Before anything else, three words get thrown around like everyone already knows what they mean. They don't, so here they are, properly.

Agent. A piece of software that does a job for you with some independence, instead of waiting for you to click every button. You give it a goal and some rules. It figures out the steps.

Model. The "brain" the agent uses to think. When people say GPT, Claude, or Gemini, they mean a model. The agent is the worker, the model is what the worker thinks with.

Token. The unit a model's thinking gets measured and billed in, roughly a word or part of a word. Every time the model reads something or writes something, that costs tokens, and tokens cost money. This is the detail almost nobody explains, and it is the one that matters most below.

With those three out of the way, here is agentic advertising with no further acronyms. A human tells an agent roughly what it wants: "buy this kind of ad space, for this budget, with these rules." The agent then goes and does the legwork a media buyer used to do by hand: finds inventory, talks to the seller's agent, agrees a price, checks the terms are inside the rules it was given.

That is the whole idea. Everything else you hear about, AdCP, AAMP, A2A, DealSheets, is plumbing for making that conversation reliable and auditable. It is not the idea itself, and you do not need to understand any of it to understand what is coming next, which is the part that actually costs you money.

Why an agent's thinking has a price tag

A language model does not think for free. Every time an agent reads a proposal, drafts a counter-offer, or checks a term against your rules, that is tokens spent, on top of whatever the media itself costs. Think of it like paying a human negotiator by the hour, except the clock runs for every email they read as well as every one they write.

A good agent does this efficiently. It reads the brief once, finds inventory that fits, makes one clean counter-offer, and closes. Call that three or four exchanges, three or four lots of tokens.

A bad agent does the same job badly. It re-reads information it already has. It gets a counter-offer and cannot decide whether to accept or push back, so it asks a clarifying question instead, which is itself another paid exchange. It goes round the same negotiation loop five or six times because it cannot tell a finished deal from an unfinished one. Every extra loop is tokens billed, and none of them bought you a better outcome. Some of them bought you a worse one, because a long negotiation under time pressure tends to end in whichever terms get agreement fastest, not whichever terms are best.

This is not hypothetical, we measured it

We built Concourse, a benchmark that runs AI buying agents through a simulated market with the same brief and records what happens, the agentic equivalent of a crash test. Eight models, 96 attempts across four versions of one brief, same rules, same inventory, same money on the table each time.

The spread was wide. Some models closed a deal every single attempt. Others failed to close at all on a meaningful share of attempts, running out of their allotted exchanges before reaching terms either side would sign. Every one of those failed attempts still cost tokens. The model read the brief, made proposals, received counters, and burned its budget without producing anything a human could use. That is the purest version of the problem: cost with zero output.

Closing is not the same as closing well, either. Of the deals that did get signed, 95 of 110 signed contracts carried a viewability floor below 70%, a number that should make any buyer wince. Viewability is simply the share of an ad that was actually capable of being seen on screen, not just served. A floor that low means a meaningful share of the inventory in that deal was never positioned to be seen in the first place. So the agents that closed fast were not necessarily the agents that closed well. Several of them got there quickly by giving up ground a human negotiator would have held.

What a bad negotiation actually looks like, step by step

Picture an agent with a £5,000 budget and a brief asking for 70% viewability minimum. A publisher's agent opens at 60% viewability for the full budget. A well-trained agent rejects it immediately, one exchange, and counters with a smaller spend at the required floor. A poorly trained agent instead asks the publisher's agent to "clarify what viewability means in this context," gets an answer, asks a follow-up about measurement methodology, gets another answer, then finally proposes a counter that still only requests 65%. That is four extra exchanges, each one billed, to land on a worse number than the well-trained agent reached in one.

Multiply that pattern across every deal a buyer runs in a month and the token bill for the inefficient agent is not a rounding error. It is a second line item sitting on top of a worse media outcome.

Two numbers, not one

Which is why "which AI model is best at media buying" is the wrong first question. Two numbers matter and they are not the same number: what it costs to run the agent, the token bill, and how good the deal is that the agent produces, everything downstream of that. A cheap agent that needs five attempts to land a mediocre deal is not actually cheap once you add up the exchanges. A slightly pricier agent that gets a strong deal in one pass almost always is, because you are not paying twice, once for the thinking and once for the inventory it got wrong.

What this means if you are the one paying for it

Before you hand an agent a real budget, ask whether it has been tested against a market where it could fail, the way Concourse tests it, not just a demo where it was always going to succeed. Ask whether it has been trained on what a bad deal looks like, which is what the Alkimi Gym does, so it recognises a 60% viewability floor as a term to reject rather than a number to negotiate around in circles. And keep a human in the loop on the mandate, meaning the rules the agent has to work inside, so a bad run costs you tokens and nothing more, not your media budget as well.

The model gets the agent talking. Whether it is worth listening to is a separate question, and it is the one worth asking before the budget is live, not after.