The short answer
Start with one recurring task, not with the tools. Ask your manager in writing which tool and which information are approved, run AI on that single task for a couple of weeks, check the output against something real each time, and report what actually changed. That is a complete and defensible answer to "use more AI", and it is far more useful to your employer than scattered use across everything.
Being told to use AI without being shown how is the normal situation, not a sign that you missed a briefing. In Gallup's May 2024 survey of over 21,000 United States employees, 70 percent said their organisation had no guidance or policy for using AI at work.1 Most people asked to do this are working it out for themselves.
The rest of this page is the message to send, the experiment to run, and what to report.
First, ask what is approved
Do this before you choose a tool, and do it in writing. It takes one message, it protects you later, and it often prompts the organisation to write down the rule it should have had. You are not asking permission to be competent. You are asking which tool and which information are in bounds.
If there is no policy yet
"We do not have a rule yet" is still a useful answer, and now it is on the record. Until there is a rule, hold the line New Zealand's Privacy Commissioner sets: do not put 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 automatically make information safe to upload, because the surrounding detail often identifies people anyway. Whether you can use ChatGPT at work, and what to check first goes through where to look for a policy and what to do when there is not one.
Pick one recurring task
One task, repeated, is what produces evidence. Ten scattered attempts produce an impression. Choose something that comes up at least weekly, is heavy on reading or drafting rather than judgement, and is cheap to get wrong on the first attempt because you will be checking it anyway.
A good first experiment
- It happens weekly or more often, so you get repetitions.
- It is mostly reading, summarising, drafting or reformatting.
- You already know what a good version looks like.
- Nobody is harmed if the first draft is wrong, because you check it before it goes anywhere.
- It does not require pasting personal or confidential information into an unapproved tool.
Avoid the task that would most impress your manager if it worked. Performance drops when AI is applied just outside what it can do: in a field experiment with 758 consultants, those given the tool for tasks inside its abilities completed 12.2 percent more work and 25.1 percent faster, while on a problem designed to sit just outside them they scored 23 percent worse than colleagues working without it.3 Save the hard, judgement-heavy work until the easy case is running well. The practical guide to using AI at work sets out how to brief the tool properly once you have chosen the task.
Decide what "good" means before you start
Write down, in one line, what a correct output looks like and what you will compare it against. This is the step that separates a trial from a story, and it has to exist before the first attempt, because afterwards everything the tool produces looks reasonable.
- What you check against. The source document, the actual figures, the policy, or a person who knows. Not your impression of whether it reads well.
- What counts as a failure. An invented fact, a wrong number, a changed date, a tone you would not send, or a draft that needed more fixing than writing.
- Who is accountable. You are. Nothing about being asked to use AI moves responsibility for the output onto the tool or onto the person who asked.
Keep the corrections you make. They are the most valuable thing the experiment produces: they tell you what to put in the instruction next time, and they are the evidence that you checked at all.
Report what happened, including the checking time
Two weeks in, you can answer the question you were actually asked. Report the full effort, not the drafting time, because the gap between the two is where these trials usually go wrong. In a randomised study of 16 experienced developers working on their own codebases, they took 19 percent longer with AI tools while believing they had been sped up by 20 percent.4 If you only report the part that felt fast, you will end up defending a claim that does not hold.
Illustrative example of a two-week report, not a customer result
- The task
- Weekly supplier update emails, five to seven a week.
- What was approved
- The employer's licensed assistant. No supplier contract files, only my own notes and the delivery dates.
- What good meant
- Dates and order numbers match the system, tone matches what I would send, nothing added that I did not supply.
- What happened
- Drafting fell from about 12 minutes to about 4. Checking and correcting added about 5. Two drafts invented a delivery date and were caught by the check.
- What I would say
- "It saves roughly a third of the time on this task once checking is included, and the check is not optional: it caught two wrong dates in two weeks."
- Why this is illustrative
- These numbers show the shape of an honest report. They are not a measured Stay Useful result, and yours will differ.
That report is a better answer to "use more AI" than a list of tools you tried. It shows the improvement, its cost, and the judgement you applied, which is the part no tool supplies.