The meeting notes say, ‘Check the supplier’s revised date; Mei will follow up.’ An AI-generated action list says, ‘Mei to confirm delivery by Friday.’ It looks tidy, but the deadline was never agreed. An employee who can spot that invented commitment has learned something more valuable than a clever prompt phrase.
This is a fictional training example. It shows the kind of judgement worth practising alongside drafting and summarising. For a Singapore SME, the aim is a working method colleagues can repeat, including what to do when the answer sounds plausible but is not supported.
Start with a small source that everyone can inspect
Choose a short, synthetic set of meeting notes or an approved product description. Include one deliberate gap, such as an unconfirmed date. Give participants the source and a clear output format. Avoid using real employee or customer records simply to make an exercise feel authentic.
Ask for a draft, then review it against the source. A useful first exercise is short enough that participants can inspect every important claim. If the source is too long or unfamiliar, the group may end up judging how professional the answer looks rather than whether it is correct.
Teach four checks together
- Facts: which statements can you point to in the source? Check names, dates, quantities and commitments. Unsupported information should be removed or marked for confirmation.
- Gaps: what information is needed to finish the task? An empty owner or deadline is sometimes the correct result, provided it is clearly flagged.
- Fit: does the draft answer the actual request, for the intended audience and format? A correct summary may still be the wrong deliverable.
- Authority: who can approve or act on it? Producing a draft does not grant permission to send it, change a record or make a commitment.
Use a task template that makes uncertainty visible
Using only the notes below, draft an action list with Task, Owner, Due date and Source detail. Write ‘To confirm’ where the notes do not specify an owner or date. Do not infer a commitment. List unresolved questions separately. This is a draft for a person to check.
That instruction is a starting point, not a guarantee. Participants still need to compare the output with the notes. Ask them to label each proposed action ‘use’, ‘revise’ or ‘reject’ and explain their decision. Compare two drafts so they can see that the same task may produce different results.
Assess the method, not just confidence
At the end of the exercise, give participants a fresh example. Can they identify the unsupported detail, preserve the uncertainty and explain the next action without a trainer pointing it out? Capture the mistakes that remain. Those observations are more useful for planning follow-up practice than a general feeling that the session went well.
Leave with the task template and its review checklist together. Saving a prompt without the checks encourages people to treat it as a reliable procedure when it is only part of one. Name where the shared version lives and who can update it.
Connect training to the next working day
Choose one approved task for participants to practise again and agree a way to bring back questions. EverX’s Practical AI Team Training uses task-based exercises to develop both tool use and review habits. The tasks, tool access, starting level and expected working materials are scoped with your team.
Further reading
NIST’s Generative AI Profile discusses risks including confabulation and the evaluation of generated content. The exercise above is an EverX teaching example, not a claim of NIST certification or compliance.