Brand safety in programmatic advertising has always involved a question of accountability. When a brand's ad appears next to content it should not be near, the question that follows is: who is responsible? In human-executed buying, that question has a clear starting point. The trader who committed to the inventory, the platform that served it, or the publisher that hosted it. In agent-negotiated buying, the question becomes more complicated, because the entity that committed to the inventory is a software system acting on behalf of a human organisation.

Who is accountable when an AI buyer commits to unsafe inventory?

The legal and commercial accountability for a deal committed by an AI buyer sits with the organisation that deployed the buyer and authorised it to act. The AI system is not a legal entity. It cannot hold a contract. The organisation that gave the system authority to negotiate on its behalf takes responsibility for the commitments the system makes, in the same way that an organisation is responsible for the actions of a human employee acting within their designated role.

This accountability structure makes the deal terms and the approval record critical. If the brief defined brand safety requirements, if the harness enforced those requirements, and if a human approved the deal at the appropriate gate, then the organisation can demonstrate that its governance system functioned correctly. If the brief was vague on brand safety, or the harness did not enforce those requirements, or the deal was approved without adequate information, the accountability is harder to defend.

What should a brief specify about brand safety?

A brief that leaves brand safety requirements unspecified or loosely defined creates a governance gap that will be exploited, not necessarily by a bad actor, but simply by a system optimising on the parameters it has. If brand safety is not in the brief as a constraint, the agent may legitimately commit to inventory that a human reviewer would immediately reject.

A well-specified brief defines brand safety at the level of the deal term, not just the campaign objective. That means explicit content category exclusions, minimum viewability floors, placement restrictions, and any publisher or domain blocklists that apply. Each of these should be stated as a constraint that the harness enforces, not as a preference that the model weighs against price efficiency.

What does the Concourse Bench v1 viewability finding reveal about current practice?

One of the most significant findings from Concourse Bench v1 is about viewability floors in the contracts committed by AI buyers. Of the 110 contracts examined in the benchmark, 95 had viewability floors below 70%. The industry standard for viewability in display advertising is typically set at 50% by MRC definition, but most sophisticated buyers aim for higher floors, particularly in brand-safety-sensitive categories where the quality of exposure matters as much as the volume.

A 70% threshold is not an unusually demanding standard. It represents a basic quality floor for inventory that a brand-conscious buyer should be maintaining. The fact that 95 of 110 contracts came in below that floor suggests that in most of the tested buying scenarios, viewability was not functioning as an enforced constraint in the deal structure. It was either absent from the brief, present but not enforced by the harness, or overridden in favour of lower CPM.

How do deal terms protect the buyer after commitment?

Brand safety protection does not end at the point of commitment. A deal structure that includes explicit performance guarantees, a clear remediation mechanism for brand safety failures, and delivery confirmation against committed terms gives the buyer meaningful recourse if the inventory does not perform as specified.

Without those terms in the deal record, the buyer has committed to volume and price but not to quality. If viewability falls short, or content adjacency fails a brand safety check, the buyer has no contractual basis for remediation. The deal record needs to capture what was agreed about quality, not just what was agreed about price and volume.

What does the approval gate need to show for brand safety?

An approval gate that does not surface brand safety information to the human reviewer is not protecting the buyer. The reviewer needs to see, at the gate: the inventory source and placement context, the viewability floor committed in the deal terms, any content category adjacency flags raised by the harness, and a comparison between the committed terms and the brand safety requirements in the brief.

If this information is not present at the gate, the human approval is effectively blind to brand safety. The reviewer approves a price and volume commitment without knowing whether the inventory meets the quality standard. That is the scenario in which brand safety incidents occur in agent-negotiated campaigns, and it is entirely preventable if the approval interface is designed to surface the right information.

What does a well-governed agent-negotiated deal look like for brand safety?

A well-governed deal starts with a brief that specifies brand safety requirements as enforceable constraints. The harness validates every proposed deal term against those constraints before any offer is made. The approval gate surfaces the brand safety terms alongside the price and volume terms, so the reviewer can confirm both. The deal record captures the committed brand safety terms bilaterally. Delivery monitoring compares actual performance against the committed floors and flags any shortfall.

This is not a more complex version of existing programmatic practice. It is the same discipline applied to a new execution layer. Brand safety governance in agent-negotiated deals requires the same clarity of specification, the same enforcement rigour, and the same complete record-keeping that good human-executed buying requires. The difference is that in an agentic system, if those disciplines are not built into the structure of the deal from the start, there is no human discretion to catch the gaps at the execution stage.