AI agents help with media buying by taking on defined tasks at each stage of a buy, from turning a brief into requests to sellers through negotiating deals, pacing spend and reconciling delivery, while people set the rules and approve what gets committed. The practical way to use one is to hand it a single narrow task, test it against realistic counterparties before it touches real money, and widen its authority one stage at a time as the record supports it.

TL;DR: An AI agent can help at seven points in a media buy, from the plan to reconciliation. Start it on a frequent, well-defined decision that is cheap to get wrong, and keep strategy, the mandate and exceptions with people. Move it through the five stages of earned autonomy (observe, recommend, draft, human-approved action, bounded automatic action) only on evidence, and benchmark what it secured as well as whether it closed.

What can an AI agent do at each step of a media buy?

The table maps the seven steps of a typical buy to what an agent can do, the earned autonomy stage a first deployment should start at, and the decision that stays with a person.

Step	What the agent can do	Where to start it	What stays with a human
1. Brief and plan	Turn a brief into a structured request and shortlist sellers	Recommend	Objectives, audiences and total budget
2. Mandate	Draft testable parameters from the brief	Draft	Approving and owning the mandate
3. Proposal checks	Check every incoming proposal against the mandate	Observe, then Recommend	Which exceptions are worth pursuing
4. Negotiation	Negotiate within price ceilings and required protections	Draft, then Human-approved action	Commitments above approval thresholds
5. Creative	Check assets against specifications and distribute them	Draft	Creative approval and brand context
6. In-flight optimisation	Pace budget, adjust bids and flag anomalies	Bounded automatic action, inside hard caps	Moves outside the caps, and pausing
7. Reconciliation	Compare agreed terms with delivery	Recommend	Disputes, make-goods and final sign-off

What are the five stages of earned autonomy?

Earned autonomy is the principle that an agent's authority grows only as its record shows it can be trusted. As a worked example, Alkimi, an enterprise platform for agentic advertising, describes five stages that a buying agent moves through.

Observe: the agent records what it would have done, with no authority to act.
Recommend: the agent suggests an option, and a person decides.
Draft: the agent prepares the deal or change, and a person executes it.
Human-approved action: the agent executes routine decisions within its mandate, and anything above set thresholds goes to a named person.
Bounded automatic action: the agent acts within its full mandate without case-by-case approval, inside hard limits a person sets and can halt.

What is earned autonomy in advertising? sets out the full model, including promotion and demotion reviews.

Step 1: How can an AI agent help turn a brief into a plan?

An agent can read a campaign brief, turn it into a structured request and send it to many sellers at once. IAB Tech Lab's AAMP framework reached this stage with AAMP 3.0, whose OpenProposal specification "enables the automation of the discovery and planning phase of a campaign". The Ad Context Protocol documentation describes a comparable task, get_products, which "sends the same structured request and brief to every connected seller".

Start the agent at Recommend here: it can propose a seller shortlist and a draft plan, but objectives, audiences and the total budget are decisions a person makes.

Step 2: What should an AI agent's media buying mandate contain?

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 sellers, audiences, brand safety rules and contractual protections such as viewability floors and make-goods. Anything left out of the mandate is something the agent may trade away.

Write it as testable parameters: "premium inventory" cannot be checked, but a list of approved domains, a minimum promised viewability and a maximum price per thousand impressions can. An agent can draft these parameters, but a named person approves the mandate and owns every change to it, as covered in How to write a media buying mandate for an AI agent.

Step 3: How can an AI agent check seller proposals?

Checking incoming proposals against the mandate is the best first job for most buying agents, because it is frequent, rule-based and harmless while a person still decides. The agent reads each proposal, compares price, volume, flight dates, viewability and brand safety terms with the mandate, and flags every mismatch. At Observe the team compares the agent's verdicts with its own; at Recommend the agent's shortlist becomes the trader's starting point.

Identity checks belong in this step too. AAMP includes the IAB Tech Lab Agent Registry, which "establishes a trusted framework for agent identity, verification, and disclosure". A buying agent should confirm who is behind a seller agent before it treats a proposal as real.

Step 4: How do AI agents negotiate and set up deals?

A buying agent can run parallel negotiations with many seller agents and apply the mandate identically in each. AAMP 2.0 added support for the Deals API, so the framework covers "Programmatic Guaranteed, Preferred Deals, PMPs, and the open programmatic marketplace, as well as Linear GRPs", and IAB Tech Lab publishes open source buyer and seller agent reference implementations. On the AdCP side, the documentation says one call to create_media_buy "executes the campaign across every seller".

Start negotiation at Draft, where the agent prepares terms and a trader commits, then move to Human-approved action once the record supports it. Approval thresholds send a deal above a set value, a new seller or any term outside the mandate to a person. How to design approval thresholds for agentic campaigns covers where to set them.

Step 5: Can an AI agent handle creative trafficking?

An agent can check finished assets against each seller's specifications and distribute them, a repetitive task that is easy to verify. In AdCP, once creative is approved, sync_creatives "distributes the finished assets to every seller simultaneously". The approval itself, and the judgement about whether a placement suits the brand, stays with people.

Step 6: How do AI agents optimise a campaign in flight?

