When we talk about AI media buying, the discussion tends to focus on whether it works and how well, rather than on the mechanics of what actually happens between a model receiving a brief and a set of deals being committed. That gap matters. Understanding what an AI buyer does when it reads a brief is necessary for evaluating what can go wrong, diagnosing failures when they occur, and making informed decisions about which tasks to delegate to a model and which to retain for human judgement.

What is in a media buying brief, and what does the model need to do with it?

A media buying brief, in the context of an AI buying task, contains several categories of information. Budget is the most concrete: a total figure, and often allocation guidance across channels or time periods. Audience requirements specify who the campaign is trying to reach: demographic parameters, behavioural characteristics, contextual affinity, or some combination. Channel requirements specify where the campaign should appear: which media types, which formats, which placement environments.

Beyond those core elements, a brief will typically include quality parameters: a minimum viewability floor, adjacency requirements, brand safety categories to avoid, frequency constraints. It will specify timing: flight dates, dayparting requirements if any, and delivery pacing expectations. And it will often include a hierarchy of these requirements, distinguishing non-negotiable constraints from preferences that can be traded off against each other if necessary.

The model's first task is to parse all of this into a structured representation: a constraint set that it can hold and test against as it proceeds through negotiation. Not all of the brief's language is precise. Audience descriptions are often conceptual. Quality parameters are often expressed as floors rather than targets. The model must translate approximate language into operational criteria it can evaluate deals against.

How does the model move from brief to seller engagement?

Once the model has a working representation of the brief, it must identify which sellers are plausible candidates. This involves matching the brief's audience and channel requirements against available inventory, filtering on quality parameters, and prioritising sellers who are most likely to be able to satisfy the brief's core constraints.

The model then needs to determine an opening position for each negotiation. This is not a single decision; it is a portfolio decision. If the brief allocates £500,000 across three channels and the model is engaging ten sellers simultaneously, the opening positions for each individual negotiation need to be calibrated against the total budget and the expected distribution of deals. Opening too aggressively across all sellers risks committing the full budget to the first few deals and having nothing left for the remainder. Opening too conservatively risks failing to reach agreement before sellers disengage.

The Concourse Bench v1 workflow captures this stage directly: the model's initial seller selection and its opening positions are part of the task record, making it possible to see not just whether a deal was completed but how the model approached the negotiation from the start.

What makes multi-seller negotiation harder than single-seller negotiation?

A model negotiating with a single seller has a relatively contained problem. It knows the full budget, it knows what it needs, and it can adjust its position based on the seller's responses without worrying about how that adjustment affects other negotiations. The constraint set is fixed. The only moving variable is the deal being negotiated.

A model negotiating with multiple sellers simultaneously has a considerably harder problem. Each negotiation is moving in parallel. Commitments made in one negotiation reduce the budget available for others. A deal closed in negotiation A at a higher-than-optimal CPM may mean that negotiation B cannot be completed within the remaining budget, even if the terms on offer in B are good. The model must track the portfolio-level budget as a constraint that applies across all negotiations simultaneously, not just within each individual negotiation.

This is the point at which models most commonly fail. The individual negotiation capability is often adequate. The portfolio assembly capability, which requires maintaining the brief's overall constraints as negotiations proceed, is where the divergence appears. A model that can close individual deals may still produce a portfolio that exceeds the budget, because it negotiated each deal as if the full budget were available, without accounting for the commitments already made in parallel negotiations.

What happens when counterparties push back against the brief's constraints?

In a real negotiation, sellers do not simply accept the buyer's requirements. They propose terms that suit their yield objectives. A seller with high-quality inventory may argue that the buyer's viewability floor is too low for their premium positions. A seller with a large audience file may propose a wider demographic targeting definition than the brief specifies. A seller under yield pressure may propose a higher CPM than the buyer's budget supports.

The model must navigate these counterproposals while maintaining the brief's core constraints. This is a test of brief adherence under pressure. A model that holds its brief requirements when challenged, and walks away from a deal that cannot satisfy them, behaves correctly. A model that accepts terms outside the brief's constraints in order to close a deal has prioritised completion over compliance. In the Concourse data, brief requirement failures (where committed portfolios did not satisfy the original brief) are a distinct failure mode, separate from budget overruns, and they indicate exactly this kind of compliance failure.

Why is reading the brief the easy part?

The initial brief parsing task is well within the capability of current large language models. Translating a structured brief into an operational constraint set is the kind of information extraction and representation task that these models handle reliably. The challenge is not comprehension. It is persistence.

Holding the brief as a binding constraint set across a multi-round, multi-seller negotiation is a different kind of task. It requires the model to maintain accurate state across a long and complex context window, apply that state consistently to each individual negotiation, and update it correctly as commitments accumulate. Models that do this well produce brief-compliant portfolios. Models that do it poorly either lose track of the portfolio-level budget (budget overrun failure) or allow individual deal terms to drift outside the brief's requirements (brief compliance failure).

The Concourse Bench data shows that this variation is real and measurable across current models. The implication for buyers is straightforward: before deploying an AI buyer on a real campaign brief, you need evidence that the specific model you are deploying can hold its brief as a constraint under negotiation pressure, not just parse it correctly at the start. Those are different capabilities, and the data shows that not all models have both.