Practise Negotiation With ChatGPT

For a long time, the case for practising difficult conversations with voice AI ran ahead of the evidence. The logic made sense. Speaking to an unpredictable counterpart should be closer to a real negotiation than typing into a chat window. What we had very little of was evidence that professionals would actually use this kind of rehearsal at meaningful scale.

A new paper from Amazon researchers changes that part of the picture.

Their system, Conversation Coach, was used by more than 40,000 managers over six months to rehearse difficult workplace conversations. It let managers speak to configurable AI employee personas and then receive personalised feedback. The researchers report that usage was concentrated around conversations people considered difficult rather than becoming a general-purpose chat habit.

That is a significant deployment. It is not evidence that 40,000 managers became better managers. The study does not establish that. There is no reported controlled comparison showing improved workplace outcomes after using the system.

What it does establish is something narrower and still important: professionals will use AI to rehearse consequential conversations aloud, and building a useful rehearsal system requires considerably more than switching on a voice mode.

That matters for negotiation training.

Conversation Coach was designed around three problems: keeping the spoken interaction responsive, giving the simulated employee a configurable personality, and producing useful feedback afterwards.

The researchers also compared two very different technical approaches.

One used a direct speech-to-speech model. The other used a chain of speech recognition, a language model and text-to-speech. The direct system was substantially faster. Its median time to first audio was 1.4 seconds, compared with 4.2 seconds for the cascaded system.

Yet the researchers deployed the slower architecture.

Their reason was reasoning quality. They judged the cascaded system better at the kind of understanding required for coaching.

That is an unusually useful finding for anyone thinking about AI practice.

A conversation simulator can sound fluid and still be a poor training partner. Low latency matters because awkward pauses destroy the feeling of a live exchange. Natural speech matters. Interruptions matter. But none of them compensates for a counterpart that loses its position, misreads what was said, gives inappropriate advice or behaves differently simply because the user pushes hard enough.

In a supplier negotiation, imagine a buyer challenges an 8 percent price increase. The supplier has claimed labour inflation, energy costs and limited capacity. Three turns later, the AI forgets the capacity argument and suddenly accepts that only labour matters. The voice may be excellent. The practice is not.

The commercial logic of the counterpart has already collapsed.

You can ask ChatGPT to role-play a supplier. That is useful.

Give it a situation, tell it to ask for a price increase, describe the supplier's constraints and speak to it rather than typing. You have already moved beyond passive preparation. You now have to formulate the answer aloud and respond to something you did not write yourself.

For an occasional warm-up, that may be enough.

Training becomes more demanding when the purpose changes from “have a conversation with me” to “expose whether this behaviour holds under pressure.”

A supplier with low margins should not accept a 15 percent reduction simply because the buyer argues confidently. A sole-source supplier should behave differently from one facing three credible competitors. A counterpart that has already stated that payment terms are a board-level constraint should not casually trade them away five minutes later.

The role has to survive the conversation.

So does the scenario.

That is why configurable personalities in the Amazon system are more interesting than they may initially sound. The AI is not simply being asked to produce plausible language. Its behaviour is being shaped around a particular type of counterpart.

For negotiation practice, that distinction becomes even more important because the value of the exercise often sits in the resistance.

If the simulated supplier exists primarily to help the buyer, the buyer gets a supportive conversation. If the supplier has a credible commercial position to defend, the buyer gets something closer to practice.

A second 2026 study adds another reason to be careful about equating fluent voice interaction with useful conversational practice.

Jiyoung Lee and Yujong Park analysed 84 screen-recorded sessions involving six Korean learners using the AI speaking application Speak. Their question was not about negotiation. It was about interactional competence in second-language practice. But the mechanism they observed matters well beyond language learning.

When a learner misunderstood the application's previous question, the system often did not perform the kind of third-position repair that a human conversation partner would use to restore shared understanding. When speech recognition captured different words from those actually spoken, the burden of repairing the exchange largely fell on the learner. In one extended case, repair repeatedly failed and shared understanding could not be restored.

The authors argue that this kind of interaction can be closer to monologic production than genuine conversational practice.

That finding should not be stretched into a claim about every voice model or every negotiation simulator. The sample was small, the domain was language learning and the study evaluated one speaking application.

The useful test is narrower: natural turn-taking is not enough if the system cannot notice when shared understanding has broken down.

In negotiation, repair is part of the skill. A supplier interprets a question as an accusation. A buyer misunderstands whether a condition is firm or exploratory. One side refers to annual volume while the other thinks the number is quarterly. Strong negotiators notice the divergence, clarify it and rebuild a common frame before continuing.

A simulator that simply generates the next plausible turn can sound fluent while training the wrong thing. Realistic practice needs a counterpart that can be misunderstood, can notice misunderstanding and can recover from it.

None of this reduces the value of a general-purpose AI assistant in negotiation preparation.

It is often very good at the work that should happen before rehearsal.

