How to Prepare a Negotiation With AI: Scenarios, Arguments, Limits

How to Prepare a Negotiation With AI: Scenarios, Arguments, Limits

To prepare a negotiation with AI, use the model as a rehearsal partner, not as the negotiator. Ask it to help you map interests, build scenarios, simulate objections and stress-test your limits, always with anonymized information. What you accept, concede and refuse stays your decision, written down before the meeting.

Key takeaways

• AI is useful for widening your preparation: interests, alternatives, likely objections and the order of concessions.
• Your limits (what you will not accept, and what you trade for what) must be set by you before you open the conversation with the model.
• Do not paste the other side's confidential information, or documents covered by an NDA, into an AI tool. Work with generic descriptions and aliases.
• Always separate documented fact, your own experience and hypothesis. AI blends the three unless you ask it not to.
• Finish with a one-page sheet: objective, alternative, limits, arguments and questions. The sheet goes to the meeting. The chat does not.

What AI does well and what it should not do in preparation

A language model is good at expanding and organizing. It lists interests you had not considered, suggests how the other side may read your proposal, reorders arguments by priority and writes different versions of the same sentence, from firmer to more conciliatory. It is also a good sparring partner: you ask it to object, and you practice your answer.

It is weak in three areas. First, it does not know the real counterpart: anything it says about them is plausible guesswork, not information. Second, it tends to agree with you and to make your plan look more coherent than it is. Third, it does not carry the cost of the decision. That is why what not to delegate to AI in your company includes the final judgment on concessions, relationships and risk.

The practical rule: let AI generate options and questions, and keep the choice and the commitment for yourself.

What data should go in the prompt, and what should stay out?

Preparing with AI usually fails on confidentiality first, not on the quality of the answers. Before you type anything, decide what is allowed to leave your company.

Generally fine to include:

• Your objective, described in words, without exact figures when they are sensitive.
• Your interests and internal constraints that you are authorized to process with the tool.
• The type of counterpart and the context: a large supplier, a recurring customer, a new partner.
• Public information about the other party, with the source noted.

Keep out:

• Documents, proposals or emails from the counterpart that are marked confidential or covered by a non-disclosure agreement.
• Personal data of people at the other company beyond what is strictly necessary. LGPD (Brazil's data protection law) still applies to the name, role and history of an identifiable person.
• Margins, costs and internal numbers that would weaken your position if they leaked.
• Passwords, full contracts and anything your internal policy prohibits.

Use aliases and qualitative ranges. Instead of the supplier's name and the contract value, write Supplier A and a high-value contract for our size. The model handles this well and you lose very little. If your company has not yet defined what can go into external tools, start with a company policy for employee ChatGPT use. If the subject is too sensitive for any cloud service, consider running a local model, and if the provider processes data outside the country, see what LGPD requires for international data transfer.

There is also a technical precaution. If you paste text from third parties (an email, a web page) into the prompt, that text may contain instructions the model tries to follow. OWASP covers this risk, called prompt injection, in its list of risks for applications built on language models. Treat external text as material to analyze, never as instructions.

How to map interests and alternatives with AI, step by step

Preparation starts with what you want, not with what you will ask for. The sequence below works for a contract renewal, a significant purchase, a partnership or a price adjustment.

• Write your objective in one sentence and the minimum acceptable result in another. Do this alone, before the chat.
• Describe the situation to the AI using aliases and ask it to list the likely interests of each side, separating interest (why they want it) from position (what they ask for).
• Ask for a list of negotiable variables beyond price: timing, volume, payment terms, warranty, exclusivity, scope, service level.
• Define your best alternative if there is no deal, meaning what you do if you walk away. Ask the AI to challenge that alternative: is it real, does it have a timeline, what is the cost of switching?
• Ask for a qualitative estimate of the other side's best alternative, labeled as a hypothesis. You will test that hypothesis with questions in the meeting.
• Tag each item as documented fact, your own experience or hypothesis. Only facts support strong arguments.

If part of your preparation involves understanding the other party's market or competitors, use public sources and record where each piece of information came from, as we describe in competitor analysis with AI.

How to simulate objections and scenarios without fooling yourself

Simulation is the stage where AI adds the most. Instead of asking for a generic answer, give it a role and constraints: ask it to act as an experienced counterpart from a company with the profile you described, to present the five most likely objections to your proposal, from the weakest to the hardest, and not to give in easily. Then answer them yourself, aloud or in writing, and ask the AI to assess your answers and point out where they were vague, defensive or contradictory.

Work with at least three scenarios:

• Favorable scenario: the other side accepts the essentials. What else do you ask for?
• Central scenario: there are trade-offs. Which concessions do you offer, and in what order?
• Difficult scenario: the other side hardens its position or imposes a deadline. What is your limit, and how do you communicate it calmly?

