AI Role-Play Training: Scenarios and Feedback Examples

AI role-play training gives people a place to rehearse workplace conversations with an AI playing a buyer, colleague or customer. A useful session has a clear task, an agreed standard for feedback and a chance to try again. The aim is to practise an observable behaviour, not to collect a flattering score.

This guide provides sample scenarios and feedback for sales, leadership and support teams. All names, dialogue and scores below are fictional learning-design examples, not client results or transcripts from a live Mimic deployment. Adapt the details to your policies and have the relevant subject experts review them before use.

Build a brief before starting the conversation

Choose one moment that people find difficult. For example, a buyer questions the price, a colleague misses a handover or a customer requests an exception. State what the learner knows, what the other person wants and what the learner is authorised to promise. A scenario with no boundaries can reward an impressive answer that would be inappropriate at work.

Write the brief in five parts: audience, situation, objective, constraints and evidence of success. For a sales exercise, the objective might be to understand the buyer’s concern and agree a useful next step. The constraints might prohibit invented product claims and unapproved discounts. Success means clarifying the concern and checking understanding, not forcing the buyer to agree.

Give the AI persona facts it may reveal when asked, a consistent position and a stopping point. Do not make every persona hostile. Include uncertainty, limited time and reasonable disagreement so learners practise listening as well as persuasion. Keep the first exercise short enough for a learner to review the conversation and repeat it.

Scenario 1: a buyer challenges the price

Brief: You sell a business service. The buyer likes the proposed scope but says the price exceeds their budget. Your goal is to understand the comparison and explore a next step. You cannot offer a discount or promise a return on investment without approval.

Buyer: “Your proposal costs more than the other one. Why should I pay extra?” An unhelpful opening is: “We are the best, so it will pay for itself.” It asserts superiority and a financial result without discovering what the buyer is comparing.

A more useful opening is: “Which parts of the other proposal are you comparing with ours? I’d like to understand whether the difference is price, scope or the support you need.” If the buyer explains that implementation support is missing from the alternative, the learner can compare approved scope details without inventing a saving.

Example feedback: “You asked an open question before defending the price. That helped establish the comparison. You then described support without checking whether it matters to this buyer. On the next attempt, summarise the concern and ask which support requirements are essential before suggesting a revised scope.”

Retry: Change the buyer’s concern from total price to payment timing. The learner should discover the new concern rather than repeat the previous answer. A coach can check whether the learner adapts their questions and stays within commercial authority. Agreeing a follow-up with the right colleague can be a valid outcome.

Scenario 2: a manager discusses missed handovers

Brief: A colleague has submitted two weekly handovers after the agreed deadline. The delay left another team waiting for information. Your goal is to discuss the specific events, understand the cause and agree a workable next step. This is a routine development conversation, not a disciplinary or medical assessment.

Colleague: “You only notice when something is late. I have too much to do.” A weak response is: “You need a better attitude.” It labels the person and leaves the practical problem unexplored.

A more constructive response is: “The last two handovers arrived after the agreed time, so the next team could not start its checks. I want to understand what got in the way. Which competing tasks were you dealing with?” The learner can then discuss priorities and agree who will resolve a conflict in workload.

Example feedback: “You described two events and their impact without judging the colleague’s character. You asked about workload, but moved to a solution before hearing the answer. Try pausing, summarising what you hear and checking that the agreed deadline is achievable.”

Retry: The colleague identifies an unclear dependency rather than excessive workload. Ask the learner to agree an owner, a next step and a review point. Do not score agreement alone: a colleague may reasonably disagree, and a respectful unresolved discussion can be better than a forced commitment.

Scenario 3: a customer asks for an exception

Brief: A customer is frustrated after contacting support twice. They request a refund that the learner cannot authorise. The goal is to acknowledge the problem, clarify the desired resolution and explain the approved escalation route without promising an outcome.

Customer: “I’ve explained this twice. Just refund it now.” A risky answer is: “I promise you will get your money back today.” Even if it calms the simulated customer, it exceeds the learner’s authority and should not receive positive feedback.

A better response is: “I’m sorry you’ve had to repeat this. I can review the information already recorded and explain the next step. I cannot approve that refund myself, but I can send the request to the team that can review it.” Any response-time commitment must come from the organisation’s actual policy.

Example feedback: “You acknowledged the repeated contact and stated your authority clearly. You offered an escalation, but did not confirm what information the receiving team would need. On your next attempt, check the relevant details and explain what the customer can expect next, using the approved process.”

Retry: The customer asks for a guarantee that the exception will be approved. The learner should keep the boundary clear while remaining helpful. Use invented account details in practice; do not paste identifiable customer records into an unapproved AI tool.

