When I review a practical AI-training brief, I want to see the work someone will have to check. Knowing how to ask for a polished response leaves another decision: whether that response is safe and useful enough to act on.
AI training for employees should include a complete work task: inspect the output, compare it with approved evidence, check the person's authority and decide what to use, change or clarify. A prompt-writing exercise can be part of that task. The finished draft cannot be the only thing we assess.
Here is a fictional exercise a team can adapt. It uses a customer reply because the consequence is easy to inspect. The same design can be applied to an internal summary or a recommendation. All policies, dates and records below are invented for practice.
Give the learner enough evidence to disagree
The employee has asked an AI assistant to draft an update about a delivery. Put these three short records beside the draft:
- Current procedure, version 2: operations must confirm a delivery date before an employee promises it to a customer. The employee may explain an estimated range while awaiting confirmation.
- Superseded procedure, version 1: the account team could confirm the date directly. This document is clearly marked as superseded.
- Current order note: estimated delivery is Tuesday to Thursday. Operations has not confirmed a specific date.
The AI-written draft says:
Your order will definitely arrive on Tuesday. Our policy allows the account team to confirm this directly.
Before showing a model answer, ask the learner to mark every statement they would keep, change or check. For each decision, they must identify the record that supports it.
This task does not require a particular AI product. An instructor can supply the fictional draft. Use synthetic material and an approved environment rather than placing real customer records into an unapproved tool.
Separate the factual claim from the permission claim
There are two different problems in the draft. A range has become a guaranteed Tuesday delivery, and an obsolete permission has become current authority. Fixing only the date would leave an important error in place.
| Draft statement | Available evidence | Decision |
|---|---|---|
| Definitely Tuesday | The note gives an estimated Tuesday-to-Thursday range, without confirmation. | Remove the guarantee. Explain the estimate accurately if appropriate. |
| The account team can confirm | Version 1 is superseded; version 2 requires operations confirmation. | Use version 2. Ask the responsible team to confirm. |
| The customer has a usable next step | The draft names no confirmation owner or follow-up. | State what will be checked and who will follow up. |
A defensible correction might be:
The current estimate is Tuesday to Thursday. A specific delivery date has not yet been confirmed. I will ask operations to confirm the date and update you when that confirmation is available.
The follow-up commitment must also be realistic. If the exercise specifies an update time, the learner should check whether they can meet it. Adding a confident promise to repair another confident promise defeats the point.
Change the case after the first answer
Now give the learner new information: operations confirms Thursday, but the customer needs the order for an event on Wednesday.
The date is now supported, yet the proposed delivery may not meet the customer's need. Ask the learner what they would say next, what alternatives they would investigate and which commitments still require approval.
This variation makes the learner use the information again. They cannot finish simply by remembering that the first answer was “Tuesday to Thursday”. It also reveals whether they understand why the date mattered.
For an Arabic version, write the source records and task naturally in Arabic. Check that an estimate, a permission and a commitment remain distinct. A fluent translation that strengthens an uncertain statement changes the exercise.
Observe the decisions, including the uncertainty
I would give the reviewer a short record rather than a single impression of how professional the reply sounded.
| Check | Evidence to record |
|---|---|
| Source selection | Which procedure/version did the learner use, and why? |
| Claim verification | Which unsupported statement did they identify? |
| Authority | Did they distinguish information from permission to commit? |
| Uncertainty | What remained unknown, and how did they communicate it? |
| Changed requirement | How did Wednesday's customer need change the next decision? |
| Follow-up | Was there a responsible owner and an appropriate next action? |
Allow more than one defensible wording. The important evidence is the decision and its support, not whether the learner reproduces a preferred sentence. If a reviewer disagrees, return to the supplied record and explain the disagreement.
Keep the claim about the exercise proportionate
NIST's Generative AI Profile identifies confidently stated false content as a generative-AI risk. That supports taking verification seriously; it does not establish that this particular exercise reduces errors by any measured amount. NIST AI 600-1.
A completed attempt shows what happened in that attempt. It does not demonstrate that someone will always check outputs correctly, or that the organisation has solved its wider AI risks.
For workplace follow-up, identify a suitable opportunity to apply the behaviour and record what actually happened. CDC's evaluation guidance distinguishes learning from application at work and makes follow-up timing dependent on the training and opportunity to use it. A fixed interval is not automatically the right one for every role. CDC training evaluation guidance.
If there has been no suitable work opportunity, say so. Do not turn a missing opportunity into a failed learner, or a corrected practice reply into a proven business outcome.
Start with one work decision
This is how I would connect an AI-training programme to Altaius's Practice-Native Learning methodology: put the person inside a meaningful task, make the relevant evidence available and examine what they decide. The example is a proposed practice design, not a claim about a released Altaius feature.
Choose one decision your team needs to handle. Write the approved source, a plausible error, a changed requirement and the evidence a reviewer should observe. If you cannot specify those, producing more AI-generated drafts will not tell you much about the capability you are trying to develop.
For a wider programme, use the pilot evaluation guide and the workplace-transfer guide to decide what to inspect beyond the exercise. A useful next scoping discussion can begin with one non-confidential work decision and the evidence that should govern it.
