The phrase "AI-negotiated deal" gets used as though the negotiation is a single event. In practice, it is a sequence. An AI buyer working through a programmatic deal does not submit one offer and wait for a yes or no. It iterates: probing price, checking constraints, revising terms, and eventually reaching a commitment or walking away. Understanding what that sequence looks like in detail matters for anyone evaluating whether to deploy an AI buyer or accept inventory bought by one.

What does a multi-round AI negotiation actually involve?

A multi-round negotiation is a structured exchange in which both sides adjust their position in response to what the other party communicates. For an AI buyer, this means the model reads incoming deal terms, evaluates them against the campaign brief, identifies the gap between what is offered and what is required, and generates a counter-position. It then sends that counter-position and waits for a response. Each exchange is one round. A negotiation might run through two rounds or eight, depending on how far apart the opening positions are and whether a deal is reachable at all.

The key distinction from a simple bid is that the buyer maintains context across rounds. It remembers what it offered in round one, what came back in round two, and adjusts its next position accordingly. This context-carrying capacity is what separates a negotiating agent from a standard bidder responding to a fixed floor.

How does the opening offer work?

The opening offer is generated from the campaign brief. The AI buyer reads the targeting parameters, the CPM range, the placement specifications, and the performance targets, and produces an initial position that reflects what the buyer would accept at best terms. This is rarely the limit of what the buyer will pay. The opening offer is a starting point, not a ceiling.

A well-designed buyer harness constrains the opening offer so it stays within the brief. If the brief sets a maximum CPM of £18, the opening offer should not come in at £22 on the grounds that it might be negotiated down. The harness enforces the brief as a boundary condition, not a target. The model works within that boundary, not around it.

What happens when the counter comes in?

The seller's counter triggers a re-evaluation. The AI buyer parses the new terms, identifies which parameters have moved and which are unchanged, and assesses whether the gap is closable. If the counter meets the brief, the buyer accepts. If it does not, the buyer calculates a revised offer. This revised offer should move toward the seller's position while still satisfying the brief constraints.

Where this gets technically demanding is in the reasoning about trade-offs. A counter might accept the CPM but add a frequency cap the buyer finds restrictive, or it might lower the floor but widen the geo targeting beyond what the brief allows. The buyer must evaluate the package, not just the price. Models that evaluate only headline CPM will miss compound disadvantages buried in secondary terms.

How does a budget constraint change the negotiation?

Budget constraints introduce a hard limit that the buyer must communicate clearly and maintain consistently across rounds. If the campaign has £40,000 left and the seller is offering inventory at a rate that would exhaust it before the flight end date, the buyer faces a binary decision: adjust the volume commitment or hold the rate and accept fewer impressions.

The constraint is not just about the current negotiation. A buyer managing multiple concurrent deals must account for what is already committed elsewhere. If £30,000 is committed in other deals, the £40,000 total budget means the available headroom for this negotiation is £10,000, not £40,000. A buyer that fails to carry this portfolio awareness into each individual negotiation will routinely overspend or make commitments it cannot honour.

What is a portfolio check and why does it matter?

A portfolio check is the step in which the buyer reviews its full set of existing commitments before advancing a new offer. It answers the question: if I commit to this deal on these terms, does my overall position still make sense? That involves checking budget headroom, reach overlap with existing deals, frequency implications across inventory sources, and whether this deal displaces something already committed that performs better.

Portfolio checks are computationally straightforward but require that the buyer harness maintains an accurate real-time record of all live commitments. If that record is stale or incomplete, the portfolio check fails silently. The buyer believes it has more headroom than it does and commits accordingly. The error only surfaces when delivery runs against actual budget limits or when inventory overlaps appear in reporting.

How does the negotiation reach commitment?

Commitment happens when the buyer's current offer and the seller's current position meet within tolerable distance on all material terms. That includes price, volume, placement, targeting, and any performance minimums set in the brief. When all terms are within range, the buyer signals acceptance and the deal moves to execution.

The commitment step is where the deal record is created. In a well-governed system, the agreed terms are written into a deal record that both parties can access and that captures the full negotiation history: every offer, every counter, and the final accepted position. This record is the evidence base for any later dispute about what was agreed and who approved it.

What does Concourse Bench v1 tell us about real negotiation sequences?

The most rigorous publicly available data on how AI buyers actually perform in negotiation sequences comes from Concourse Bench v1. The benchmark tested eight model-harness combinations across 96 registered negotiation attempts, of which 55 reached completion. The variance in completion rates across models is significant: some combinations completed every attempt in the test set, while others completed none. This is not primarily a model capability gap. It reflects how well the harness layer enforces brief compliance, maintains context across rounds, carries portfolio state, and handles edge conditions like seller-side delays or ambiguous term changes.

The negotiation sequence described above is not a theoretical construct. It is the sequence that the Concourse Bench evaluation is designed to stress-test. Each of the failure modes observed in the benchmark corresponds to a specific point in the sequence where the buyer's reasoning or the harness's enforcement broke down: opening offers that exceeded brief constraints, counters that missed compound term changes, budget overruns caused by portfolio state errors, and commitments made without adequate record of the agreed terms.

Understanding the sequence in detail is the starting point for understanding where AI buyers succeed and where they fail. The negotiation is not an event. It is a process with specific steps, specific decision points, and specific failure modes at each stage. Buyers and sellers evaluating AI-negotiated deals should be asking which steps the system handles reliably and which it does not.