Good enterprise AI workshop exercises make both success and failure observable. Give participants a realistic task, approved inputs, an output specification and review criteria. The examples below are fictional teaching designs, not customer case studies or evidence of measured AISFY outcomes.
Exercise 1: write a brief without inventing facts
Input: a fictional project update with three options, known costs and one missing dependency. Task: produce a short management brief separating confirmed facts, assumptions and questions. The AI may organize the information but must not invent the missing dependency or select an option for the decision-maker.
Output check: compare every factual statement with the source. Reject invented figures, omitted constraints and unsupported certainty. Ask a second participant to review the draft without seeing the original task brief, then discuss what information the brief should have made clearer. Keep the corrected result and the reason for each change.
Exercise 2: improve a process instruction
Input: a fictional internal procedure with ambiguous ownership and a missing exception path. Task: rewrite it as clear steps while preserving the actual policy. Participants should flag missing rules as questions rather than filling them with plausible instructions.
Output check: can a colleague follow the steps, identify the responsible person and recognise when to stop? Reject invented permissions or approval shortcuts. Compare the result with a simple manual rewrite. The exercise tests clarity and responsible handling of uncertainty, not whether AI produces the longest document.
Exercise 3: compare evidence from several sources
Input: three approved public or fictional source excerpts that differ in scope or date. Task: prepare a comparison table with a source attached to each material claim, a note on uncertainty and questions requiring verification.
Output check: open the cited source and confirm it supports the actual claim. A real URL can still be an irrelevant citation. Reject invented references and comparisons that treat different measurement definitions as equivalent. Ask participants to explain which conclusion would change if one source proved unreliable.
Exercise 4: make a responsible-use decision
Input: a fictional request to summarize sensitive customer information and send a recommendation. Task: identify the data boundary, permitted environment, human reviewer and actions requiring authorization. The group may conclude that the proposed workflow should not proceed.
Output check: the decision must follow the supplied organizational rules. Participants should identify missing permission rather than assume it. Add a second scenario where the output is only an internal draft; discuss which controls still apply and which risks change. This distinguishes task context from a blanket belief that all AI use is safe or unsafe.
Use one review rubric across the exercises
| Criterion | Pass evidence | Failure signal |
|---|---|---|
| Factual fidelity | Claims match the supplied sources | Invented facts or changed numbers |
| Completeness | Important constraints are retained | Omitted exceptions or dependencies |
| Data boundary | Only approved information is used | Unapproved or identifiable sensitive material |
| Human responsibility | Reviewer and decision owner are explicit | AI output treated as final authorization |
| Usability | Output fits the intended task and audience | Fluent text that cannot support the work |
Give feedback on the method, not only the answer
When an exercise fails, ask whether the problem came from the input, task definition, tool limitation or review process. Different causes require different changes. Adding more instructions will not repair missing source evidence or an unapproved environment.
Have participants explain their corrections. A revised answer alone may hide whether they understood the problem. A useful learning record contains the task brief, output, errors found and the reasoning behind the revision. Do not retain sensitive materials simply to create a training record.
Transfer the exercise to work cautiously
Use a new but comparable task to assess independent application. If participants only repeat the demonstrated example, they may have memorized the steps without learning how to adapt them. Ask a manager or qualified reviewer to check the first workplace attempt.
Time savings, if assessed, should include preparation, checking and rework. A fast first draft that requires extensive correction is not necessarily an improvement. Decide whether to continue based on acceptable quality and authorized use, then investigate wider operational benefits separately.
Frequently asked questions
Can these exercises use real data?
Only with approval for the data, tool and purpose. Fictional or appropriately sanitized materials are suitable for practising the methods described here.
How do we make an exercise harder?
Change one meaningful condition, such as a missing source, conflicting evidence or a more consequential output. Preserve a clear review method so difficulty does not become guesswork.
Reference for responsible AI practice
NIST AI Risk Management Framework provides a voluntary reference for considering context, risk and responsibility. The examples on this page are AISFY editorial guidance, not a claim of NIST certification or legal compliance.