Take a sanitised supplier case and ask it to attack your logic. Give it your proposed opening and ask where you are over-explaining. Ask what evidence the supplier is likely to challenge. Ask which parts of your BATNA sound less credible than you think. Ask it to identify the concessions you might give without receiving anything in return.

That work can improve the quality of the preparation considerably.

A buyer facing a supplier price increase might arrive at a practice session with a much better map of the conversation because an AI assistant has already helped them think through the case.

The danger is confusing that improvement in preparation with an improvement in execution.

Knowing that you should refuse an unsupported inflation argument is one thing. Hearing the supplier say, “Then I need to be very clear that we cannot guarantee capacity for Q1,” and answering without immediately softening your position is another.

One happens at the desk.

The other happens in the conversation.

A new review draws a sharper line around live coaching

A September 2026 review by Jeanne Brett in Current Opinion in Psychology makes the distinction between preparation, practice and live assistance much more explicit.

Brett reviews empirical research on AI in negotiation training and groups interventions into practice, coaching and feedback. Socratic coaching and post-negotiation feedback and reflection appear to support learning. Interactive, explicit AI coaching during planning or during the negotiation may instead create dependency.

The proposed mechanism is not simple laziness. Motivated learners may defer to the apparent expertise of the system, or they may become overloaded by dynamic advice while already trying to manage the task.

That is a useful boundary for negotiation tools. AI can challenge the case before the conversation, provide a counterpart to practise against and help analyse what happened afterwards. Those uses still require the person to produce the behaviour.

A system that tells the buyer what to say while the supplier is speaking is doing something different. It may improve the immediate response while reducing the need for the buyer to exercise the judgment the training is supposed to develop.

The review is not proof that every form of live assistance harms learning, and Brett explicitly calls for more research on the mechanisms behind learning and dependency. But it gives a stronger research basis for a distinction that matters in practice: support around the negotiation is not the same thing as guidance through the negotiation.

That distinction is explored directly in why we do not put an AI copilot inside the live negotiation.

The number 40,000 deserves attention precisely because production deployments at this scale are unusual in published research on AI conversation training.

It also deserves restraint.

We do not know from this paper whether the managers subsequently handled real employee conversations better. We do not know how performance compared with human coaching, conventional role-play or no rehearsal. The paper reports adoption and system design, not a causal training-outcome study.

That distinction is important because AI training products can easily make a leap from “people used the system” to “the system developed capability.”

Those are different claims.

Still, adoption answers a practical objection that has followed simulated conversation training for years: will busy professionals actually rehearse conversations if practice is available privately and on demand?

At least in this deployment, tens of thousands did.

That suggests the friction of traditional role-play may be part of the problem. A manager does not necessarily want to schedule a workshop next Thursday to prepare for a difficult conversation tomorrow morning. A buyer does not want to assemble colleagues every time a supplier escalation needs rehearsing.

On-demand simulation changes the economics of repetition.

The Amazon paper shows that voice rehearsal can operate at serious scale. A separate question is whether a particular simulation is good enough to create useful practice.

For negotiation, I would judge that less by how human the voice sounds than by what happens after the buyer starts testing it.

Does the supplier remember its economics?

Does it protect the interests it was given?

Does it notice what the buyer actually said?

Does pressure change its behaviour without destroying the scenario?

Can the buyer repeat the same situation and see whether their own behaviour improves?

Does the feedback identify what happened in the conversation rather than simply praising the user and summarising the topic?

Those questions move the evaluation away from conversational novelty and toward training quality.

They are also why Voice2Evolve treats the AI counterpart, the scenario logic and the post-session analysis as separate parts of the training environment rather than asking a general assistant to improvise everything from one prompt. The aim is not to prevent variation. Variation is essential. The aim is to let the conversation vary without losing the commercial reality underneath it.

Yes, within limits.

ChatGPT can help structure a case, attack your reasoning, generate objections and give you a useful spoken warm-up. Modern voice AI is good enough that the conversation can feel surprisingly natural.

The new evidence makes the larger opportunity clearer. More than 40,000 managers voluntarily used a purpose-built voice system to rehearse difficult conversations. That is strong evidence that spoken AI rehearsal has moved beyond the prototype stage.

It is not yet evidence that any voice chatbot automatically becomes an effective coach.

The more serious question is now about training design: whether the counterpart has a role worth practising against, whether it can hold that role when the conversation becomes difficult, whether feedback is grounded in what happened, and whether the user can try again.

Talking is the entry requirement.

Practice starts after that.

Brett, J. M. (2026). AI and negotiation: Facilitating learning or fostering dependency?. Current Opinion in Psychology, 73, 102437. Published 7 September 2026.

Lee, J., & Park, Y. (2026). Conversation by design: repair and interactional competence in L2 talk with an AI speaking application. Applied Linguistics. Published 15 September 2026.

Wu, F., Maharjan, S., Yessenalina, A., Chen, D. X., Srivastava, R., & Sengamedu, S. H. (2026). Conversation Coach: A Voice-enabled AI System that Helps Practice Difficult Workplace Conversations. EMNLP 2026 Industry Track / arXiv. Primary source on arXiv