In my last article I argued that AI is good news for game theory and game theory is good news for AI. Part of the reason was that as AI takes more decisions on our behalf, more strategic interactions will have an AI on both sides of the table.

In procurement that has stopped being a forecast. Walmart has been running autonomous supplier negotiations with Pactum since 2022, starting with more than 2,000 suppliers and expanding into transportation rates and goods for resale. Platforms now offer agents that negotiate price, payment terms and rebates across thousands of suppliers at once, inside guardrails the buyer defines.

This is almost always discussed as an automation story. You can finally negotiate the tail spend that never justified a buyer’s time. Cycles shrink from weeks to days. Your people concentrate on the suppliers where human judgement earns its keep. All true, and all beside the point I want to make.

Because deployment does something no procurement tool has done before. It puts the same strategic player into the market several thousand times, under one fixed decision rule. Markets learn players.

Every negotiation reveals something

Take a simple procurement agent. You give it a target, a reservation point and a set of trade-offs. It values extra payment days, rebates and price reductions at some internal exchange rate. You tell it how fast it may concede and when to escalate to a human buyer.

From the inside, those are configuration choices. From the supplier’s side of the table, they are a strategy waiting to be learned.

Nobody needs your prompt or your code. They need observations. After enough rounds it may become apparent that the agent barely moves in the first two rounds and turns noticeably flexible in the third. Or that triggering a human escalation reliably improves the supplier’s outcome. Or that price and volume are always traded at roughly the same rate.

The asymmetry is in the repetition. A human buyer meets a particular commercial situation a handful of times a year and brings different judgement each time. An agent runs a structurally identical negotiation thousands of times and brings the same judgement every time. A supplier-side AI can read all of it.

Why this is a general problem, not a procurement quirk

Most negotiation and auction formats in use today work better than theory says they should. Not because the theory is wrong, but because the people playing them are imperfect strategists. Participants overlook openings, use rules of thumb, misread the format, and quite often decide that exploiting a weakness is not worth the effort.

That has been a quiet subsidy. Badly designed formats have been protected for decades by the fact that nobody plays them very well.

Capable agents remove the protection. Calvano and co-authors showed that reinforcement-learning pricing algorithms in a standard oligopoly model consistently learn to charge supracompetitive prices, without communicating and sustained by punishment strategies nobody taught them. Fish, Gonczarowski and Shorrer repeated it with LLM-based pricing agents: same result, reached faster, no training period, and the effect carried into auction settings. Nobody programmed collusion in either case. The rules made it rational and the agents found it.

The paradox worth thinking about

The instinctive fix is to make the agent less predictable. That is not obviously good negotiation design.

Schelling’s insight was that constraint creates power. There are situations where I want the supplier to know my negotiator cannot move. When a human buyer says he cannot approve anything above 100 euros, the supplier has to judge whether that is true. When an autonomous agent is genuinely unable to accept more, the commitment is far harder to test. No relationship argument, deadline or well-timed pitch shifts the boundary. Delegating to an agent can make a constraint credible in a way a person never quite manages.

Now suppose the supplier works out that 100 is exactly where the line sits. Why would it offer 95? It offers 99.99.

The same property does both jobs. Consistency is what makes the commitment believable, and consistency is what makes the agent readable. Which narrows the real design question to this: what should the agent be predictable about, and what should a supplier never be able to infer with confidence?

That is a game-design question, and it has to be answered before deployment, not after.

Configuration is strategy

Once an agent is authorised to negotiate, prompts, guardrails, escalation rules and objectives stop being settings.

Instruct an agent never to exceed a threshold and you have made a commitment. Tell it to prioritise agreement over walking away and you have moved your bargaining position. Define when it escalates to a person and you have created a signal a supplier can learn to trigger. Run the same logic repeatedly and you have handed the other side a training set.

A recent benchmark makes the sensitivity concrete. Liang and Xu ran 9,840 LLM-to-LLM negotiations across nine models in a supply-chain bargaining problem with private buyer information, measured against a Perfect Bayesian Equilibrium. The agents created surplus reliably. How they split it was another matter: under common prompts, the buyer’s share averaged around 40% with OpenAI models, 50% with Google’s and 70% with Alibaba’s Qwen. Reversing which provider sold moved the division by 7 to 18 percentage points.

I would not extrapolate those numbers to a billion-euro supplier negotiation. But the direction is hard to ignore. Which model you deploy turns out to be a distributional decision. In most companies it is currently made on other grounds entirely.

The test most companies are running is the wrong one

The standard evaluation is a backtest. Run the last thousand negotiations through the system, compare outcomes, count savings, check the guardrails held.

That tells you how the agent performs against the world as it is today. Strategic environments do not hold still. If suppliers find that pushing to round four pays, some will push to round four. If asking for escalation works, escalation requests will rise. If they identify which variable your agent values most, they will build their offers around it. And no dataset contains the answer, because the behaviour you want to predict has not happened yet.

A better test points forward. If I were a sophisticated supplier who knew I would face this agent a thousand times, what would I change?

That question is harder and considerably more useful.

War-game the agent before your suppliers do

This is where the two halves of the previous article meet something practical.

Before releasing an autonomous negotiator into thousands of supplier interactions, build the game around it. Give supplier-side agents different costs, objectives and outside options. Let them interact with your agent repeatedly, probe it and learn. See whether they can reconstruct its concession logic. Change the escalation rules and watch whether they get exploited. Vary what the agent reveals. Make some constraints genuinely hard and leave others deliberately loose. Then look at what happens once the other side has adapted.

AI is what makes this affordable: simulating thousands of adapted opponents used to be the expensive part, and no longer is. Game theory is what tells you where to look. Incentives, information, commitment, outside options, learning and strategic response are the variables that decide whether a format survives contact with a competent counterparty.

In the last article I argued that once an AI starts asking how the other side will respond, it has entered the domain of game theory. The mirror image is the part worth planning for. The moment you put an agent into a negotiation, your suppliers start asking exactly that question about your AI. The answer is already written down in its configuration.

---

Sources

Van Hoek, R. et al. (2022). How Walmart Automated Supplier Negotiations. Harvard Business Review.

Calvano, E., Calzolari, G., Denicolò, V. & Pastorello, S. (2020). Artificial Intelligence, Algorithmic Pricing, and Collusion. American Economic Review 110(10), 3267–3297.

Fish, S., Gonczarowski, Y. A. & Shorrer, R. I. (2024). Algorithmic Collusion by Large Language Models. arXiv:2404.00806.

Liang, C. & Xu, F. (2026). When LLM Agents Negotiate: Private Information and Dynamic Bargaining in Supply Chains. arXiv:2608.07538.

Schelling, T. C. (1960). The Strategy of Conflict. Harvard University Press.