An AI learning session that rehearses the next decision

By the end of an AI learning session, participants should be better able to work through a decision relevant to their jobs. Making that objective observable requires an appropriate task within the session. A collection of impressive demonstrations does not yet reveal how someone will handle an unclear requirement, missing information or an unsuitable answer.

The sequence below can be adapted for an internal session of approximately an hour. The timing is a proposal, rather than a scientifically established optimum. The example is designed for managers and scientific staff; a clinical education team can adapt the structure to its own professionally approved objectives.

Select a single decision

In our fictional example, a group needs to decide whether an internal project proposal is ready for a decision. Participants should identify a missing assumption, explain its significance and ask a targeted follow-up question. This is a narrower and more observable objective than becoming generally more confident with AI.

Prepare two short artificial proposals. Each omits information that matters to the decision. The second changes the context sufficiently that simply reusing wording from the first case will not suffice. The first might contain an unsupported assumption about available staff time, while the second leaves ownership of a necessary step unclear.

The responsible educator defines in advance which reasoning would be sound and which questions allow several defensible responses. A proposal need not be made artificially unambiguous. A justified request for further information may itself be the intended outcome.

Why an independent attempt belongs in the sequence

In 2011, Karpicke and Blunt studied undergraduates learning science texts. In their first experiment, involving 80 participants, active retrieval led to stronger performance on a comprehension assessment one week later than the comparison conditions. The paper investigates a learning mechanism under those conditions, rather than workplace AI education. [1]

My design inference for this session is that participants should generate the relevant reasoning themselves and apply it again later. Experiencing a good explanation is insufficient evidence of that capability. The effectiveness of the particular exercise needs to be assessed through its own results.

Phase one: form an independent judgement

Give the group the first proposal and a few minutes of uninterrupted working time. Each person records three short points: a provisional decision, the most important reason and one unresolved question. Initially, they work without AI. The purpose is to make their starting point visible.

Explain why this step exists. The initial answer supports the later discussion and should not become a concealed individual performance assessment. Where experience levels differ considerably, prepare different levels of difficulty. The shared learning question remains the same, while the amount of material and technical language suit each role.

Invite volunteers to explain their reasoning. The educator collects different approaches without immediately resolving every open question. This leaves something concrete for the subsequent assistance to address.

Phase two: use bounded assistance

Participants now receive access to an AI environment approved for the exercise, or to prepared AI responses clearly labelled as such. A task might be: formulate one question that could help me examine an assumption in my current decision. Participants compare the suggestion with their own reasoning and explain whether it advances the decision.

The boundary is deliberate. If the objective is to recognise a missing assumption, the exercise must provide an opportunity to work on precisely that capability. A complete, extensive solution can be discussed later, but should not retrospectively replace the independent attempt.

Kestin and colleagues reported in 2025 on carefully designed AI tutoring in a Harvard physics course. Their study used prepared subject content and structured interaction. It supports examining the instructional design of assistance, while testing neither the sequence described here nor its effects on managers. [2]

Participants now mark what they changed in their decision and why. An adopted suggestion is useful when its significance can be understood and justified. A reasoned rejection can also be a valuable learning outcome.

Phase three: work through a new task

Distribute the second proposal. Again, participants initially work without AI. They need to apply the distinction they have learnt to a changed case and explain their decision. Anyone requiring additional information should be able to identify it specifically.

The educator examines three things. Was the relevant gap recognised? Was its significance for the decision explained coherently? Does the proposed follow-up question address that gap? This small assessment framework can be discussed with another knowledgeable person using sample answers beforehand. Disagreement between educators should be resolved before the group review, or acknowledged as genuine room for professional judgement.

A good second answer can indicate learning. It does not establish a permanently available capability. That requires a later observation and, depending on the task, several different cases.

Keep the discussion focused on decisions

Ask about the decisive change in the reasoning. Which information was originally missing? What made that visible? Which hint helped, and which was distracting? The discussion should lead participants towards an explanation that someone else can follow.

Difficulties also provide information for the organisation. Perhaps the proposal was ambiguous. Perhaps participants lacked prior knowledge that the educator assumed. Perhaps the AI assistance did not suit the objective. Examine these possibilities before turning a weak result into a judgement about the individual.

Only artificial or explicitly approved materials belong in the exercise. Confidential cases from everyday work cannot be inserted spontaneously as substitutes. The educator should explain this boundary at the beginning and have suitable alternatives available.

Agree the next contact during preparation

Plan a short follow-up task for a later date while preparing the session. It can reveal whether the central distinction remains available and where another explanation is required. The interval should suit the objective and the working context.

Make participation requirements and the use of the results clear. A follow-up that arrives without context may feel like an unrelated test. Connecting it explicitly to the earlier decision helps participants use it to recognise progress, identify uncertainty and request support where needed.

After the first run, the responsible people decide which elements to change. Their basis is the decisions observed, the quality of the reasoning and the remaining learning needs. This turns an AI learning session into a concrete opportunity to develop professional judgement and improve the educational design through evidence of its effects.