Agentic AI in advertising: what really changes for publishers

Generative AI answers questions. Agentic AI goes further: it breaks a goal into steps, looks up sources, calls tools, coordinates with other agents, asks for approval and then acts. In digital advertising this touches everything at once: buying and selling, ad operations, yield, data, reporting, creative and measurement.

This article opens a series based on our 2026–2029 reference dossier on agentic AI for publishers. It sets out what is changing, what is not, and where we think a publisher should invest first.

No longer a thought experiment

In 2026 the subject has moved from slides to specifications. Two families of standards now exist side by side:

• AdCP, the Ad Context Protocol, launched in October 2025 and governed by AgenticAdvertising.org. It defines tasks such as product discovery and media buy creation that buyer and seller agents exchange over MCP or A2A, and Prebid stewards an open-source sales agent for publishers.
• AAMP, IAB Tech Lab’s Agentic Advertising Management Protocols, formally named in February 2026. It brings together buyer and seller agent SDKs built on existing IAB Tech Lab standards such as AdCOM and OpenDirect, an Agent Registry, and new specifications. On 22 September 2026, AAMP 3.0 added a draft called OpenProposal to standardise the step between a request for proposal and a buying commitment, open for comments until 22 October 2026.

Large platforms are opening up too: Amazon Ads put an MCP server in open beta in February 2026, so that AI agents can call its advertising API in natural language.

Adoption is a different matter. In September 2026, Adform’s CTO said that very little money runs through either protocol so far, and that the two overlap heavily. The standards exist; the volumes do not yet. For a publisher, that is a good moment to prepare without betting on a single protocol.

Two transformations, not one

A publisher has two separate questions to answer:

• Internal: how can agents observe, diagnose, recommend and then act on advertising operations? The target is autonomous operations under human control.
• External: how does the publisher’s offer become discoverable and buyable by the agents of advertisers and agencies? The target is a sales agent, a machine-readable catalogue, an advertising API and a gateway that controls who may do what.

Both rely on the same foundations, which is why they should be planned together.

A maturity ladder

“Agent” covers very different levels of autonomy. We use a simple scale to say what an agent is allowed to do:

Level	Capability	Publisher example
L0 Information	Search and summarise	Answer a question about an ad-ops procedure.
L1 Analysis	Interpret data	Explain a drop in fill rate.
L2 Recommendation	Propose an action	Recommend a floor price adjustment.
L3 Approved action	Prepare, then execute after validation	Create a deal once sales has agreed.
L4 Controlled autonomy	Act within a defined policy	Fix a low-risk incident automatically.

Most useful work today sits between L1 and L3. Moving an agent up a level should be a decision based on measured results, not on a good demo.

Don’t put a chatbot in front of the ad server

The tempting shortcut is to connect a language model to Google Ad Manager and let it answer questions or change settings. We think that is the wrong investment. A publisher should instead build a stable business layer that any model and any agent can use:

• clean, documented data, with shared metric definitions;
• a product catalogue with stable identifiers;
• an advertising API that hides the internal systems;
• a policy engine that decides what each agent may do;
• identity and audit for every caller and every action.

Models will change every few months. These assets will stay, and they are what makes a publisher’s offer understandable to both humans and machines.

Six principles

• API-first and model-agnostic. Do not tie the strategy to one language model.
• Read before write. Start with observation, analysis, recommendations and shadow mode.
• Policy before autonomy. No critical action without deterministic rules, permissions, limits and an audit trail.
• Data provenance. Every important recommendation states its sources, period, freshness and confidence.
• Human agency. People keep the strategic, legal, relational and sensitive financial decisions.
• Discoverability as a product. Being visible to agents is a capability to design, test and measure.

Hybrid, not disintermediated

Agent-to-agent buying does not mean that DSPs, SSPs or agencies disappear. Over the next three years, the more likely path is hybrid: agents orchestrate existing systems, use existing standards, and open new direct routes for some transactions. The open questions are economic:

• Who owns the advertiser relationship when the brief is written in a third-party assistant?
• Who controls how offers are compared, and on which criteria?
• How does a publisher protect its pricing, margins, negotiated terms and reserved inventory?
• Which intermediaries remain essential for identity, verification, measurement, clearing or reach?
• How do you stop an external agent from reverse-engineering your pricing rules with thousands of forecast requests?

Where to start

The next articles in this series go through the market figures, discoverability, the target architecture, the data foundation and a first operational agent. On our side, we test these ideas end to end on two demo publisher sites before proposing them to clients: see how.

Sources

• IAB Tech Lab gives industry until October 22 to weigh in on AAMP 3.0, PPC Land.
• AAMP and AdCP overlap on 13 functions, IAB Tech Lab COO says, PPC Land, August 2026.
• Neither of ad tech’s two AI agent protocols runs at scale, Adform CTO says, PPC Land, September 2026.
• Amazon Ads MCP Server open beta, Amazon Ads, February 2026.