If you are going on leave and a colleague has to take over a report you prepare with AI, forwarding your last chat leaves them to work out which parts of the conversation matter. They need to know where the original records live, which instructions are current, and what to do when an answer looks plausible but lacks evidence. A good handover lets someone run the task, check the result and handle a problem without waiting for you to come back.
Write down what the next owner actually delivers
Describe the job in terms of its inputs, finished result and audience. For a fictional facilities team, that might read: Every Thursday, turn the approved maintenance tracker into a short internal update. Include open jobs that affect building access, name the team responsible for each job, and flag any completion date the tracker leaves blank. The facilities supervisor approves the update before it goes to staff. The tracker is the official record.
Now attach one example of a finished update and point to the underlying rows that support it. Beside the example, explain one deliberate omission. For instance, a tentative contractor visit stayed out because no one had confirmed it. A clean sample shows how the update should look. The omission shows where the owner exercised judgment, which is usually harder to convey in a list of steps.
Show the next owner where to work
Put the current prompt or instructions in a shared location alongside any approved template. Say which AI tool and workspace the team uses. Identify the input file by its location and owner so a colleague can tell the official tracker from an old export. Ask them to confirm that they can open each resource with their own account before the handover. For customer or employee data, use the minimum fields needed and follow the organisation’s approved handling process. Use a fictional or suitably sanitised record for practice.
Document decisions the tool can help with and actions reserved for staff. In the facilities example, AI can draft wording and flag blank fields. A supervisor settles conflicting completion dates, and staff send the approved notice. If the team already maintains a decision log for AI-assisted work, point to its location so the next owner knows where to record a changed rule or approval.
Run one normal case and one awkward case together
Have the incoming owner drive while you observe. Give them an ordinary example and ask them to produce the update from scratch. Notice where they have to ask for a missing link or guess how to name a file. Add those details to the handover. Then give them an awkward case, such as two entries for the same job with conflicting completion dates. The expected action might be to mark the date unconfirmed, contact the tracker owner and hold that item out of the final notice until it is resolved.
Use the exercise to verify what leaves the team as well as what the tool returns. Ask the colleague to trace each date in the draft to its row, open the current approval channel, and show where they will leave the final copy. Compare the finished update with the sample only after they have checked the records. Two notices can look nearly identical while one repeats a date from an outdated tracker.
Leave a short card for the moments things go wrong
At the top of the handover, record the next run date, task owner, backup contact, approved work location and final approver. Include a one-sentence stop rule. For example, if the tracker is unavailable or the draft contains a date that cannot be verified, do not circulate the update. Add the link to the manual fallback for times when the tool is unavailable. Write down how an unresolved item is carried over so the same problem does not disappear between weeks.
The NIST Generative Artificial Intelligence Profile discusses human oversight and how organisations manage generative AI risks across a system’s lifecycle. The handover card and facilities drill here are practical examples for a small team. Set a date to check the document after the first independent run, while the new owner can still remember the steps they had to improvise.
Use one departure test
Before you sign off, ask the colleague to show a prepared update and identify one item they held back, the record they checked and the person who can resolve it. That demonstration gives you evidence that they can make the workflow’s routine judgment calls, instead of just following the steps on a clean example. If they get stuck, fix the instruction or access problem and repeat that portion of the exercise.
