Why We Don't Put an AI Copilot in Your Negotiation

There is an obvious product feature an AI negotiation platform could build: listen to the live conversation, transcribe it, detect the supplier's move and then whisper the next response to the buyer in real time. The prompts could be as simple as “challenge the cost claim”, “ask for evidence”, “hold the silence” or “offer volume only against price”. Technically, that feature is increasingly feasible. From a capability-building perspective, we think it is the wrong direction.

The reason is not that AI has no place in negotiation. It has a very useful place before the conversation and another useful place after it. The problem begins when the system becomes a hidden participant inside the live negotiation and starts telling the negotiator what to do while the pressure is happening.

A September 2026 review by Jeanne Brett in Current Opinion in Psychology gives that distinction a much stronger research basis. Brett reviews empirical work on AI interventions in negotiation and separates them 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 create dependency.

The proposed mechanism is not laziness. Motivated learners may defer to what looks like AI expertise, or they may become overloaded by dynamic advice while already trying to manage the negotiation. That is remarkably close to the product boundary we would choose even without the paper.

The person still has to own the live judgment

A negotiation is not a sequence of isolated objections with one correct response attached to each. The supplier says that three percent is the minimum. Whether the buyer should challenge, accept, trade or pause depends on information that may never be spoken during that moment: the mandate agreed with finance, the value of upcoming projects, qualification lead times, the strength of the BATNA, a maintenance shutdown next month, a legal constraint, the political importance of the supplier internally, or an earlier concession already exchanged.

A transcript can tell you what was said. It does not automatically tell you what the commercially correct move was.

That matters because a live copilot has to make a recommendation now. It cannot wait for the missing context to arrive later.

Consider a buyer who moves from two percent to three percent after the supplier raises continuity risk. A system watching only the dialogue could label that as an unnecessary concession. But perhaps three percent was already inside the buyer's authorised mandate and the supplier had simultaneously agreed fixed hourly rates for the project. Perhaps another percentage point would save less than the expected cost of delaying mobilization.

The observed behaviour is clear. The quality of the decision is not.

This is the same judgment problem we discuss in AI and procurement judgment. Commercial context is distributed across people, systems and prior decisions. The AI only sees what reaches it.

A useful analysis can therefore say, “You moved after continuity risk was introduced without first asking for evidence.” It should be much more careful about saying, “You should not have moved,” because the second conclusion requires more context.

Live advice competes with the conversation for attention

There is a second problem even when the advice is correct. A tense negotiation already consumes attention. The negotiator has to listen, interpret what matters, remember the mandate, track several variables, notice changes in tone, decide whether a threat is credible, regulate their own reaction, choose a response and deliver it clearly.

Now add another stream of work: read the AI suggestion, interpret it, decide whether the AI has enough context, compare it with your own judgment, accept or reject it, and then return attention to the counterpart, who may already have continued speaking. The system has become another participant to manage.

Brett's review explicitly identifies overload from dynamic AI advice as one possible path to dependency. Separate CHI 2026 research on human-AI negotiation provides a useful supporting mechanism. In a controlled rental-negotiation experiment with 32 participants, human performance declined as the number of simultaneously managed issues increased and cognitive load rose. The study was not testing a live coaching overlay in a human-to-human supplier negotiation, so it is not direct evidence against copilots. It does show how quickly negotiation performance becomes sensitive to additional cognitive demands.

Interestingly, that CHI study also found that a non-prescriptive visualization could reduce cognitive burden in the multi-issue task. That is an important distinction. Support that organizes complexity is not the same thing as a system continuously prescribing the next move.

The question is not whether technology may ever appear during a negotiation. It is whether it helps the person see the situation more clearly or starts substituting for the person's judgment.

Dependency is a learning problem even when the deal improves

Suppose the live copilot works perfectly. The buyer handles the supplier more effectively because the system feeds them the right questions, reminds them to stay silent and warns them when they are about to concede. The immediate commercial outcome might improve, but the learning question becomes harder: if the difficult moment is always the moment when the system intervenes, the learner gets fewer repetitions of exactly the judgment the organisation needs them to own.

This is why AI-assisted performance and capability building are not automatically the same objective. A navigation system can improve today's route without teaching the driver the city. A live negotiation copilot can potentially improve today's responses without making the buyer more capable of producing those responses unaided.

That distinction becomes more important if the organisation expects the person to remain accountable for the decision. It is also why the evidence Brett reviews is so useful. The promising learning mechanisms occur where the learner still has to think: Socratic coaching, practice, feedback and reflection. The risk of dependency appears where explicit AI advice moves closer to executing the task for them.

The research is still developing, and Brett explicitly calls for more work on the mechanisms behind learning versus dependency. The current evidence does not justify saying that every form of live support harms skill. It does justify asking a much harder design question than “Can we build it?”

Before the negotiation, AI has room to challenge the thinking

Before the conversation, time works differently. The buyer can give the system a structured brief, ask it to separate facts from assumptions, attack the negotiation logic, identify missing information and build the strongest credible supplier case. There is room to verify a claim, check with finance, discover that the supposed BATNA will take nine months to qualify and reject bad AI advice without losing the thread of a live conversation.

This is where tools such as ChatGPT can add real value. Our article on practising negotiation with ChatGPT covers that preparation role in detail. The important thing is that preparation should strengthen the person's map of the situation rather than become a script they must obey. The negotiation still belongs to the negotiator.

After the negotiation, AI can work with evidence instead of interrupting performance

After the conversation, the time pressure is gone. Now the system can examine the transcript, the original brief and any context the organisation is permitted to provide. It can identify where the conversation moved, which assumptions survived contact with the counterpart, where the buyer deviated from the intended strategy and which moments deserve another look.

That is a much better place for forensic analysis than the middle of a sentence. There is also a crucial difference between describing behaviour and prescribing a commercial answer. “You offered movement before establishing what the supplier would trade in return” is an observation grounded in the conversation. “You should have rejected the offer” may require knowledge the transcript does not contain.

For simulated Voice2Evolve sessions, this separation already matters: the analysis is grounded in the scenario, the conversation and the known training context. Voice2Evolve does not currently capture or analyse real supplier negotiations. If that boundary is extended in the future, the direction we would consider is post-conversation analysis against the original brief and known context, using the same discipline as sparring analysis. It would not be a hidden AI voice telling the buyer what to say next during the real negotiation. That is an important product boundary, not a missing feature.

The development loop is different from a copilot

The learning loop we favour is simple: prepare the case, train the conversation, let the person own the live decision, then analyse what happened and decide what to train next. The AI is heavily involved around the conversation, but it does not become the negotiator.

That design is less seductive than an interface that flashes the perfect sentence at exactly the right moment. It may also be much better aligned with the capability organisations say they want to build.

If the goal is merely to maximise the next response, live guidance is tempting. If the goal is to create buyers who can handle pressure when no one is whispering in their ear, the boundary matters. AI should make the negotiator better rather than become the negotiator.

Sources

• Brett, J. M. (7 September 2026). AI and negotiation: Facilitating learning or fostering dependency?. Current Opinion in Psychology, 73, 102437.
• Parmar, M., & Silpasuwanchai, C. (13 April 2026). From Overload to Convergence: Supporting Multi-Issue Human-AI Negotiation with Bayesian Visualization. CHI '26.