In-flight optimisation suits agents because the decisions are frequent, small, measurable and usually reversible. An agent watching delivery, remaining budget and remaining flight time can adjust spend to avoid underpacing and overpacing, shift budget toward the placements producing results, and flag anomalies such as a sudden spike in spend. AdCP's update_media_buy task handles mid-flight changes, including the ability to "shift budget between packages, adjust flight dates, swap creative assignments".

This is usually the first place an agent reaches Bounded automatic action. The bounds must be enforced by the system, not reported by the agent: hard caps on daily spend and on how much budget can move, a stop control a person can use at any time, and an alert whenever a limit is hit.

Step 7: How do AI agents help with reporting and reconciliation?

After launch, an agent can pull delivery from every seller into one view and compare it with what was agreed, line by line. In AdCP, get_media_buy_delivery "aggregates performance from every seller into one response". The harder problem is agreeing what was agreed in the first place.

If each side keeps only its own log, a disagreement becomes one record against another and reconciliation turns into a dispute. The more reliable approach is a single deal record written at agreement, timestamped, held by both parties and retrievable by either side independently. Each mismatch between that record, the mandate and delivery should be traced by a person before the next deal.

What should you hand an AI agent first?

Hand it one decision that is frequent, well defined and cheap to get wrong, and nothing else. Proposal checks (step 3) and pacing within hard caps (step 6) are the usual candidates, because the right answer is checkable and a mistake is small and reversible. A hypothetical first scope: the agent may evaluate and recommend run-of-site display deals from five named publishers.

Anthropic's engineering guide Building effective agents recommends "finding the simplest solution possible, and only increasing complexity when needed". In media buying that means one deal type and a short seller list before anything broader.

What should stay with a human?

Three things stay with people however well an agent performs. The first is strategy: objectives, audiences and brand context are judgement calls the mandate expresses but the agent should not originate.

The second is the mandate and its exceptions. An agent that can widen its own limits has no limits, so every mandate change needs a named owner, and a deal outside the mandate is flagged and stopped while a person decides.

The third is verification. Delivery should be checked against a shared deal record and independent measurement rather than the agent's own report, and the organisation that deployed the agent remains answerable for what it buys. Article 14 of the EU AI Act, written for high-risk AI systems rather than advertising, frames oversight the same way, saying measures should be "commensurate with the risks, level of autonomy and context of use".

How do you test an AI buying agent before it spends?

Testing has four layers, and an agent should pass each before moving to the next: can it buy well, can it hold its ground, do the controls work, and how does it behave with real money.

Benchmark what it secured, not just whether it closed

A useful benchmark scores completion, what was committed and which contractual protections were secured. In Concourse Bench v1, eight AI buyers each negotiated with three sellers controlled 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, instructions or mandate change.

Spar it against counterparty agents

Before live spend, let the buying agent negotiate against seller agents in a practice environment, a process often called sparring. Vary the counterparties (one firm on price, one that concedes quickly, one offering weaker protections at a lower price) and check that the agent walks away when terms breach the mandate. IAB Tech Lab's open source seller agent gives teams a documented baseline counterparty, and Anthropic's guide, which warns of "the potential for compounding errors", recommends "extensive testing in sandboxed environments".

Test the controls, not just the agent

Push the agent against its limits on purpose: a deal just over the price ceiling, an unapproved seller, a pacing scenario that would breach the daily cap. Each should produce a refusal or an escalation, never a commitment, and the stop control should halt the agent mid-negotiation. A control that has never been triggered in testing has not been tested.

Run a small supervised pilot

Run a time-boxed live pilot in which a person approves every commitment, on a budget the team would accept losing as the cost of learning. An early public example followed this pattern: according to AgenticAdvertising.org, which publishes AdCP, the first agent-to-agent media buy was executed on 16 October 2025 with "real money, real inventory from LG Ads, real AI agents handling creative, targeting, and approval with human oversight". Adam Broitman, a partner at McKinsey & Company, is quoted on the same page: "It felt like you were watching the first banner ad being served."

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.

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. Write the promotion criteria before each stage begins (a hypothetical example: 100 consecutive deals with no mandate breaches and no reconciliation mismatches), and have the person with budget authority sign off the review. Demotion should be just as explicit, and even at Bounded automatic action a person sets the bounds and can halt the agent.

Where does an agentic advertising platform fit in?

Alkimi's three products follow this guide's path. Concourse, a live benchmarking product for AI media-buying agents, ran the Bench v1 test cited above, 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.

The Marketplace is neutral: it works within IAB Tech Lab's AAMP framework today, uses AAMP identity checks, is being built to be AdCP compatible, and leaves humans with approval rights at the deal level. Deals there are recorded on a DealSheet, the shared record described in step 7, and the company took part in the OpenProposal v3.0 public comment period, filing comments and change requests.

Frequently asked questions

How can AI agents help with media buying?

AI agents can turn briefs into requests to sellers, check proposals, negotiate within a mandate, pace spend and reconcile delivery, applying the same rules every time. 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.

Which standards cover AI agents that buy media?

IAB Tech Lab's AAMP framework covers agentic buying from the brief, through OpenProposal, to deal execution, with open source buyer and seller agents and an Agent Registry for identity. AgenticAdvertising.org publishes AdCP, an open protocol whose tasks cover discovery, creative, buying, governance and delivery reporting.

How Can AI Agents Help with Media Buying? A Step-by-Step Guide