The Wrong Question Every Negotiator Is Still Asking About AI

Dear Negotiation Explorer,

Welcome to issue 38 of our NegoAI series.

I keep getting the same question. From procurement leads, from sales VPs, from senior partners running cross-border deals. Sometimes in workshops, sometimes by email after a course, sometimes at the start of a coaching call before we've even covered what they're working on.

The question is reasonable. There are several capable models, the landscape changes every quarter, and people don't want to invest in the wrong one. But it's still the wrong question.

Back in issue 28, I quoted something that's stayed with me — a line that's also become the headline for the live Lightning Lesson I'm running on May 11:

In issue 28, I was making the broader point about the window closing on AI literacy in negotiation. Today I want to take it one level deeper, because most readers I talk to have heard that line and still — still — believe their main lever is which model they're on.

It isn't.

Here's what happens. A senior negotiator hears about AI in negotiation. They open ChatGPT. They type something like: "I'm renegotiating a $15M consulting contract. Help me prepare."

The output is fine. Structured. Confident. Generic.

So they try Claude. Better tone, maybe. Still generic.

So they try Gemini. About the same.

They draw a reasonable conclusion: "AI isn't ready for negotiation yet."

That conclusion is wrong — but not because they tested the wrong models. It's wrong because they were never testing the model. They were testing what an empty context stack produces, across three different LLMs. The output was generic because the input was generic. The variable was never the AI.

This is the same point I made in issue 29 ("It's Not About the Prompt"), but with one extension. Issue 29 said: the prompt is maybe 20% of what determines output quality. The other 80% is everything you feed around the prompt. That's true. And it's also true that the model itself isn't the 80% either. Context still is. Switching from ChatGPT to Claude with the same empty stack changes the surface — the tone, the structure — but not the substance. You're still getting 80% of nothing.

If you've concluded "AI isn't there yet" after trying a couple of chatbots cold, you didn't diagnose the technology. You diagnosed an empty context stack.

Two early-2026 papers put empirical numbers on this. Meta-Harness (Stanford–MIT) reports a 6× performance gap from changing only the wrapper around a fixed LLM — same model, same prompts. Princeton's Holistic Agent Leaderboard, after 21,730 agent rollouts across 9 models and 9 benchmarks (coding, math, web nav, customer service), finds that once the scaffold is fixed, no single model dominates. Negotiation hasn't been tested at this scale yet — which is what the CEF paper (issue 33) and the May Lightning Lessons address.

Issue 29 introduced four context layers: Knowledge, Memory, Instructions, the Prompt itself. Issue 30 went deep on memory. Issue 37 named two of those layers as the two drawers that every serious AI copilot tool exposes — Knowledge Base (what the AI knows) and System Instructions (how it behaves).

The full picture, with everything we've covered across this series, is five layers:

Strip any layer and the output gets weaker in a specific way. Strip several and you're back to generic. The compounding lives across the layers, not inside any one of them.

Memory deserves a special mention here, because it's the layer that turns the others into a system. KB and SI together make the AI capable on day one. But without memory, the AI starts from zero every Monday morning. It re-meets your counterpart every time. It re-discovers the partner's anchoring pattern every quarter. KB and SI are foundations; memory is what makes them compound across deals.

That's what we did in issues 30, 35, and 36 — the memory templates, the five context files, the structured MEMORY.md file that turns a flat list into navigable institutional intelligence.

So when I get the question — "which AI is best?" — what I really hear is: "which of my five layers should I focus on?" And the answer is never "the model."

Let me show you what this looks like concretely.

You've met Sarah Chen across this series. CPO at a major pharma. Renegotiating an $18.5M consulting contract with a top-tier firm. Decade-long relationship, three years of measurable underdelivery, AI tools internally now matching about 40% of the firm's routine strategic work.

I ran the same prompt through the same model three times. The prompt was a one-page preparation request — Sarah's interests, the issues, her BATNA, scenarios for the counterparty, ZOPA, value creation, objective arguments, concession map. The kind of structured prep we covered in issue 32.

The only thing I changed between runs was the context stack behind the prompt.

Below, I'm showing you just one section of each output: the Counterparty Scenarios. The output had everything else, but scenarios are where the difference is most visible.

No Knowledge Base. No System Instructions. No memory. No documents. Just the prompt, into a blank chat.

Read that and ask yourself: which deal is this for?

Right. It could be any of them. A consulting renegotiation. A vendor contract. A licensing deal. A salary discussion. The labels are interchangeable. The drivers are textbook. There's nothing in here that someone with five minutes and a negotiation textbook couldn't have written.

This is the output that makes capable negotiators conclude AI isn't ready.

Now I gave the model a Knowledge Base — negotiation expertise organized around the scholars I rely on, the same compendium Deepak draws from in my work. And I gave it System Instructions — be a prep partner, challenge weak assumptions, give options not conclusions, structure responses cleanly. The drawers from issue 37, both filled.

Same prompt. Same model. New stack.

This is unrecognizable from Run 1. The scenarios are sharply named. The interests are precise. The BATNAs are real. The drivers connect to specific commercial logic — precedent risk, reputational upside, pricing power across a portfolio. This is what negotiation-shaped output looks like.

