Last updated: 13 October 2026

An AI agent in advertising is software that takes actions on behalf of an advertiser, agency or publisher within limits a human has set, without a person approving each individual step. What separates an agent from the automation programmatic has run on for fifteen years is not how clever the model is. It is the mandate: a defined scope of authority, a route back to a human when a decision falls outside that scope, and a record of what was decided and why.

TL;DR. AI agents in advertising fall into three roles: buy-side agents acting for advertisers, sell-side agents acting for publishers, and infrastructure agents that check, route and record the transactions between them. The tasks they execute today are operational: finding inventory, bidding, negotiating within set limits, pacing spend and reporting delivery. Setting the mandate, approving spend above a threshold and carrying legal accountability stay with people. Two open frameworks now give agents built by different companies a shared way to find each other and transact: AAMP from IAB Tech Lab, and AdCP, the Ad Context Protocol, from AgenticAdvertising.org.

What is an AI agent in advertising?

An AI agent in advertising is a software system that works from a mandate set by a human principal, makes decisions inside that mandate on its own, hands decisions outside it back to a person, and leaves an auditable record. The mandate defines what the agent may buy or sell, at what price, against which audiences, in which environments, and which commitments need human sign-off before they are final.

The term is stretched in vendor marketing. "Agent" is used for anything from a bid-shading algorithm to a system that negotiates deal terms with another company's software. A simple test separates the two: if a system cannot state what it is authorised to do, cannot pass a decision back to a person, and cannot show a record of what it agreed, it is automation with a new label.

How is an AI agent different from a DSP or programmatic automation?

A demand-side platform reacts. A bid request arrives, the DSP prices it within milliseconds, and the auction clears. An AI buying agent works over a longer horizon: it can start a conversation with a seller, compare offers across several sources, negotiate terms, manage one budget across many transactions and report back, all within its mandate.

Rules-based automation applies the same rule every time. An agent interprets a brief and chooses its actions. That flexibility is the point of using an agent, and it is also why an agent needs tighter governance than a rule ever did.

What are the main types of AI agent in advertising?

The clearest way to classify advertising agents is by whose interests they represent.

Buy-side agents act for an advertiser or agency. They take a brief, discover inventory, submit bids or negotiate deal terms, apply brand safety and frequency rules, pace spend and report on delivery.

Sell-side agents act for a publisher or inventory owner. They package inventory into products, respond to buyer requests, enforce price floors and deal rules, and accept, counter or decline proposals. IAB Tech Lab's AAMP framework publishes separate reference implementations for a buyer agent and a seller agent, which reflects how the industry now divides the two roles.

Infrastructure agents sit between buyer and seller. They check a proposed action against the mandate before it executes, route decisions to a person when a threshold is crossed, and write the record of what was agreed. In AAMP, the agentic-direct component provides the communication and orchestration layer for agent interaction, and a registry example shows how agents are discovered and registered before they transact. AdCP's governance task, check_governance, does a comparable job at the point of execution: according to the AdCP documentation it validates budget limits, brand safety and targeting compliance before a buy runs, and escalates to a human when an action exceeds the agent's authority.

Some taxonomies classify by function instead: planning agents, buying agents and optimisation agents. The two views are compatible. Function describes what an agent does. Role describes who it answers to, and for accountability, role is the one that matters.

What does an AI agent actually do in a media buy?

A buying agent's work across a campaign follows the same stages a trader's does. It takes the brief, finds inventory that fits it, requests or reads proposals from sellers, negotiates terms within the mandate, executes the buy, applies frequency and brand safety rules while the campaign runs, and reports delivery. IAB Tech Lab's OpenProposal specification, part of AAMP, standardises how publishers define advertising products and packages and present them in RFP responses, so a buyer agent can read offers from different sellers in one format. AdCP expresses the same lifecycle as named tasks, from get_products for discovering inventory to create_media_buy for executing a buy and get_media_buy_delivery for reporting. These are execution tasks: high volume, rule-bound work where speed and consistency count.