Also ask the model to attack your own arguments: where do they depend on an unverified assumption? This is the most valuable use, because it works around the model's tendency to agree with you. If an answer comes back looking too good, ask for the pessimistic version.

Watch out for a common mistake: treating the simulated objection as a prediction. It shows possibilities, not what the real person will say. Prepare to hear something different and write down open questions to uncover real interests.

How to set your limits before the meeting

A limit is a decision made calmly, before the pressure starts. AI helps you test it, but not choose it. Use this decision guide and fill in each item in writing:

• Objective: what a good deal looks like, in terms of both outcome and relationship.
• Minimum acceptable: below this, you prefer not to close.
• Alternative: what you do if there is no deal, and how quickly.
• Concessions that are cheap for you and valuable to them: what to offer first.
• Expensive concessions: what only comes in exchange for something equivalent.
• Non-negotiables: legal, ethical, safety or reputation items.
• Who decides: if the decision goes through a partner, finance or legal, put it on the sheet so it is not improvised at the table.
• Trigger to pause: situations in which you ask for a break instead of answering on the spot.

Once it is written, ask the AI to look for inconsistencies. Does any limit contradict the objective? Is any cheap concession actually expensive because of a side effect, such as setting a precedent for other contracts? You make the corrections, not the model.

If the negotiation involves a contract, clauses or legal obligations, use AI to generate questions for your legal team, not to replace them.

Illustrative scenario: renewing with an important supplier

Illustrative scenario: a mid-sized distributor is about to renegotiate its annual contract with a packaging supplier. The purchasing manager writes down, on her own, that she wants predictable delivery times and is willing to pay somewhat more for that, but will not accept exclusivity.

In the chat, she uses the names Supplier A and Distributor, with no figures and no excerpts from the proposal she received. She asks for the supplier's likely interests, and the AI suggests, among others, predictable volume, on-time payment and lower logistics cost. She tags all of it as hypothesis.

Next, she asks the AI to simulate objections to her proposal of guaranteed delivery times in exchange for a minimum volume. The AI points out that the supplier may ask for exclusivity in return. She already had that item as non-negotiable, so she writes a ready answer and rehearses it: offer volume priority, not exclusivity.

Finally, she builds the one-page sheet with the objective, the alternative (getting quotes from another supplier, which takes time), her limits and three open questions to find out what really matters to the supplier. The final decision at the table is hers, and the sheet, without the chat history, is what she brings.

How this method applies to working with AI in your company

The method we use in AI projects applies here too, on a small scale. Map: describe the current process, meaning how you prepare negotiations today and where time or information gets lost. Limit: define which data goes in, which actions the AI never takes and who reviews. Pilot: test the routine on one real negotiation of moderate risk, with an agreed success criterion, for example whether the preparation was more complete and whether fewer surprises came up at the table. Transfer: if it worked, turn the routine into a usage template for the team, with clear rules, and decide whether to expand, adjust or stop.

If your team records meetings to review them later, sort out notice and legal basis first: recording meetings with AI and LGPD covers what to look at.

Final checklist before you walk into the meeting

• The one-page sheet is ready and can be read quickly.
• Each argument is tagged as documented fact, experience or hypothesis.
• No confidential data from the counterpart was pasted into a tool without authorization.
• Your limits were set by you and reviewed with whoever decides alongside you.
• You have at least three open questions to uncover real interests.
• You know what you will do if the conversation does not move forward.

For a structure to manage the risks of AI systems, the NIST AI Risk Management Framework is a useful reference. For the use of personal data, Brazil's data protection authority (ANPD) publishes official guidance and materials.

Frequently asked questions

Can I use ChatGPT or another assistant to prepare a negotiation?

Yes, as long as you respect your company's data policy and any confidentiality agreements. Use anonymized descriptions, no documents from the counterpart and no sensitive internal numbers. If the topic is very delicate, consider a model running locally.

Can AI predict what the other side will say?

No. It generates plausible possibilities based on general patterns, and it does not know the real person or company. Use the simulations to prepare for variations and to formulate questions, never as a prediction.

Is it safe to paste the email or proposal I received from the other company?

It depends on the content and on what was agreed. If there is a confidentiality clause or personal data, the prudent route is to summarize in generic, anonymized terms. Also, third-party text can contain instructions the model tries to follow, so treat it as material to analyze.

Should AI decide my negotiation limits?

No. It helps test coherence and surface consequences, but your limits depend on risk, relationships and alternatives that only you and your company know. Set them in writing before you open the chat and use AI to question them.

Sources

• NIST: AI Risk Management Framework
• ANPD: Autoridade Nacional de Proteção de Dados
• OWASP: Prevenção de prompt injection