You could walk into a meeting with this and have a serious conversation.

But notice what's still missing. There's nothing here about this firm. Nothing about this partner. No anchoring pattern. No override authority. No behavioral tells. Run 2 is what a smart negotiation generalist would write before they'd met the people across the table.

Now I added two more layers. Counterpart memory built across past quarterly reviews — the partner's anchoring patterns, his override authority over procurement (counterpart memory shows he overrides above 10% movement, but ceiling is unclear past ~15–20%), Kahneman's behavioral profile (Cool Blue dominant, Fiery Red secondary; the "stewardship" and "holistic impact" language flagged as polished professional persona, not authentic preference). And documents — the transcript from the last quarterly review, where the partner himself said "junior consultants are increasingly capable of strategic work."

Same prompt. Same model. Same KB and SI. Memory and documents added.

Now Sarah is preparing for this meeting. Not a meeting. This meeting. With a partner whose linguistic signals she'll recognize in the first thirty seconds. With a probability shift she can defend (Scenario 1 likelihood went from 45% to 55% based on past anchoring patterns). With a brand-new third scenario — Authority Above the Table — that didn't exist in Run 2 because Run 2 didn't know about the override threshold.

Read the three side by side again. Same prompt. Same model. Same prompt structure, same questions asked. The variable across runs was never the AI. It was the stack.

Run 1 to Run 2 takes you from generic to negotiation-shaped. The drawers — KB and SI — do that work.

Run 2 to Run 3 takes you from negotiation-shaped to seeing this counterpart. Memory and documents do that work, and this is where the compounding lives. After ten deals with the same counterpart, your Run 3 is unreachable by anyone else — not because they have a worse model, but because they don't have your stack.

If you came into this newsletter from a Lightning Lesson invitation, or because cohort 2 enrollment is open, here's the through-line.

The Lightning Lesson on May 11 is called Build Your AI Negotiation System — No Code, Any Tool. Read the title carefully. Any tool. Not "Claude." Not "ChatGPT." Not "Copilot." The session is built around three live demonstrations and the third one is the one most people don't expect: the same Knowledge Base and System Instructions, the same architecture, dropped from Claude into Microsoft Copilot, producing the same negotiation IQ.

Why does that matter? Because if your company runs Microsoft, you're going to use Copilot — that's the tool you have. If your company gives you ChatGPT Enterprise, you're going to use that. And if you've built your stack right, none of that matters. Your method travels. Your KB travels. Your SI travels. Your memory travels.

The model is the runtime. Your stack is the asset.

That's the practical version of what this issue is arguing in concept.

If you take one thing away from this issue, let it be this:

Stop optimizing for the model. Start investing in the stack.

The reader question I keep getting — "which AI is best?" — has the right energy and the wrong target. The energy is correct: people want to invest in the right thing. The target is wrong: it's pointed at a layer that's basically interchangeable across the major providers. (They're all fine. They're all improving. Pick whichever your company gives you and move on.)

What deserves your investment is the four layers behind the prompt:

These four layers are what produce Run 3 instead of Run 1. None of them depend on which model you happen to be using.

Open whichever AI tool your company gives you. Open a blank chat. Run a one-page preparation prompt for a negotiation you have coming up — interests, issues, BATNA, three counterparty scenarios, ZOPA, value creation, objective arguments, concession map. Use whatever prompt structure you normally use.

Read the output. Look specifically at the counterparty scenarios.

Now ask yourself: could this output be for any deal in my industry, or only for this one?

If the scenarios feel interchangeable — Cooperative / Aggressive / Strategic, Fear / Pride / Control — that's not a signal about your AI. It's a signal about your stack. The stack is what we'll spend cohort 2 building, and what you'll see assembled live on May 11.

Monday May 11 — Build Your AI Negotiation System: No Code, Any Tool. Sixty minutes, free. We'll fill both drawers — Knowledge Base and System Instructions — live on Sarah's PharmaCrest case, and then drop the same system from Claude into Microsoft Copilot to show that the method travels. If the argument in this issue is about the stack mattering more than the model, LL1 is where you watch the proof. Save your seat: https://maven.com/p/d82a4e/build-your-ai-negotiation-copilot-no-code-any-tool

Monday May 18 — Build Your AI Negotiation Workflow: No Code. One week later, we extend. A single copilot is powerful. But a real deal needs a process — preparation, rehearsal, debrief — and sometimes several agents working together. LL2 takes the copilot from LL1 and wires it into a three-step workflow. Save your seat: https://maven.com/p/c70efd/build-your-ai-negotiation-workflow-no-code

Cohort 2 — Build Your AI Negotiation System, starts May 25. Five weeks, live, hands-on. The full stack built with you over five weeks — KB, SI, memory templates, the workflow, the multi-agent architecture behind it. Lightning Lessons show you the system in 60 minutes; cohort 2 is where you build yours. Details: https://maven.com/nego-ai/build-your-ai-negotiation-system

Same model, same prompt, three different stacks, three different deals. The next time you hear someone say "AI isn't ready for negotiation yet," you'll know what they actually tested.

Questions? Reply directly — I read every response.