Nothing matches that. Try a different word or tag.
Sep · 4 min read
Agent negotiation changes what a constrained budget can access by removing intermediary layers and reducing the execution cost per buy. Here is what that looks like in practice.
Sep · 5 min read
Concourse Bench v1 evaluates model-harness combinations, not models alone. The software layer that wraps an AI buyer is as consequential as the model it runs.
Sep · 6 min read
A step-by-step account of how AI buyers negotiate media deals: opening offer, counter, budget check, portfolio review, and commitment.
Sep · 5 min read
55 completions from 96 attempts. 95 of 110 contracts with sub-70% viewability floors. The first rigorous public data on how AI media buyers actually perform.
Sep · 5 min read
95 of 110 AI-negotiated contracts had sub-70% viewability floors. In CTV, where standards are more stringent, that gap is a serious problem. Here is what the contract needs to fix it.
Sep · 5 min read
Fragmented supply, deal-based inventory, and tight targeting make CTV the most demanding environment for a model-driven buyer. Here is why, and what the data shows.
Sep · 6 min read
The claim "AI can buy media" was untested. Concourse Bench v1 is the evidence base that makes every other piece of Alkimi's content defensible.
Sep · 6 min read
Reading a brief is the easy part. Holding it as a binding constraint across multiple parallel negotiations is where models diverge.
Sep · 6 min read
Each stage has distinct failure modes. A completion rate alone does not tell you which stage broke.
Sep · 7 min read
41 of 96 attempts in Concourse Bench v1 did not produce a completed buy. The failure modes are distinct, and the distinction matters for what you fix.
Sep · 4 min read
Concourse Bench v1 tested eight model-driven buyers in a simulated market. Fifty-five of 96 attempts produced a completed buy. Two models completed every attempt. Two completed none. Which AI, and how reliably, is the right question.
Sep · 4 min read
Concourse Bench v1 tested eight model-driven buyers across four buying situations. Sol and Fable completed all 12. Opus completed 10 of 12. Astra and Sonnet each completed 9 of 12. Terra completed 3 of 12. Luna and Haiku completed zero.
Sep · 5 min read
'Agentic media buying' describes a wide range of things. Concourse Bench v1 provides the first structured evidence on what an AI buyer actually does in a workflow. The picture is more specific and constrained than most uses of the phrase suggest.
Sep · 5 min read
Concourse Bench v1 recorded three ways an AI buyer fails: budget overrun, missed brief requirement, and technical interruption. These are not the same failure. V1 reports them separately. That separation is the point.
Sep · 4 min read
Across 110 contracts in the 55 completed buys, 95 promised a viewability floor below 70%. The agents finished the workflow but left the buyer poorly protected on a basic contractual measure. Completion and quality are not the same score.
Sep · 5 min read
Concourse Bench v1 recorded model API costs for eight AI buyers. Among completed buys, API expense ranged from $0.74 to $14.38 per buy. Total known API cost was $390.16. These figures are model API expense only, not full operating costs.
