We Let Two AIs Haggle Over a Purchase Order. Here's Who Won.

What is agent-to-agent commerce?

It is a transaction where software negotiates with software. Instead of a person emailing a supplier to argue over price and terms, an AI buyer talks to an AI seller, they settle on a deal, and the people involved read the outcome later like a receipt.

To be clear about the framing: think of “we let two AIs haggle” as an illustrative thought experiment, not a claim that we ran a specific lab experiment of our own. The serious version of this exists in research. A Columbia Business School simulation study looked at AI agents negotiating with each other, which is where a lot of the current interest comes from.

The idea sounds futuristic and is mostly plumbing. Buying already involves a back-and-forth over price, quantity, and terms. Agent-to-agent commerce just asks what happens when both sides of that back-and-forth are automated and fast.

So who wins when two AIs negotiate a purchase order?

Whoever was given the better rules. That is the honest answer, and it is more useful than a settlement figure. A negotiation between two agents is not decided by charm or persistence. It is decided by the constraints each one is carrying: its floors, its ceilings, its instructions on what to concede and when.

We will describe the dynamics qualitatively, because inventing a settlement number would be dishonest. But the shape is intuitive. An agent told to protect a hard budget ceiling behaves very differently from one told to close at almost any cost. An agent with a firm floor on price will not sell below it, no matter how the other side pushes. The “winner” is usually the side whose owner thought harder about the boundaries before the negotiation started.

Which flips the interesting question. It is not “which AI is smarter.” It is “whose human wrote better rules.” The negotiation is downstream of the setup.

Where does the risk actually live?

In the rules you give the agents, not in the negotiation itself. An automated negotiation is only as safe as the boundaries around it. Get those wrong and speed just helps you lose faster.

The boundaries that matter:

Budget floors and ceilings

A ceiling caps what your buyer agent will ever agree to pay. A floor stops your seller agent from giving product away. Without both, an agent optimizing to “close the deal” can close a bad one, quickly and repeatedly.

Terms, not just price

Price is the obvious lever, but terms carry real money too: payment timing, minimum quantities, delivery windows. An agent that wins on headline price and loses on terms has not actually won. The rules need to cover the whole deal, not one number.

Approval limits

Above some size, the deal should stop and wait for a person, the same principle that protects any automated purchasing. Let agents handle the small and routine. Keep a human above the limit for anything large enough that being wrong hurts.

The pattern here is the same one that runs through all automated purchasing. The opportunity and the risk both live in the constraints. Good constraints turn automation into leverage. Missing constraints turn it into a fast way to make expensive mistakes.

What OpsMavix builds for this

Purchasing and supplier negotiation are getting automated whether or not you are ready, and the useful lesson is not “agents are coming.” It is that automated buying is only as good as the rules underneath it. That is the layer OpsMavix builds.

We build custom internal software for purchasing and order management where the boundaries are explicit: budget ceilings, spend limits, approval workflows, reorder controls. Whether a human or an agent is doing the buying, the system holds the line on what is allowed. You do not have to trust that everyone, or every bot, will behave. You encode the limits once and let the software enforce them.

If you want to see where your purchasing has no rules protecting it, Book a free Operations Leak Audit. If you are thinking about handing software more control over buying, our Control Pilot approach is built to keep the limits, and a human, in charge.

FAQ

Did OpsMavix really run two AIs against each other?

No. Treat that framing as an illustrative thought experiment. The real research it points to includes a Columbia Business School simulation study on AI agents negotiating, which is worth knowing about as this idea develops.

Is agent-to-agent commerce actually happening?

It is an emerging idea getting serious attention around September 2026, more explored than widespread. The reason to care now is that the controls you would need already apply to normal automated purchasing.

What decides which agent wins a negotiation?

The rules each one carries: budget floors and ceilings, terms, and approval limits. A negotiation between agents is decided by the constraints their owners set, not by the agents being clever.

How do I keep automated purchasing safe?

Encode explicit boundaries: spend limits, budget ceilings, and approval thresholds above which a human signs off. The negotiation can be automated. The limits should be deliberate and enforced by the system.