instead of "make this better".
Checking the output
Checking means comparing the output against something outside the tool: the source document, the actual figures, the policy, a person who knows. Reading it and finding it plausible is not checking. Example: if it summarised a contract, the clause numbers and dates get checked against the contract itself, and any figure gets recalculated.
Saving and reusing a useful workflow
A one-off improvement disappears when you close the tab. The behaviour worth having is keeping the instruction that worked, in a note or a saved prompt, so the second time costs minutes rather than the same half hour. Example: the brief that produced a usable weekly report is saved in a note with the corrections you made, then reused.
Measuring the full effort
The honest number includes reading, checking, correcting and any rework somebody else had to do. Felt and measured effort come apart badly here: in a randomised study of 16 experienced developers on their own codebases, they took 19 percent longer with AI tools while believing they had been sped up by 20 percent.3 Survey research also finds the review and rework landing on whoever receives weak AI output, not whoever sent it.4 Example: a draft that took two minutes to generate and eleven to fix did not save you nine minutes.
Explaining the improvement with evidence
The last check is whether you can describe what changed and how you know, in a sentence somebody could challenge. Aggregate savings are small because most people use AI on a thin slice of the week: United States workers self-reported savings equal to about 1.6 percent of all work hours in 2025.5 A checkable claim about one task beats a general claim about your productivity. Example: "The monthly report used to take a day and now takes about half a day, and I keep the checking step because it caught two wrong figures in June."
One check, worked through
Illustrative example, not a customer result
- Task
- A team coordinator writes the same project status update every Friday.
- What she checked first
- Permission and sensitivity. The tool is on her employer's approved list, and because the update names clients she pastes the milestone dates and her own notes, not the client file.
- The output check
- Dates and owner names verified against the project plan. One date was wrong in the draft.
- The full effort
- Six minutes drafting, nine minutes checking and correcting, against about 35 minutes before.
- What she can now say
- "The Friday update takes about 15 minutes instead of 35, and the check has caught a wrong date twice."
- Why this is illustrative
- These figures describe a plausible workflow to show the shape of the checks. They are not a measured Stay Useful customer result, and yours will differ.
What this checklist is not
It is not a score, a certification or a prediction. Seven observable behaviours are a prompt for honest reflection, not a validated instrument. Nothing here tells you whether your job is safe, or gives you permission to use a tool on particular information: only your employer and the provider terms answer that, which what happens to the data you enter into AI tools at work covers in detail.
It is also retrospective by design. For the method rather than the diagnosis, see the practical guide to using AI at work. For the evidence side, the check most people skip, see how to stay valuable at work as AI gets better.