AI agents help with media buying by taking on a defined set of trading decisions, such as reading a brief, negotiating terms with sellers and assembling a portfolio within budget, while people set the rules and approve what gets committed. The practical way to use them is to start with one narrow decision, test the agent before it touches real money, and widen its authority only when the evidence supports it.

TL;DR. Using an AI agent for media buying is a staged process, not a switch. Scope one decision, write the mandate, set approval thresholds, benchmark the agent, let it practise against counterparty agents, run a small supervised pilot, reconcile the deal record, and only then move it up the earned-autonomy stages. A person approves commitments at every stage until the record shows the agent can be trusted with more.

1. Which media buying decision should an AI agent own first?

Pick one decision that is frequent, well defined and cheap to get wrong, and give the agent only that. Good candidates are checking incoming proposals against a list of requirements, or negotiating a single deal type with a short list of known sellers. Strategy, audience choice and creative judgement stay with people.

Anthropic's engineering guide Building effective agents recommends "finding the simplest solution possible, and only increasing complexity when needed". A hypothetical first scope might read: the agent may negotiate run-of-site display deals with five named publishers, and nothing else.

2. What goes into an AI agent's media buying mandate?

The mandate is the written set of limits the agent works within, and it should be specific enough that software can test every proposed action against it. At minimum it covers budget, price ceilings, permitted inventory and sellers, audiences, brand safety rules, and the contractual protections the buyer expects, such as viewability floors and make-goods. Anything left out of the mandate is something the agent may trade away.

Write the mandate as testable parameters rather than intentions. "Premium inventory" cannot be checked, but a list of approved domains, a minimum promised viewability and a maximum price per thousand impressions can. A fuller template is in How to write a media buying mandate for an AI agent.

3. Where should human approval thresholds sit?

Approval thresholds are the points at which the agent must stop and ask a named person before committing. Set them on value, novelty and deviation: a deal above a set amount, a seller the agent has not traded with before, or any term outside the mandate should go to a human. In the early stages, the threshold should be every commitment.

A hypothetical starting rule set might send every deal to a trader for sign-off, any deal over £10,000 to a second approver, and any brand safety exception back to the campaign lead. Each approval should be logged so it becomes part of the record. How to design approval thresholds for agentic campaigns covers the design in more depth.

4. How do you benchmark an AI media buying agent before trusting it?

Benchmark the agent on realistic briefs, and score more than whether it finished. A useful benchmark records completion, what was committed, which contractual protections were secured and how much each run cost.

Concourse Bench v1, a controlled benchmark published by Alkimi, shows why the second layer matters. Eight AI buyers each negotiated with three sellers run by fixed software, and 55 of 96 registered attempts ended in a completed buy. Yet 95 of the 110 contracts signed in those completed buys promised a viewability floor below 70%, so an agent can close deals and still leave the buyer weakly protected.

Test the agent together with its harness, the software that supplies the brief, checks each action and records the deals, because the two behave as one system. Re-run the benchmark whenever the model, the instructions or the mandate change.

5. How should an AI buying agent practise against counterparty agents?

Before any live spend, let the buying agent negotiate against seller agents in a practice environment, a process often called sparring. Sparring shows how the agent reacts to a firm counter-offer, whether it walks away when terms breach the mandate, and whether it tracks budget across several negotiations at once.

Anthropic's guide makes the general case: "The autonomous nature of agents means higher costs, and the potential for compounding errors. We recommend extensive testing in sandboxed environments". For advertising specifically, IAB Tech Lab's AAMP framework publishes reference implementations of both a buyer agent and a seller agent, which gives teams a documented baseline for how each side of an agentic deal is expected to behave.

Vary the counterparties: a seller that holds firm on price, one that concedes quickly, and one that offers weaker protections at a lower price.

6. How do you run a first live pilot with an AI buying agent?

Run a small, time-boxed live pilot in which the agent negotiates and a person approves every commitment before it becomes final. Keep the budget to an amount the team would accept losing as the cost of learning, and work with sellers who know an agent is on the other side.

