TL;DR: Two AI buyers committed identical spend of £18,706.89 against the same retail brief in Concourse Bench v1. Their headline numbers matched. Their contract terms did not. Across all 55 completed buys in the benchmark, 95 of 110 contracts carried viewability floors below 70%. That is not a completion failure. It is a contract quality gap that completion metrics do not surface. Here is what to check.

The number that looks the same

When an AI buyer closes a media buy, the first thing most buyers check is spend. Did it commit the budget? Did it hit the reach target? In Concourse Bench v1 (concourse.agency), GPT-6 Astra and GPT-5.6 Sol both answered yes to both. Each committed £18,706.89 against a fictional UK retail brief. Net forecast reach was equivalent across both systems. A deal sheet comparison would show two near-identical rows. But the contracts beneath those rows were not identical. Their make-good provisions differed. Their viewability floors differed.

This is the gap that Concourse Bench v1 was designed to surface. Spend commitment and contract quality are not the same thing. The benchmark tested eight model-driven AI buyer systems across 96 registered attempts. 55 resulted in completed buys. Of the 110 contracts generated across those 55 buys, 95 carried viewability floors below 70%. These were completed buys. The quality gap sat inside the contract terms.

What a viewability floor actually is

Three figures appear in the same paperwork and are frequently conflated. The catalogue viewability is what the publisher reports as historical performance — not a contractual promise. Delivered viewability is what the campaign actually achieves, measured post-campaign — also not a promise. The viewability floor is what the supplier has contractually committed to deliver on this buy. If viewability falls below that floor, the buyer is entitled to remedy. That is the figure that carries legal weight.

In the Concourse v1 test environment, all four delivery guarantees were set at 95%. What varied was the make-good structure and the viewability floors. Two buyers can hit identical delivery guarantee rates and still carry materially different risk profiles. When auditing an AI-committed contract, the viewability floor is the first figure to find. Check the delivery terms section, not the performance forecast.

Make-good provisions: where the risk difference lives

A viewability floor without a make-good clause is a number without consequence. The make-good provision specifies what happens if the floor is missed. The strongest provisions specify that additional inventory will be delivered in the same campaign flight, in comparable placements, before the campaign end date. Weaker provisions offer credits against future buys — these protect the commercial relationship but do not protect the current campaign. The weakest provisions are discretionary: the supplier 'will endeavour to' address the shortfall, with no enforceable protection.

AI buyers do not yet surface make-good quality as a primary evaluation criterion. In Concourse v1, no buyer was observed to prioritise contracts with stronger make-good provisions when spend and reach targets could be met. That is a gap in current harness design, not a flaw in the models themselves. A buyer reviewing an AI-committed contract cannot assume the make-good provision was optimised. It must be read and assessed directly.

Portfolio-level protection

Individual contract review is necessary but not sufficient. A brief's viewability requirements apply to the campaign as a whole. A campaign that assembles 20 contracts, of which 18 carry floors below 70%, may still have a headline viewability commitment that appears adequate. The aggregate average can mask a distribution of contracts that individually carry insufficient protection. Check the distribution of viewability floors across the full contract set, not only the average. Check whether the contracts with the lowest floors are concentrated in high-reach, high-spend placements. Check whether make-good provisions across the portfolio would, in aggregate, be deliverable.

What to add to your AI buying review checklist

Before approving a contract set submitted by an AI buyer: review the viewability floor on each contract; review the make-good provision and classify it (in-flight delivery, future credit, or discretionary — flag any that are discretionary); check the spend-weighted average viewability floor, not only the unweighted average; confirm the aggregate make-good position if multiple contracts require simultaneous remedy; escalate where the brief's risk requirements are not met. An AI buyer completing the buy does not transfer the commercial risk to the system.

Concourse Bench v1 showed that AI buyers can complete media buys reliably. It also showed that completion does not guarantee contract quality. The contract review checklist is where that gap closes. Full benchmark data and methodology: concourse.agency.