An AI course draft is the beginning of the work

Start with the performance problem

A tidy outline, a friendly introduction and ten quiz questions can make a draft feel finished. What they cannot tell you is whether the course solves the right problem. Before reviewing the writing, write down what a learner should be able to do afterward and the situation in which they need to do it.

Consider a fictional support team that makes mistakes when deciding which refund requests to escalate. If the escalation rule is buried in an internal system, better access may matter more than a course. If employees can find the rule but misapply it in ambiguous cases, realistic practice is a stronger candidate. AI cannot settle that diagnosis from a request to “make training about refunds.”

Give the model a bounded assignment

Prepare an approved source pack, one observable objective, an audience description and the delivery constraints. Ask for alternatives with their assumptions exposed. A draft that marks an unsupported detail is easier to review than one that confidently fills every gap.

Try this prompt: “Using only the supplied policy, propose three practice activities for new agents deciding whether to escalate a refund. For each, explain the decision being practiced, the feedback needed and the policy section that supports it. Mark missing information as NEEDS REVIEW.”

Only use source material you are permitted to share with the selected tool. De-identify examples and follow your organization’s data rules. An assistant that is useful for a public writing exercise may not be approved for confidential case histories.

Review the decisions, not just the prose

Check whether learners actually practice the intended skill. A multiple-choice question asking for the definition of escalation does not necessarily prepare someone to recognize an escalation case. Ask whether each wrong answer reflects a plausible mistake and whether feedback explains the reasoning.

• Accuracy: can a reviewer trace each consequential statement to the approved source?
• Alignment: does the activity require the target behavior?
• Feedback: does it explain the consequence and the better decision?
• Access: can learners use the experience with their devices and access needs?

Test the smallest useful version

Build one decision and its feedback before producing the entire course. Ask representative learners to attempt it. Watch what they do, where they pause and how they explain their reasoning. Separate confusing interface behavior from difficulty with the actual decision.

Revise the prototype and record why you changed it. If an expert says the policy is correct but learners consistently interpret the situation differently, investigate that mismatch. Both kinds of feedback are useful; neither substitutes for the other.

Count review time as part of AI’s cost

Track the effort to create, check, correct and test the material. A draft produced in minutes is not a time saving if it creates hours of avoidable rework. Keep a short log of recurring errors so the next prompt or source pack addresses them.

Your final deliverable is an approved, usable learning experience. The generated draft is one input into that work.