An early public example followed this pattern. According to AgenticAdvertising.org, which publishes the Ad Context Protocol (AdCP), the first agent-to-agent media buy was executed on 16 October 2025 with real money, inventory from LG Ads, and AI agents handling creative, targeting and approval under human oversight. The organisation calls its governance principle Embedded Human Judgment: humans stay in the loop on decisions with real consequences.

During the pilot, log every proposal, counter-offer, escalation and approval. The aim is not volume but a record of how the agent behaves when the money is real.

7. How do you check and reconcile an agent-negotiated deal?

After each deal, compare what the buying agent says it agreed with what the seller says was agreed, and with what was delivered. If each side keeps only its own log, a disagreement becomes one party's record against the other's, and reconciliation turns into a dispute.

The more reliable approach is a single deal record written at the moment of agreement, timestamped, held by both parties and retrievable by either side without asking the other. Check that record against the mandate line by line: price, volume, flight dates, viewability floor, make-good terms and brand safety conditions. Each mismatch is a finding about the agent, the harness or the counterparty, and it should be traced before the next deal.

8. When should an AI buying agent get more autonomy?

Move the agent up one stage at a time, and only when the record from its current stage supports the move. The earned-autonomy model sets out five stages: observe, recommend, draft, human-approved action, and bounded automatic action.

Write the promotion criteria before the agent starts each stage, for example a hypothetical requirement of 100 consecutive deals with no mandate breaches and no reconciliation mismatches. Demotion should be just as explicit, so a breach sends the agent back a stage. What is earned autonomy in advertising? explains the model in detail.

Regulation frames oversight the same way. Article 14 of the EU AI Act, written for high-risk AI systems rather than advertising, says oversight measures should be "commensurate with the risks, level of autonomy and context of use". Even at the final stage, a person sets the bounds and can halt the agent.

What should AI agents not do yet in media buying?

Some decisions should stay with people however well an agent performs.

Set strategy. Objectives, audiences and the creative context that suits the brand are judgement calls that the mandate expresses but the agent should not originate.

Write or change its own mandate. An agent that can widen its own limits has no limits, so every mandate change needs a named human owner.

Approve its own exceptions. When a deal falls outside the mandate, the agent should flag it and stop, and a person decides whether to approve, decline or amend the mandate.

Mark its own homework. Delivery and outcomes should be checked against the shared deal record and independent measurement, not against the agent's own report.

How does an agentic advertising platform support these steps?

Alkimi is an enterprise platform for agentic advertising whose three products follow the same path as this guide. Concourse is a live benchmarking product for AI media-buying agents (step 4), and the Gym, available now, lets buyers and sellers build their own agents, set the rules they work to and spar against counterparty agents before any real money moves (step 5).

The Marketplace is neutral: it works within IAB Tech Lab's AAMP framework today, uses AAMP identity checks, and is being built to be AdCP compatible. Deals there are recorded on a DealSheet, which is written at agreement, timestamped, held by both parties and retrievable by either side independently, and humans keep approval rights at the deal level. The company is also taking part in the OpenProposal v3.0 public comment period.

Frequently asked questions

How can AI agents help with media buying?

AI agents can check proposals, negotiate terms with sellers and assemble a portfolio within budget, applying the mandate consistently across many negotiations at once. People still set strategy, write the mandate and approve commitments above agreed thresholds.

Do you need to build your own AI buying agent?

Not necessarily, but whoever supplies the agent, the buyer still owns the mandate and the approval thresholds. Any agent, built in-house or bought in, should be benchmarked and tested against counterparties before it trades.

Who is accountable when an AI agent makes a bad media buy?

The organisation that deployed the agent is accountable. An agent is not a legal entity, so responsibility stays with its principal, which is why mandates, approval thresholds and deal records matter.

Can an AI agent buy media without any human approval?

It should not until it has reached the bounded automatic action stage on evidence. Even then, a person sets the bounds and can stop the agent at any time.

How to Use AI Agents for Media Buying: A Step-by-Step Guide