Human + AI Usually Loses. Here's When It Wins.

Dear Negotiation Explorer,

For a year now, I've prepared every negotiation with the same AI. The common assumption is that this kind of partnership gets better on its own — more time together, sharper work. For most people, it doesn't. They plateau, or quietly give up.

That tracks with what the research shows. When researchers pooled more than a hundred studies — people working alone, AI working alone, and the two together — the combination usually came out behind whichever of the two was stronger on its own.* The pattern has held: it was replicated again this year. Synergy, it turns out, is rare. It has to be built.

And that's exactly where the research has moved — from whether to combine human and AI, to how to structure the combination so it pays off. That second question is the one I've spent the year living, and it's what my new paper answers. It's my fourth on AI and negotiation, out this week and free on SSRN — I'll link it at the end.

So why does it go right for a few people and not the rest? Why do some describe their work with AI getting sharper month after month — closing better deals, not just faster ones — while most plateau?

It comes down to one thing almost no one builds.

A while back, in issue 36, I showed you how to structure the memory file an AI reads at the start of every session — grouping entries into rooms, marking what's been superseded, checking for contradictions as you write. That was the how: how to keep the file sharp.

This is the why — and a bigger claim I couldn't make until I'd lived it for a year. Because structuring the file is the easy part. The real question is what goes in it. And here's where most people get it wrong: they give the AI a memory of facts. What compounds — what actually makes you a better negotiator over time — is a memory of the partnership.

Most work on AI memory is about stopping the AI from starting blank — feeding it what it needs to know about your project, your case, your preferences. Useful. But that's the AI's knowledge.

What I learned to keep is something else: a record of how the two of us have learned to work together. Which decisions we settled, and which we threw out. How I think, and how I want to be pushed back on. Who's better placed to lead on what. The methods we worked out the hard way.

I call it collaboration memory — and it's the thing that turns a capable assistant into a particular partner. Yours. Shaped by a year of work.

I assumed this memory was a fix for the machine's bad memory. It isn't. It's for both of us.

The AI forgets everything. Every session starts blank — it remembers nothing of yesterday. That much is obvious. What's less obvious is that I forget too. Not the way it does — I keep my judgment, my expertise, the things years of negotiating taught me. But I lose the state: where a piece of work landed four days ago, which option we'd already weighed and dropped, the reasoning behind a decision I now only half-remember making.

So the record we keep isn't a crutch for the machine, with me as the reliable one. It's the memory of the partnership, and it re-grounds both of us at the start of every session.

I call this the shared mind: the joint sense of where things stand, written down and kept outside both our heads, so the work resumes instead of restarting. It's the same thing any close team builds over years — except a team holds it silently, in two human heads, and here you have to write it down. Because one of the partners won't carry it on its own.

Once I started looking, the collaboration memory had a clear shape — four layers, each deeper than the last:

That last layer is the engine. I bring the negotiation judgment; the AI brings a way to build something — a workflow, a structure — and I keep that skill afterward, while it keeps a deepening sense of how I reason. The earlier layers let the partnership persist across sessions. This one lets it improve.

Back to that research finding about synergy. The pair producing more than either alone — more even than a clean split of the work would predict — is not something you get for free. The research is blunt: most of the time you get the opposite. It has to be built.

And negotiators already know how to build it. Synergy is exactly what integrative negotiation creates: value that exists only because two sides combined their differences well — value neither could have made alone. The augmented negotiator runs that same negotiation, every day, with their own AI: trading on what each is better at, creating something that lives in the combination.

The memory is what makes the trade hold. Without it, you renegotiate the same terms every session — re-explaining how you work, re-deciding what you'd already settled, re-teaching what the AI forgot. With it, each round starts where the last one ended. That's when one and one start making more than two — and when the deals you walk into are better prepared than the ones you walked into a year ago.

You don't need Claude Code, or any particular tool, to do this. If you keep any running notes for your AI — a project file, a case file, a document you paste into every session — you can start keeping the partnership in it, not just the facts. The test is simple: think about your work with AI over the last few months. Is it compounding — each round sharper than the last — or are you starting over every time?

That's the whole claim, in a sentence: an augmented negotiator and their AI become more than the sum of the two — because every round of work, kept and curated, makes the next one sharper. Not the AI alone. Not you alone. The pair. Getting a little better every day.

I've written the whole argument up properly — the four layers, the research behind them, and how far the idea reaches past negotiation, into a lawyer's practice across matters or a consultant's across clients. Nineteen pages, free to read and download. It's the only place the argument is fully made:

Structuring Collaboration Memory for AI-Augmented Negotiation → https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6882538

Think about your own work with AI over the last few months. Is it compounding — each session sharper than the last — or are you starting over every time, re-explaining who you are and what you'd already decided?

Just hit reply and tell me which. I read and answer every one.

* When researchers pooled more than a hundred studies, human–AI combinations underperformed the better of the two alone on average (Vaccaro, Almaatouq & Malone, Nature Human Behaviour, 2024, https://doi.org/10.1038/s41562-024-02024-1). The pattern was replicated again in 2026 (Zhu & Zou, Journal of Economic Behavior & Organization, https://doi.org/10.1016/j.jebo.2026.107414), and research has since turned to how to structure the collaboration so it pays off (Gonzalez et al., PNAS Nexus, 2026, https://doi.org/10.1093/pnasnexus/pgag030).