Practising with an AI role-play: going through the difficult conversation safely

The problem with conversation training

A bad-news conversation, an angry customer, a colleague who feels unheard: these are the situations where training is needed most and hardest to organise. Role-play with colleagues is awkward, costs the time of two people and often stays at one go. An actor is expensive. Many people therefore hardly practise, and then do it for the first time for real.

How a conversation with a CrewMate works

In a conversation slide the author chooses a CrewMate and describes the situation. Examples are a bad-news conversation, a customer conversation and motivational interviewing. The learner talks, by typing or by voice. The character responds, with a 3D avatar with a talking head.

After the conversation it is assessed and a final report is written. For those who want hands-free: every CrewMate chat has a headset button, the system hears when you are done, answers aloud and listens again. You can also interrupt.

For motivational interviewing there are eight worked-out practice scenarios, for example around workload and a new system. That is content to start with and adapt, not a fixed catalogue.

Why practice with feedback works

Anders Ericsson’s research on deliberate practice shows what good practice looks like: a concrete, bounded task, immediate feedback and repetition with small adjustments. An AI role-play delivers those three almost by itself. The task is a situation, the feedback comes right after and you can do it again straight away with a different approach.

Then there is psychological safety. A character does not judge and does not tell anyone. That lowers the threshold to experiment, even with what you would not dare to do for real.

What it does not replace

Staying honest matters. An AI character is not a person. It does not see your face, does not feel your hesitation and does not know the real relationship. For the nuance of tone, timing and trust, a human coach stays valuable. See also the role of the trainer.

So use AI practice for the first five rounds: the basic structure, the words, the hesitation that has to be overcome. Use the human for refining and for situations where a lot is at stake.

Designing a good practice scenario

1. Choose one behaviour. For example "asking follow-up questions before you solve". Not "communication".
2. Describe the situation concretely. Who is the character, what do they want, what do they know, what is the tension?
3. Decide what good looks like. What is the conversation assessed on? Make criteria visible to the learner.
4. Allow multiple rounds. A second attempt with a different approach is where the learning is.
5. Discuss the report. Let learners share a piece of their report and discuss it in the group.

By way of illustration: a conversation about workload

A manager notices that an employee, let us call her Sanne, is more and more often late and has less energy. In the practice scenario the CrewMate plays Sanne. She is defensive, says it is fine and wants to keep the conversation short. The learner has to ask follow-up questions without judging, listen to what Sanne is really saying and find a first step together.

Afterwards the system assesses the conversation and writes a report: how often did you ask an open question, where did you jump in with a solution before Sanne had finished, when did you manage to stay quiet? The learner reads it, picks one point and plays the conversation again. That is the core: a short loop of doing, looking back and doing it again.

Safety and agreements

A practice conversation feels personal, even with an AI. So make agreements. Tell learners who sees the report: only the learner, or the trainer too? Use fictional situations and people. Make it voluntary where something is sensitive. And say that it is practice: a bad conversation in practice is a gain, not a failure.

For use with voice the same applies: speech is recognised in the browser and the answer is read aloud. Add that the learner may calmly interrupt and start again. See also feedback that works.

Which conversations are suitable?

Not every conversation suits an AI role-play. It works well for conversations with a recognisable structure and a clear goal: an intake, a bad-news conversation, an appraisal conversation, a sales conversation, a conversation about workload. It works less well for conversations that depend mostly on the relationship and the history, such as a conflict between two colleagues who have known each other for years.

Start with a situation many people dread and that comes up often. Keep the scenarios short, start simple and build up to harder variants. And let learners choose which scenario to do first.

Outside the lesson: a CrewMate that thinks along

A CrewMate is more than a counterpart. A personal CrewMate or a CrewMate of the learning journey answers questions and carries out actions, in the dashboard chat or on the phone. So the same technology that plays a character in the lesson also becomes a help in the work.

Also read feedback that works on what a good report contains.