9 Nov 2026 · 4 min read
What to Ask Before You Trust an Agency's AI Buying Agent
Agencies pitching AI buying agents are not short on confidence. What they are often short on is anything a client can independently check. An AI buying agent, in plain terms, is software an agency uses to carry out parts of the media buying process: choosing inventory, setting a price, pacing a budget, with less manual intervention than a person doing the same job by hand. When an agency says it works, that is a commercial claim made by the party selling the service, and it only means something once a second party can verify it. Here is what a client should ask for before accepting it.
Ask where the benchmark came from
A benchmark is a structured test that measures an agent's decisions against a defined standard or against other agents, under conditions set in advance, rather than measuring the agency's own live campaign and calling the result proof. This matters because live campaign numbers have no comparison point built in: a CPM looks good or bad only relative to something, and if the agency chooses what it is relative to, the comparison is not independent. Ask who ran the benchmark, the agency itself or a separate party, and what the agent was tested against. Concourse, a live agent benchmarking product, is one example built to sit outside the agency doing the selling. The question is simple: show the benchmark, not the headline drawn from it.
Ask for the decision audit trail, not just the outcome
An audit trail is the record of what the agent actually considered and chose at each step, distinct from the final result it produced. A client asking whether the agent performed well should also ask why it made a specific choice, and what alternatives it passed over first. An agency that can show impressions delivered and the price paid, but not what the agent evaluated beforehand, is offering an outcome claim rather than a verifiable one. Outcomes can be good by chance; a defensible decision process is good by design.
Ask who holds the deal record, and whether it can be retrieved independently
When a buy-side agent and a sell-side agent agree a deal, that agreement needs a record neither side can quietly edit afterwards. A deal record is only useful if timestamped at the moment of agreement and held by both parties, rather than filed inside the agency's own system where only the agency can see it. The DealSheet, Alkimi's own deal record, is written at the point of agreement, time-stamped, and held by both the buy-side and sell-side agent, so either party can retrieve it directly rather than relying on the other side's account. The test is whether the client, not just the agency, can pull that record alone.
Ask exactly where a human has to approve, not just where one is available
Human oversight as a concept is not the same as a human approval point. Ask, deal by deal, where a named person is required to sign off before the agent proceeds, as against where the agent acts first and a person reviews a report afterwards. The second arrangement is not the same protection, because by the time a person reviews it the deal has already happened. A credible answer names the specific stage, typically the deal itself, at which approval sits with a person, and names that person within the agency's process. A vague answer along the lines of our team monitors performance is not an answer.
Ask how the agent fits a published standard, not whether it has a badge
Two efforts now exist to standardise how buying and selling agents exchange information with each other. AAMP, the Agentic Advertising Management Protocols announced by IAB Tech Lab in early 2026, sets out a shared framework for agentic transactions. AdCP, the Ad Context Protocol governed by AgenticAdvertising.org, does related work from a separate body and has reached version 3.2 as of October 2026. Neither protocol certifies that an agent makes good decisions; both standardise how agents pass information to each other, a different question from whether the decisions produced are any good. There is no compliance badge for either standard, so be cautious of any agency presenting a certification as proof of performance. A fairer question is which framework the agent operates within today, and what it is being built toward.
None of this assumes an agency is lying. It is the ordinary discipline applied to any commercial claim made by the party being paid for it: ask for the independent benchmark, the record of the decision rather than just its result, a deal record either side can retrieve on its own, and the point at which a person, not a dashboard, had to say yes. An agency with a real answer to all four will usually give one without being asked twice.