What an agent does not do matters as much. An agent does not decide whether the campaign objective is right. It cannot judge whether a creative suits the page it lands on unless that rule was written into the mandate. It cannot establish the lawful basis for using personal data: under UK GDPR, the data controller has to establish that before processing begins. And it carries no legal responsibility for what it commits to. The organisation that deployed the agent remains accountable.

Where do humans stay in the loop?

Well-designed agent workflows put the human checkpoints in on purpose rather than discovering them after a failure. The first checkpoint is the mandate itself, which a budget owner approves before the agent spends anything. The second is the spend threshold: a commitment above a set value is held until a person signs it off. The AdCP documentation illustrates this with a buy above a $20,000 auto-approval limit being escalated to a campaign operations manager, who reviews it, adds conditions and approves. The third is anything new: an audience segment, a data source or a deal type the mandate did not anticipate defaults to human review.

Autonomy in this model is earned, not assumed. Alkimi describes the progression as five stages, each earned with evidence from the one before: observe, recommend, draft, human-approved action, and bounded automatic action.

Are AI agents already buying media?

Yes, in limited volume, and the industry infrastructure is now being built in the open. IAB Tech Lab announced AAMP, its Agentic Advertising Management Protocols, in early 2026 as the umbrella for all of its agentic work. The framework is open source and made up of separate components, each versioned independently: agentic-direct for agent communication and orchestration, reference buyer and seller agents, OpenProposal for publisher products and RFP responses, a registry example for agent discovery, the Agentic RTB Framework for agent services running inside a host platform, and agentic-audiences for interoperability between advertising agents.

AgenticAdvertising.org, which publishes AdCP, reports that the first agent-to-agent media buy using its protocol ran on 16 October 2025, with inventory from LG Ads and AI agents handling creative, targeting and approval under human oversight. The organisation lists more than 154 member companies, including Yahoo, PubMatic, Scope3 and Samba TV. It quotes Adam Broitman, a partner at McKinsey & Company, on that first buy: "It felt like you were watching the first banner ad being served."

Adoption is real but early. Most programmatic spend still runs through DSPs and SSPs. For buyers, the open question is no longer whether an agent can execute a buy. It is whether the record of what the agent agreed will hold up when delivery is disputed.

How should buyers evaluate a vendor's "AI agent"?

Three questions cut through most claims. First, what does the mandate interface look like, and who approves it? Second, what happens when a decision exceeds the mandate: does the agent stop, escalate, or carry on? Third, what record of the agreed terms exists, and does the counterparty hold the same record? A vendor that cannot answer the third question is asking the buyer to trust its logs.

Alkimi, an agentic advertising marketplace, is built around that third question: buy-side and sell-side agents agree terms on a shared DealSheet that both parties hold, and humans keep approval rights at the deal level.

Frequently asked questions

What does agentic advertising mean?

Agentic advertising is advertising in which AI agents carry out buying, selling, planning or optimisation tasks within mandates set by people, without a person approving every individual transaction.

Is an AI buying agent the same as a DSP?

No. A DSP responds to auction bid requests. An AI buying agent can start negotiations, compare offers across sellers, agree deal terms and manage a budget across many transactions, all within a mandate.

What protocols do advertising agents use to transact?

The two main frameworks are AAMP, IAB Tech Lab's Agentic Advertising Management Protocols, and AdCP, the Ad Context Protocol published by AgenticAdvertising.org. AAMP covers agent communication and orchestration, reference buyer and seller agents, agent discovery and OpenProposal. AdCP defines tasks across the campaign lifecycle and runs over MCP or A2A.

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

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

What can an AI media-buying agent not do?

It cannot set its own mandate, establish the lawful basis for using personal data, judge creative suitability it was not told about, or approve spend above its threshold without a person signing off.

What Are AI Agents in Advertising?