Most KPI frameworks for media buying were designed when the buyer was a human. Some of those metrics carry forward unchanged when the buyer is an AI agent; a few become more important; and several new ones emerge that the standard framework does not capture at all. The organisations getting this right are distinguishing between the performance of the underlying campaign and the performance of the agent doing the buying, and reporting on both separately.

TL;DR: Campaign performance KPIs largely hold when buying is automated. What changes is the layer above: the metrics that measure the agent's decision quality, deal completion behaviour, and audit trail integrity. These are not optional additions; they are the only way to know whether the automation is working correctly.

Which KPIs stay the same?

Viewability, completion rate, cost per completed view, cost per click, and brand safety metrics carry forward unchanged. These measure the performance of the ad against the audience, not the process of acquiring the inventory. Whether a human or an agent agreed the deal, the ad either viewed or it did not. Any framework that abandons these in favour of agent-specific metrics has misunderstood where automation sits in the process.

Which KPIs become more important?

Deal completion rate becomes more important, because agents negotiate and close deals at a scale and pace that humans cannot monitor individually. A human buyer who abandons a negotiation is visible to the team; an agent that fails to complete deals may not be. Similarly, deal accuracy (the ratio of deals executed to agreed terms versus deals that required amendment after the fact) becomes a primary metric rather than an audit footnote. Deviation at scale is a signal that the agent's parameters need adjustment, and it is very difficult to catch without systematic measurement.

What new metrics does agent buying introduce?

Three categories of metric emerge that the standard media buying framework does not capture. The first is approval rate: what proportion of agent-proposed deals are approved by the human-in-the-loop before execution, and what proportion are rejected or amended. A low approval rate suggests a misalignment between the agent's parameters and the buyer's actual intent. The second is audit trail completeness: for every deal closed, does the record contain the full decision chain, the agreed terms, and the approval authority? Incomplete audit trails represent an operational and compliance risk that grows with deal volume. The third is agent error rate: the number of deals executed outside approved parameters, expressed as a proportion of total deals. In a production deployment this should be close to zero; if it is not, the agent's configuration requires review.

What does Concourse Bench v1 suggest about agent performance metrics?

Concourse Bench v1 is, to date, the most detailed public benchmarking exercise of AI models performing agentic advertising tasks. The benchmark tested 8 models across 96 attempts, resulting in 55 completions. The distribution of completions is instructive for anyone designing an agent evaluation framework: Sol and Fable completed 12 out of 12 tasks; Opus completed 10 out of 12; Astra and Sonnet completed 9 out of 12; Terra completed 3 out of 12; Luna and Haiku completed 0 out of 12.

The spread from 100% to 0% completion across models on the same task set makes model selection a first-order variable in agent performance, not an implementation detail. Completion rate, in Concourse Bench terms, is the primary metric, and the variance between models suggests it should be measured continuously in production, not assumed from a pre-deployment test against a single model version.

How should teams report on agent-negotiated campaigns?

Reporting should be structured in two distinct layers. The first layer covers campaign performance as standard: delivery, viewability, cost efficiency, and audience metrics. The second layer covers agent performance specifically: deals negotiated, deals completed, deals approved and rejected, audit trail completeness, and error rate. The two layers should be reported separately and reviewed by different owners. Campaign performance sits with the media planning team; agent performance sits with the technology or operations team. Conflating them makes it impossible to isolate whether a performance problem is a campaign problem or an agent configuration problem, which are very different things requiring very different responses.

Alkimi tracks audit trail completeness and deal completion as core metrics in every agent-negotiated campaign, alongside standard campaign performance reporting. Both layers are available to buyers and sellers from the shared DealSheet record, giving each party visibility into the full picture without relying on one side's system log.