There is a Wednesday call in your diary. The client has gone quiet on a £180k engagement that is six weeks behind. You have been telling yourself you will think it through on the drive in. You will not. You will arrive thirty seconds before the call connects, the kettle still going, and you will open with whatever your face does when their face appears on the screen. Then you will spend the next forty minutes recovering ground you should have held from the first sentence.
The conversation you are avoiding tomorrow is the conversation that goes worst when you wing it. Many owner-operators already know that, in the part of themselves that is honest before bedtime. What they don’t know is what to do about it that doesn’t feel ridiculous.
Twenty minutes of role-play with AI playing the client is what to do about it. Not theatrics, not a sales-training programme, not therapy. Calm preparation. By the end of the session you have surfaced three things: the version of the conversation you are scared of, the version that actually happens, and the opening line that holds.
What does AI role-play actually look like?
It is a focused twenty minutes with three movements. Two minutes of setup, where you brief the model on the client. Twelve minutes of role-play across two or three cycles, where the AI plays the client and you play yourself. Six minutes of debrief, where the AI steps out of role and reflects on what worked and where your reasoning was thinnest.
The brief is the load-bearing piece. A weak prompt produces a generic client. A careful one produces a specific, awkward, recognisable person. You are giving the model the sector, the history, the last four interactions, what has gone wrong, and what you suspect is really driving their silence.
The shape that works in practice: “You are the operations director at [client firm]. You signed a £180k engagement six months ago. Three weeks in, your COO told you to apply more budget pressure. Two weeks ago, the project missed a milestone we attributed to your team’s data delays. You haven’t replied to my last two emails. Play the call I’m about to start. Push back hard where I am vague. Ask the questions you would actually ask. Stay in role until I say ‘pause’.”
Why does this beat rehearsing it in your head?
Because your own brain is the wrong place to rehearse a conversation you are dreading. When you run it solo, you invent a softer client, a more articulate version of yourself, and a graceful resolution. None of those will be in the room on Wednesday morning. The AI client is indifferent to your self-esteem and will give you the version that actually shows up.
The psychology is well documented. Ziva Kunda’s 1990 paper on motivated reasoning showed that when we rehearse alone, we bias the evidence we gather toward outcomes that protect our self-image. The conversation we imagine is the conversation we wish we were having, not the one we are about to have. A friend would notice and stay polite about it. The AI is briefed to push back, and a model in role does not soften the way a friend who likes you will.
That is the part that surprises owner-operators the first time they try it. The AI client will respond to your first weak explanation with the question your real client is going to ask, and it will do it without flinching. The debrief afterwards is where the second piece of value lands: the AI can step out of role and tell you, specifically, where the line you wrote in advance fell apart.
When is the technique worth using, and when isn’t it?
It is worth twenty minutes for any conversation where the cost of getting the opening line wrong is material: a fee renegotiation, a delivery slip you have to own, a stakeholder relationship that has frayed, a renewal where you suspect the client is shopping. For a typical owner-operator running a £1M to £10M services firm, that is roughly one to two calls a month, not every weekly check-in.
The skill, like any rehearsal practice, compounds across sessions. The frameworks Chris Voss teaches in Never Split the Difference (mirroring, labelling, the calibrated question) start arriving without conscious effort once you have used them in role-play four or five times. The first session feels effortful and slightly silly. By the third one the moves have started to land in your real conversations without you having to think about them, which is the point.
It is not worth using for routine status calls, for internal team conversations, or for moments where the issue is purely informational and there is no relationship risk. It is also not the right tool for the conversation where you do not yet know what you think; for those, the sparring-partner approach is closer. Role-play is for the conversation where you already know the position you need to hold, and you need to find the line that holds it under pressure.
What do the platforms add that a good prompt doesn’t?
Specialist platforms exist. Hyperbound and Second Nature are the two most widely cited as of 2026, the former trained on over two million hours of recorded B2B sales conversations. They give you scoring rubrics, rep-level analytics, voice avatars, and a library of scenario templates. They earn their cost in sales organisations running consistent, scored practice across many reps.
For an owner-operator preparing for one call tomorrow night, none of that is the binding constraint. A carefully briefed prompt in Claude or ChatGPT covers the ground for a fraction of the cost, sometimes with voice mode if you want to practise the actual sound of the conversation rather than the typed version of it. The variable that matters most is the quality of your brief, not the polish of the platform.
If you find yourself running this technique weekly, or you have a client-facing team of three or more who would all benefit from the same library of scenarios, the platforms become worth a look. Below that volume, a saved prompt template and a quiet half-hour will get you the bulk of the result.
What does the debrief usually surface?
Three things, almost every time. The first is that the version of the conversation you were scared of is sharper than the version that actually happens, because the AI client puts the worst objection in your face in the first thirty seconds. The second is the line that does not hold under pressure. The third is the line that does, the one you take into the real conversation.
The first and the third are connected. Once you have answered the worst objection once, even badly, the dread softens. You stop spending mental energy on the version of the call that ends with the client walking, because you have already lived through a version of that scene and noticed the floor did not give way. The conversation you walk into the next morning is no longer the catastrophic one. It is just the difficult one.
The technique sits inside the wider category of using AI on your own desk before you ever ask it to do anything for the team or the firm. It is the same lineage as the pre-mortem with AI, applied to a different surface. The pre-mortem rehearses what could go wrong in a project before it starts. The role-play rehearses what could go wrong in a conversation before it happens. Both turn private dread into specific, named, addressable risk, and the bulk of the work of difficult client management lives in that translation.
When the call connects on Wednesday morning, the kettle still going, your face still adjusting, you will not be reaching for the right line. You will already have it.
