How much does running an AI media-buying agent cost?

Running an AI media-buying agent costs between $0.74 and $14.38 per deal negotiation in API compute cost, based on Concourse v1 operational data : a figure that is separate from the media cost itself and covers only the cost of the agent reasoning through and completing a deal.

Agent operating cost: The compute cost of running the AI model that executes the buying agent's reasoning and actions; separate from media cost, which is what the agent spends on buying inventory.

Per-deal cost: The total API compute cost of negotiating and committing to one deal, from mandate check through to DealSheet commitment; the unit that Concourse v1 data is measured in.

Cost variance: The range between the cheapest and most expensive deals reflects negotiation complexity: straightforward deals with immediate acceptance cost less than multi-round negotiations with parameter adjustments.

Human buying equivalent:The labour cost of a human media buyer completing the same deal tasks : for comparison purposes, a single human-negotiated deal in a large agency team costs significantly more than $14.38 in loaded labour time.

Where the cost figure comes from

Concourse, the agentic advertising platform, published operational data from its v1 agent deployment showing per-deal compute costs ranging from $0.74 to $14.38. This is the cost of the API calls made by the buying agent while completing one deal: reasoning about the inventory brief, generating the deal enquiry, processing the publisher's response, iterating on terms, and committing to the DealSheet.

The range reflects real variation in deal complexity. A deal with a publisher whose terms immediately match the buyer's parameters completes in fewer reasoning steps and costs toward the lower end. A deal that requires multiple rounds of counter-offer and parameter adjustment costs more because the agent must process more input and generate more output.

These figures are from 2025 v1 deployments. As model efficiency improves and agentic workflows are optimised, per-deal costs are expected to fall.

What is and is not included in the operating cost

Agent operating cost covers the model inference cost: the compute required for the AI to read context, reason about options, and generate outputs during the deal negotiation sequence. It includes mandate checks (does this deal fall within my authorised parameters), deal term generation (what specific terms do I want to propose), response processing (what did the publisher's agent offer), and commitment actions (execute the DealSheet commitment).

It does not include the media cost, which is what the buyer pays for the inventory itself. A $1,000 CPM deal negotiated by the agent at a cost of $2.50 in agent compute is still a $1,000 CPM deal; the $2.50 is the overhead cost of the negotiation process.

The operating cost also does not include infrastructure costs: the servers, APIs, and platform subscriptions required to run the agent. These vary by deployment model.

How does this compare to human buying?

A human media buyer completing the equivalent deal tasks : reading a brief, reaching out to a publisher contact, negotiating terms over email or phone, processing a counter-proposal, and confirming the final deal : takes materially more than $14.38 in loaded labour time even at junior buyer rates.

The efficiency case for agentic buying is not primarily that agents are cheaper per deal at current compute costs. It is that agents can run in parallel across dozens of publisher relationships simultaneously, without the coordination overhead, scheduling constraints, and information loss that characterise human deal management. The cost per deal is secondary to the throughput advantage.

At scale, a buyer running 500 deals per month through agents at an average of $4 per deal spends $2,000 in agent operating cost, replacing work that would require several full-time media buyer headcount.

What drives cost up or down?

Deal complexity is the primary cost driver. Deals that require more reasoning steps cost more. The variables are: number of negotiation rounds before acceptance, number of parameters being evaluated (targeting, pricing, volume, safety constraints all add context), and whether the agent needs to escalate to human approval mid-negotiation.

Model choice matters too. More capable models cost more per token but may complete deals in fewer steps; less capable models cost less per token but may require more iterations. Optimising agent cost involves calibrating model capability against deal complexity requirements.

Frequently asked questions

Is agent operating cost charged on top of media cost?

Yes. Agent compute cost is separate from media spend. A buyer pays for the media inventory at the agreed deal price and separately bears the compute cost of running the agent that negotiated the deal.

Do agentic advertising platforms charge for agent usage?

Platform pricing varies. Some platforms include agent compute in their service fee; others pass through API costs directly. Concourse's published figures reflect raw API compute cost before platform markup.

At what volume does agentic buying become cost-effective?

At current compute costs, agentic buying becomes cost-competitive with equivalent human labour at deal volumes above roughly 50-100 deals per month, assuming typical loaded buyer rates. The crossover point varies by deal complexity and buyer market.