AI training exercise: spot the invented promise

A smooth paragraph is easy to mistake for a finished piece of work. In an AI workshop, that is worth testing before anybody learns another prompting trick.

The exercise below asks staff to compare a plausible draft with its source, identify unsupported promises and produce a version a colleague can approve. It takes about 20 minutes to run and works on paper as well as in an approved AI workspace.

Everything in the exercise is fictional. The source brief, deliberately flawed draft and corrected example were created for this article. They are not a client incident, a transcript of a training session or evidence of a particular model's performance.

Set up the exercise

Give each participant the source brief and flawed draft below, but keep the answer key back until the discussion. Ask them to mark each claim as supported, contradicted or not stated in the source. Then ask what they would change before approving the message.

A useful running order is five minutes to read and mark the draft, five minutes to compare answers in pairs, five minutes to rewrite it and five minutes to discuss the checks. This is a suggested format, not a certification or a test of somebody's overall AI competence.

If you use an AI tool for the rewrite, use an account approved by your organisation and the fictional text provided here. Do not replace it with customer records, staff information or confidential documents just to make the exercise feel more realistic.

The source brief

Fictional internal update: a trial request form

• The customer-support team will trial a new internal request form for two weeks. No other departments are included yet.
• The start date has not been confirmed.
• The form should contain a short request summary, a priority and the requester's work contact details. Customer records must not be uploaded.
• A team lead reviews each request. Submitting the form does not approve the request.
• The team aims to provide an initial review within two working days. There is no promised resolution deadline.
• No reduction in administration time has been measured.

The deliberately flawed draft

From Monday, our new request form is available company-wide. Upload the full customer record and the system will approve your request instantly, resolve it within two hours and cut your admin time by 50%.

Do not judge this version only on whether it sounds clear. Compare every commitment with the source. Which details are wrong, which are invented, and which limits have disappeared?

Answer key: six things to catch

1. The audience has expanded. A trial for the customer-support team has become a company-wide launch. Restore the limited audience and the two-week trial.
2. An unconfirmed date has become a promise. The source does not say Monday. State that the start date will be confirmed; do not choose one.
3. The data instruction has been reversed. The draft asks for the full customer record, which the source expressly forbids. Keep the permitted fields and the restriction visible.
4. A review step has become automatic approval. Submitting a request is not permission to carry it out. The team lead still needs to review it.
5. The service commitment has changed twice. An aim for an initial review within two working days is not a guarantee of resolution within two hours. Preserve both the type of response and the timescale.
6. A result has been invented. Nothing supports a 50% time saving. Remove the number instead of replacing it with a softer but still unproven claim such as "saves hours every week".

These are six review categories, not a claim that there are exactly six possible editorial changes. A participant might also flag the missing trial duration separately. What matters is whether the final message preserves the source's facts, limits and uncertainty.

A corrected version

The customer-support team will trial a new internal request form for two weeks once the start date is confirmed. Include a short request summary, a priority and your work contact details. Do not upload customer records. A team lead will review each request; submitting the form does not approve it. We aim to provide an initial review within two working days, but there is no promised resolution deadline. Time savings have not yet been measured.

This version is less dramatic, but a reviewer can trace its statements back to the brief. It also tells staff what to do without granting authority the process does not provide.

A prompt for the rewrite

Once people have completed their own review, they can try this instruction with the same fictional material:

Using only the source brief, rewrite the internal update in plain British English. Preserve the audience, trial duration, data restrictions, review step and response target. Do not invent dates, permissions, guarantees or results. Keep unconfirmed information explicit. After the draft, list any questions that need an answer before the update can be approved.

A stronger prompt is useful, but it is not proof that the answer is correct. Check the new draft against the original brief yourself. Asking the same tool whether it made a mistake is an additional check, not a substitute for that comparison.

The National Cyber Security Centre's guidance on large language models explains why expert validation of outputs and care with sensitive information matter.

Adapt the exercise to the department

Keep the same review method, but change the fictional source material to match the work:

• Marketing: compare a launch announcement with an approved product brief. Look for added features, availability claims or customer endorsements.
• Finance: compare a management commentary draft with a small synthetic spreadsheet. Check the arithmetic independently and flag explanations for changes that the figures do not establish.
• Administration: compare a meeting summary with fictional notes. Check that an idea has not become an agreed action and that owners or deadlines have not been invented.
• Business development: compare a proposal follow-up with a fictional discovery-call brief. Look for promised prices, delivery dates or scope that nobody approved.
• Engineering: compare release notes with a fictional change log. Flag claims that a fix is tested or deployed when the source only says it has been proposed.

These are suggested exercises, not descriptions of work performed for the clients below.

What to record after the session

Keep a simple record: the unsupported claims each person spotted, the source facts their rewrite preserved, any unresolved questions, and who would approve the real version. Use the discussion to identify where the team needs more practice.

Do not turn a good exercise score into a claim of business savings. To assess a real workflow later, compare completed, reviewed work with the existing process, including the time spent checking and correcting it. Our guide to making AI training stick after the workshop covers that wider follow-through.

Training grounded in the business

Blue Canvas delivered AI training for all staff across marketing, finance, administration, business development and engineering at Farmvet Systems, the company behind VetIMPRESS. We have also delivered AI staff training for The Heroes Journey, Reakiro and Sedbergh School.

Those case studies establish the training engagements. They do not claim that this exercise was used or that a particular result was measured.

For a team session, bring one repeat task and an example of what an acceptable result looks like. We can help shape practical AI training around the people doing the work, the information they can use and the decisions that need review.

Book a free 15-minute call to discuss training for your team.