How I Use an LLM to Prepare for a High-Stakes Negotiation

An LLM can make negotiation preparation more systematic. It cannot tell you what the other person thinks, make an unreliable deal safe, or replace your judgment in the room.

I learned that distinction while preparing for a difficult commercial conversation. I used a model to test my assumptions, rehearse objections, and improve my questions. The counterpart later took a break and returned with a better offer than I expected.

That is the observation. It is not proof that the model caused the outcome, that the counterpart “collapsed,” or that experience stopped mattering. My earlier version of this story made those leaps. This version documents the method that is actually reusable.

What an LLM can—and cannot—contribute

Useful contribution	Unsafe inference
Generate plausible objections for rehearsal	Predict what this counterpart will say
Expose contradictions in your own brief	Determine whether your facts are true
Offer alternative wording and questions	Choose your reservation point or accept a deal
Run repeated, low-cost simulations	Reproduce the incentives and emotions of a real room

The model is a sparring partner. It produces possibilities, not privileged access to another person's motives. Treating fluent output as intelligence about the counterpart is the fastest way to rehearse the wrong negotiation.

Step 0: establish the information boundary

Before prompting, decide what the model is allowed to see. A high-stakes brief may contain personal data, prices, contract terms, legal advice, trade secrets, or information covered by an NDA. “It helps the prompt” is not permission to disclose it.

• Prefer synthetic facts: replace names, companies, dates, and exact values with roles and ranges.
• Minimize: include only the fact required for the exercise. The model rarely needs the complete relationship history.
• Respect policy and authority: use only an approved system with the retention, access, and data-use controls your organization requires.
• Keep restricted material out: if you cannot confirm that disclosure is permitted, do not paste it. Rehearse the structure with fictional data instead.

I now prepare a redacted brief first. The unredacted source stays outside the chat, and any generated advice is treated as untrusted until I verify it.

Step 1: write the decision before the dialogue

Role-play is premature if you have not defined the decision you are trying to make. Write these fields without an LLM:

1. Objective: the outcome you want and why it matters.
2. BATNA: your best realistic alternative if this negotiation ends without agreement.
3. Reservation point: the worst acceptable package, including non-price terms. Below it, you choose the BATNA.
4. Target: an ambitious but defensible package.
5. Authority: what you may decide in the room and what requires approval.

Do not ask a model to invent these numbers. They come from your economics, constraints, evidence, and governance. A model can challenge them: “Which assumption would make this BATNA unavailable?” That is useful. “Choose my minimum price” is abdication.

Step 2: map interests, evidence, and trades

Separate what you know from what you merely suspect. For every counterpart hypothesis, record the evidence and a question that could disconfirm it.

Item	Example, deliberately generic	How to test it
Known constraint	Delivery must complete this quarter	Confirm the deadline and consequence
Hypothesized interest	Predictable support may matter more than a discount	Ask how support risk will be evaluated
Trade	Faster payment in exchange for phased scope	Price both terms before offering either
Evidence	Comparable delivery data	Bring the source, not an invented statistic

This is where a model helps most: ask it to find hidden assumptions, missing stakeholders, asymmetric trades, and ways your proposal might fail. Require it to label every claim as a fact from the brief, an inference, or a question.

Step 3: rehearse adversarially, not theatrically

A useful simulation is not “play a ruthless negotiator.” That prompt rewards drama. Give the model a mandate, constraints, and a scoring rubric:

You are a rehearsal partner, not an adviser.
Use only the redacted facts below.
Do not invent the counterpart's private motives.

Round 1: challenge my proposal with three plausible objections.
After each response, score me on:
- whether I asked before defending;
- whether I separated fact from assumption;
- whether I protected my reservation point;
- whether I proposed a reciprocal trade instead of a concession.

At the end, list unsupported assumptions and the strongest
question I failed to ask. Do not recommend accepting a deal.

Run at least three variants: collaborative, constrained, and skeptical. Change one assumption at a time. If the “winning script” works only against the model's cooperative persona, it is not a strategy.

Step 4: prepare questions, not a script

Scripts break when the conversation changes. I carry a small set of prompts that help me listen:

• “What would make this difficult to approve?”
• “How would this work operationally after the agreement?”
• “It sounds like predictability matters more than speed—is that accurate?”
• “What am I missing about the constraint?”

Labels and mirrors are not magic words. Used mechanically, they sound manipulative. Their purpose is to check understanding and invite correction. Silence is not a weapon either; it gives both sides time to think.

Step 5: keep the model out of the decision loop

During the meeting, I use the preparation—not a covert real-time agent. The human negotiator owns attention, trust, truthfulness, consent, and the final decision. I will not use the model to impersonate someone, fabricate leverage, conceal material facts, or pressure a vulnerable person.

If recording or transcribing is involved, obtain the required consent and follow the applicable policy and law. For employment, legal, medical, financial, or other consequential negotiations, involve the qualified professional your situation requires.

Step 6: debrief against evidence

Immediately afterward, record what happened before asking a model to interpret it:

• Which questions produced new information?
• Which assumptions were confirmed, contradicted, or left unknown?
• What concessions and reciprocal trades were proposed?
• Did the outcome stay above the reservation point?
• What will I test differently next time?

One of my own weekly audits provides a useful negative result. I sent seven outreach emails and created no real back-and-forth negotiation. The messages used a yes-oriented request, with no accusation audit, mirroring, or deliberate silence. Calling that activity “negotiation practice” would have hidden the absence of the behavior I wanted to improve. I changed the template to include an objection forecast and a genuine how-or-what question.

That small audit taught me more than the dramatic success story: a technique is not part of your practice until the log shows that you used it, and an outcome is not attributable to the LLM without a comparison.

Download the preparation canvas

I turned this protocol into a fillable CSV. It includes classification, BATNA, reservation point, counterpart hypotheses, evidence, reciprocal trades, question design, stop conditions, authority, and debrief fields.

Download the LLM negotiation preparation canvas

Keep sensitive values in your approved system—or offline—and use redacted placeholders in model conversations.

Where this work goes next

The next useful question is not whether an LLM can produce convincing negotiation language. It can. The question is whether structured rehearsal improves observable behavior and outcomes across repeated, controlled scenarios. That requires baselines, defined metrics, and failures that are allowed to count.

I am applying the same standard to my agent-to-agent email negotiation protocol: publish the measurement plan before claiming a result. If you are designing a high-stakes AI-assisted workflow and want an independent review of its boundaries, contact me.