AI Role-Play Training for Sales, Support, and Interviews

Some skills only improve with repetitions against another person: handling an objection, calming an angry customer, answering "tell me about a time you failed." Colleagues are busy, and live calls are an expensive place to practice. An AI model can play the other side as many times as you like. How much that helps depends almost entirely on how you set up the scenario, the persona and the scoring.

Design the scenario before the persona

A scenario is the situation and the stakes. Write it down before you think about who the other person is. Include:

• Your goal. What counts as success, stated so you could tell afterward. "Get agreement to a second meeting that includes the budget owner" works. "Have a good call" does not.
• The setting. Phone, chat or video, how much time you have, and what you know going in.
• Your constraints. The policy you must follow, what you can offer, and what you cannot.
• Hidden information. What the other side knows or wants but will only reveal if you ask well. This is what turns a role-play into a test of skill instead of a reading exercise.
• The end condition. When the scene stops, so the conversation does not drift on.

Keep one skill per scenario. A single practice call that tests discovery, objection handling, pricing and closing produces feedback too scattered to act on. Split it into four short scenes instead.

Build personas with motives, not adjectives

"A difficult customer" gets you a cartoon. A useful persona has a role, a goal, a fear, private facts, a way of talking, and a clear rule for what would change their mind. Here is a sales example:

You are playing Dana, operations manager at a regional food distributor.
Situation: I'm a sales rep calling about warehouse scheduling software. You agreed to 20 minutes.
You want: fewer missed delivery windows. You fear: another tool your team won't use.
Hidden facts (reveal only if I ask good questions): a pilot with a competitor failed last year,
and your VP must approve anything over the budget you control.
Behavior: polite but busy, short answers, push back on vague claims.
Rules: stay in character. Don't coach me during the call. Don't agree to a next step
unless I've uncovered the failed pilot and who approves the purchase.
When I type END, step out of character and wait for scoring instructions.

The same structure works for support. A customer persona might have been charged for a renewal they did not expect, want a full refund, and be willing to accept a partial one only if the agent explains the policy clearly and apologizes without blaming them. Write your organization's actual policy into the setup so the persona and the scorer both know what is allowed.

For interviews, borrow from how structured interviews are built. The US Office of Personnel Management describes structured interviews as using rules for eliciting, observing and evaluating responses, and distinguishes two question types. Behavioral questions ask you to describe past behavior in a situation relevant to the skill. Situational questions ask what you would do in a scenario like ones on the job. Ask the interviewer persona to use both types for the role you want, one at a time, with a follow-up probe after each answer.

Score with a rubric, not a vibe

Without a rubric, AI feedback drifts toward generic encouragement. OPM's guidance scores interview answers against benchmarks of proficiency, meaning concrete examples of what a weak, adequate or strong answer looks like. Do the same for any skill: three to five criteria, each with a described low, middle and high level.

A rubric for a support de-escalation scenario might look like this:

Then ask for scoring with evidence attached:

Score the transcript on each rubric criterion from 1 to 5.
For every score, quote the exact line that justifies it.
List what was missing for the next level up.
End with the single change that would most improve the next attempt.

Requiring a quoted line for each score keeps the feedback tied to what you actually said.

Get feedback you can act on

John Hattie and Helen Timperley's 2007 review "The Power of Feedback" argues that effective feedback answers three questions: Where am I going? How am I going? Where to next? A role-play setup maps onto them directly. The scenario goal answers the first, the rubric scores with quotes answer the second, and the one change answers the third.

Run practice as a short loop:

1. Play the scenario.
2. Score it against the rubric.
3. Pick one change.
4. Replay the same scenario with the same persona, trying only that change.
5. Once the change holds, make the persona harder or swap in a new objection.

For spoken skills, practice out loud if your tool supports voice, since typing hides pace, filler words and long pauses. Every so often, have a manager or colleague score one of your transcripts with the same rubric. If their scores and the AI's differ a lot, fix the rubric wording before you trust either.

Know the limits of realism

AI role-play is useful practice, not a faithful simulation. Plan around these limits:

• The persona gives in too easily. Research on sycophancy in language models by Mrinank Sharma and colleagues, published in 2023, found that five leading AI assistants of the time consistently showed sycophancy, giving responses that match a user's beliefs over truthful ones, across four varied tasks. The authors link the tendency partly to human preference judgments used in training. In practice, a skeptical buyer may soften quickly and a scorer may grade generously. Counter it with hard rules in the persona ("do not agree unless...") and by asking for weaknesses before strengths.
• Score in a fresh session. Paste the transcript into a new conversation and ask for an assessment of "a trainee's call." The scorer is then less anchored to the friendly back-and-forth.
• Characters drift. Long conversations wander out of role. Keep scenes short and restate the rules if the persona starts coaching you.
• Facts get invented. A persona may make up product details or policies. Supply the facts that matter in the setup.
• Real people are messier. They are tired, distracted and emotional in ways a model will not reproduce, and text leaves out tone and body language entirely.
• Scores are for practice. An AI rubric score is a training aid, not a validated assessment. Do not use it to rate employees or candidates.

Keep private information out of the scenario

Realistic practice tempts people to paste real material. Build composites instead.

• Leave out real customer names, contact details, account numbers and recordings or transcripts of real calls. Change the identifying details.
• Leave out confidential deal terms, pricing agreements and internal documents unless your organization has approved the tool for that kind of data.
• For interview practice, describe your experience without confidential details about a current or former employer.
• Check the tool's data settings. As of September 2026, OpenAI's Data Controls FAQ says that once you turn off "Improve the model for everyone," new conversations will not be used to train its models. Other tools have their own settings, and work accounts may be governed by separate terms, so ask whoever administers yours.
• Remember that training is not the only exposure. Chat history is stored, and on work accounts it may be visible to administrators.

The safest habit is also the simplest: write each scenario as fiction from the start, with invented names and invented numbers, and you will never have to decide whether a detail was safe to paste.