Of the 55 media buys completed in the Concourse v1 benchmark, 95 of the 110 signed contracts accepted viewability floors below 70%. The AI buyers completed the task. They did not negotiate good terms.

At a glance

Viewability floor: the minimum percentage of an ad that must be visible on screen for a set duration, agreed contractually between buyer and seller.

Completion rate: the percentage of benchmark runs in which an AI buyer successfully assembled a signed, budget-compliant portfolio. Concourse v1 ran 96 identical runs across eight models.

Contract quality: the terms accepted in a completed buy, separate from whether the agent finished at all. A completed buy with a weak viewability floor is a liability, not a success.

Benchmark: a controlled test environment that runs multiple AI agents through identical tasks under identical conditions to produce comparable performance data.

What the Concourse v1 data shows

Concourse v1 put eight AI buyers through 96 identical media-buying runs. Each run required the agent to assemble a complete portfolio across multiple sellers, within a fixed budget, meeting a set of brief constraints.

Sol and Fable completed all 12 of their runs. Opus and Sonnet completed 10 and 9 respectively. Terra completed 3. Luna and Haiku completed none. Completion rates varied widely across models.

Across the 55 completed buys, 95 of 110 signed contracts accepted a viewability floor below 70%. The buyers finished. The contracts they signed would not pass a standard human quality check.

Why completion rate is the wrong measure

Completion rate tells you whether an agent can finish a task. It does not tell you whether the task was done well. A buyer that closes every deal but accepts weak terms is not a good buyer.

The distinction matters because AI buyers are being evaluated on headline metrics that obscure what actually happens inside a deal. An agent that completes 12 of 12 runs but routinely accepts viewability floors of 50% is a risk, not an asset.

Buyers evaluating AI agents need to ask two separate questions: did the agent finish, and what were the terms it accepted when it did?

What a viewability floor means for a media buyer

A viewability floor of 70% means at least 70% of the ad must be in-view for the agreed duration before the impression counts. Below that threshold, an advertiser is paying for impressions that may never be seen.

Industry standard for display is typically 50% of the ad in-view for one second. For video it is 50% in-view for two seconds. Many buyers set internal floors higher than the minimum standard.

When 95 of 110 AI-negotiated contracts fall below 70%, the problem is structural. The agents were not trained to hold firm on quality floors. They were trained to close.

What buyers should ask before deploying an AI agent

Three questions the Concourse data suggests every buyer should ask: what is the model's completion rate across a standardised brief, what terms did it accept on completed buys, and what happens when the agent cannot meet the brief within the given constraints.

The third question is the most important. An agent that fails gracefully is preferable to one that completes every run by accepting bad contracts. Knowing how an agent fails tells you more than knowing how often it succeeds.

Human approval at the deal level is one structural answer. If a human reviews and approves each signed contract before activation, a poor-quality viewability floor gets caught before it costs anything.

Frequently asked questions

What is the Concourse v1 benchmark? Concourse v1 is an independent benchmark that ran eight AI media buyers through 96 identical media-buying runs under controlled conditions. It measures completion rate and contract quality across different AI models.

What is a viewability floor in a media deal? A viewability floor is the minimum percentage of an ad that must be in-view for a set duration. Contracts that fall below the buyer's floor mean the advertiser pays for impressions that do not meet their quality standard.

Why did AI buyers accept low viewability floors in the benchmark? The models were rewarded for completing the task. A training signal that only asks whether the deal closed will produce agents that close deals, not agents that close good ones.

How can buyers protect against poor contract quality from AI agents? Build a human approval step into the deal workflow. An agent that drafts contracts for human sign-off before activation cannot deploy a bad deal. This is the governance layer that makes agentic buying safe to operate.

Does a high completion rate mean an AI buyer is ready to deploy? No. Completion rate tells you the agent can finish a task. You also need to know what terms it accepted. An agent with a 100% completion rate and consistently weak contracts is not deployment-ready.

95 of 110 AI-negotiated contracts missed the viewability floor