For an employee learning to review AI output, prepare a checked worked example before deciding how much help to remove. Then leave a specific decision unfinished for the learner to complete. Keep the approved source material available throughout. Removing an answer and removing the evidence needed to reach it are different training choices.
This guide proposes a way for learning teams to design that transition. The practical deliverable is a small sequence of exercises with an explicit record of what the learner sees, what they must decide and when support returns. It is a proposal to trial locally, not a validated workplace AI curriculum.
Our earlier guide to learning after AI assistance explains why a completed output and retained capability need separate checks. Here the question is instructional: what should happen before that independent check, especially when an employee does not yet know how to begin?
What worked-example research can establish
A worked example shows a problem and the steps leading to its solution. Barbieri and colleagues’ 2023 mathematics meta-analysis synthesized 55 studies from 43 articles, spanning elementary through postsecondary settings. It found a moderate average benefit for mathematics performance compared with conditions without worked examples. This does not estimate the benefit of workplace AI training.
Its cautions matter for design. Evidence about prior knowledge was mixed and could only be synthesized descriptively. Conditions with self-explanation prompts had smaller effects overall, but that moderator comparison was not causal: studies differed in other respects. The paper does not establish that asking someone to explain always harms learning, or that advanced learners should always skip examples.
Renkl and colleagues’ 2002 study examined progressively removing worked solution steps. The authors’ university research record reports benefits for near transfer: applying learning to similar problems. Only the abstract and bibliographic record were accessible for this article; we do not treat it as proof of broad workplace transfer or prescribe its experimental sequence as a universal rule.
Rebecca Adler’s January 2024 practitioner explanation makes the mathematics findings accessible to teachers and highlights the mixed prior-knowledge results. It discusses the same meta-analysis, so it is an interpretation aid, not an independent replication.
Define the step that must become the learner’s work
Avoid a vague instruction such as “give less help each round.” Name the reasoning that the employee must eventually supply. Reviewing an AI-drafted recommendation may require locating the applicable rule, comparing a claim against a fact and deciding whether the draft can proceed.
Those are separable actions. An employee might find the right document but miss the contradiction. Another might recognize the contradiction once the relevant passage is highlighted but struggle to locate that passage. Giving both people the same extra slide deck hides the difference.
Daiki Nakamura’s October 2026 conceptual review proposes distinguishing cognitive work peripheral to a learning goal from work that constitutes the goal itself. It offers testable propositions about AI-supported STEM learning, rather than new experimental results. Our application is to state exactly which operation each exercise leaves with the employee.
For example, if the objective is checking a recommendation against a rule, requiring the learner to format a polished memo adds work without directly exercising that objective. A simple response field may suffice. If the objective is finding the governing source, highlighting the decisive paragraph supplies part of the answer and should be recorded as support.
A hypothetical software-request exercise
Consider a fictional operations team learning to review an assistant’s draft recommendation about a software purchase. The invented training policy says that tools receiving customer records require a data review before purchase. Tools used only with fabricated demonstration data do not require that review under this training rule. Other purchasing requirements remain outside this exercise.
The first case states that a proposed add-on will receive customer records. The assistant’s fictional draft recommends proceeding because the subscription is inexpensive. A checked worked example shows three steps: identify the intended data, find the applicable rule and return the recommendation for the required review. Price does not answer the data-review question.
The facilitator prepares the correct decision and its source basis before showing the case. Calling an AI-generated solution “worked” does not make it correct. For a real organization, the designated policy owner would need to confirm the rule and example; no such review is claimed for this invented scenario.
Now prepare the following sequence. Every stage retains the same training policy and the case facts.
| Exercise | Guidance supplied | Work left to the learner |
|---|---|---|
| Checked example | Completed data classification, relevant passage and disposition | Inspect how the conclusion follows from the facts |
| Completion case | Data classification and relevant passage; disposition blank | Decide what happens next and identify the draft’s unsupported reason |
| Less guided case | Case facts and policy; no highlighted passage or completed classification | Find the rule, apply it and give the disposition |
This sequence is our design proposal. Moving forward depends on the response, not on completing a fixed number of screens. If someone cannot make the final decision in the completion case, restore the explanation and try another comparable case before removing more support.
Change the decision, not just the names
For the next case, state that the add-on uses fabricated demonstration data only. Under the invented rule, the customer-data review prerequisite does not apply. The learner should recognize that difference while leaving other purchasing requirements unresolved.
Add a third case in which the request does not say what data will be supplied. The expected response is to obtain that information before settling whether the prerequisite applies. Neither automatically approving nor automatically asserting that customer data is involved follows from the supplied facts.
These cases prevent the exercise from rewarding one memorized response, such as always returning the request. They also let the facilitator distinguish finding a rule from identifying when its conditions are unknown. Do not introduce a new policy exception and remove guidance in the same step unless handling that extra complexity is the intended learning target.
Keep answer feedback specific. “Correct” teaches little about the boundary. Feedback can state which supplied fact activates the rule, or which missing fact prevents a decision. If the training policy cannot resolve the case, revise the material or refer the question to its owner. Do not let the assistant invent a missing rule to complete the exercise.
Record support without treating it as failure
A useful practice record can be brief: case version, decision made, support supplied and the next exercise chosen. Distinguish a passage the learner found independently from one the facilitator highlighted. Both may lead to a correct result, but they answer different questions about what guidance is still useful.
Do not assign a permanent “novice” or “expert” label from job title, confidence or one response. A senior colleague can be unfamiliar with a new purchasing policy; a recent hire can already know how to verify a source. Let people request an example and allow them to attempt a less guided case when appropriate.
Also separate solution guidance from normal working resources. Removing the completed classification is consistent with practising classification. Removing the policy tests recall as well. Keep accessibility support and approved reference tools available unless changing those conditions is necessary for the stated objective.
An explanation prompt should earn its place. Asking the learner to identify the fact that changes the decision gives a focused response to inspect. Requiring a long explanation after every small step may consume time without resolving the learning team’s uncertainty. Pilot the prompt and look at what its answers actually reveal.
Keep the example stable while testing the lesson
An approved assistant can help draft alternate cases, but check each one’s facts, rule and expected response before use. Keep the training version identifiable. A live-generated explanation that changes the policy halfway through a session makes the exercise difficult to interpret.
If an assistant delivers practice, state whether it may reveal a complete solution, offer a hint or only give feedback after an attempt. Check that behavior in the actual training setup. An instruction requesting restraint does not establish that a product enforces it. A static checked exercise is a reasonable starting point when that behavior has not been verified.
Begin with one target decision, one checked example and two changed cases. Review where learners needed support, whether the cases were answerable and whether less guidance left the intended reasoning with them. Use those observations to revise the next lesson. Completion speed alone cannot tell a training owner which step is ready to become independent.