Scenario 4: a team member challenges a change

Brief: A team is moving to a new workflow. A colleague worries that the transition will add work. The learner must explain the confirmed change, hear the concern and separate known facts from questions that need an answer. They should not claim that every detail has already been resolved.

Colleague: “This will slow us down. Nobody asked the people doing the work.” A useful opening is: “Which step do you expect will take longer? I can explain what is confirmed, and we can record the part that needs testing.” This invites a concrete concern without promising that the change will be cancelled.

Example feedback: “You invited a specific example instead of dismissing the objection. You then described a benefit as certain, although the brief only listed it as an intended outcome. Distinguish the aim from evidence and agree how the team will check the concern during the trial.”

Retry: Ask the learner to end with an accurate summary of the concern, the person responsible for checking it and a follow-up point. The exercise tests clear communication under uncertainty; it should not reward persuading someone to accept unsupported claims.

Use a feedback rubric with observable evidence

A practical rubric describes what a reviewer can see or hear in the conversation. For these examples, use four criteria: clarifies the situation; acknowledges the other person’s concern; stays within approved facts and authority; and agrees an appropriate next step. Define each criterion for the specific role rather than applying the same wording to every team.

An illustrative scale is 0 for not demonstrated, 1 for partly demonstrated and 2 for clearly demonstrated. Add “not assessable” when the transcript is incomplete or the scenario did not create an opportunity to show the behaviour. A missing recording is not evidence of poor performance. These anchors are a starting point for review, not a validated assessment instrument.

Worked example for the customer-support scenario: clarification 1, acknowledgement 2, authority boundaries 2 and next step 1. The explanation matters more than the total: the learner recognised frustration and avoided a refund promise, but did not establish the details needed for handover. The coaching priority is to clarify the request and explain the escalation, not simply to increase a number.

Use the same feedback structure after each attempt: identify a specific moment, describe its effect, suggest one change and offer a retry. “Be more empathetic” is too vague on its own. “You moved straight to policy after the customer described repeated contact; acknowledge that inconvenience before explaining the limit” gives the learner something concrete to practise.

Review the AI feedback before relying on it

Have human reviewers score a small set of sample conversations independently, then compare their reasoning with the AI output. Investigate disagreements and revise unclear criteria. Test different valid ways of answering, including concise responses and alternative phrasing. A learner should not need to repeat a preferred script to demonstrate understanding.

Check transcription errors, language requirements and accessibility before interpreting scores. Do not infer personality, honesty or emotional state from a voice or an automated score. Give learners a way to flag an incorrect transcript or challenge feedback. Keep managers responsible for contextual judgement and avoid using an unvalidated practice score as the sole basis for an employment decision.

Tell learners what is recorded, who can see it and how long it is kept. Use fictional or appropriately de-identified practice information. Review privacy, permissions and retention with the relevant owners before rollout. The NIST AI Risk Management Framework linked below provides broader risk-management guidance; it does not certify these sample rubrics or prove training effectiveness.

Run a focused pilot and check transfer to work

Start with one role and one conversation, then agree the learning objective and acceptance criteria before testing. Ask representative learners whether the persona behaves consistently, the dialogue makes sense and the feedback is actionable. Record failures such as invented policy, contradictory instructions or an escalation that never triggers, and resolve them before expanding the pilot.

Use a baseline attempt, a coaching discussion and a later attempt with a different version of the situation. Keep conditions comparable. Improvement on the same prompt may reflect familiarity with that prompt, so a new scenario and a human observation of relevant work provide more useful evidence than repeated scores alone.

Separate participation from learning and business outcomes. Attempts and completion show usage. Evidence against the rubric shows performance in the exercise. Workplace observations may indicate transfer. Sales, satisfaction or escalation figures have other influences, so do not attribute a change to role-play training without an appropriate evaluation design.

Agree how managers will use the findings: one behaviour to practise, a follow-up conversation and a review date. Define integration needs only after the learning workflow is clear. A text or voice pilot may be sufficient; a digital human or immersive environment should serve a specific training need rather than determine the project by default.

Plan your first AI role-play exercise

Bring one recurring conversation, the relevant policies and a few examples of what good handling looks like. Mimic Business can use that brief to scope the persona, practice flow, feedback criteria and human review requirements. Start with a focused prototype and decide how it will be evaluated before committing to a broader programme.

Use the service link alongside this guide to discuss a pilot. For the wider sales-enablement programme, continue to the related sales-training guide rather than treating these sample dialogues as a complete rollout plan.

For a focused manager-development brief, scope a manager-feedback pilot with a fictional workplace situation, conversation variants and observable coaching criteria.