Agentic AI in Advertising: What Agents Do and What You Control

In September 2026 Amazon renamed its combined buying platform and ads console Amazon Ads Agent. In its Full-Funnel Campaigns the advertiser supplies a budget, products and creative, approves the creative, and “Amazon’s AI handles the rest” across sponsored ads, display, video and streaming TV. A week later, at Advertising Week in New York, Meta said it is testing agents that create campaigns, generate creative and change targeting and budgets. Its Advantage+ products, which already automate much of a campaign, reached an annual revenue run rate of more than $75 billion in the second quarter. Google’s Ads Advisor diagnoses accounts and suggests changes it applies “with your review and approval,” and The Trade Desk added agents for campaign setup, audiences and troubleshooting to its Kokai platform in August.

The big agency groups followed in January 2026. WPP launched Agent Hub, whose first agents handle strategy and insight rather than buying, and Omnicom unveiled a new version of its Omni platform with “autonomous agent systems.” The same week the ad tech company PubMatic launched AgenticOS for agent-to-agent trading. Since then Omnicom has run agent-to-agent media buys for several clients, and in July its chief executive, John Wren, spoke of “the agentic environment, which is nascent,” calling it “something that will be part of the future.” Gartner, as reported by Adweek, says highly autonomous deployments were under 5% of marketing use cases in 2025, and predicts that by 2028 only 15% of agentic AI deployments will be highly autonomous.

So agents are arriving inside the tools you already buy from, mostly with a person still signing off. The useful questions are what an agent is allowed to do in your account, what you still decide, and what to check before it spends your money.

What makes an advertising agent different from automation

Ad platforms have automated single decisions for years. A bidding algorithm sets the price for each auction, and broad match decides which searches your keyword covers. An agent goes further: it takes a goal, looks things up, takes several actions in a row, checks what happened, and acts again. It builds a campaign, sets the budget, picks audiences, writes the ads, notices under-delivery and fixes it. It does this through the same tools a person would use, often connected through the Model Context Protocol (MCP), an open standard that lets AI models call other software.

The difference that matters to you is the right to act. Every agent sits somewhere on a line:

Most of what is sold as agentic sits in the second and third positions. Know which one you are switching on before you switch it on.

What agents do inside the big ad platforms

At Google, Ads Advisor works inside Google Ads, and Ask Advisor, launched in beta in Google Ads in May 2026, is due to work across Google’s products later this year. In Google’s example, it “will automatically pull product details from Merchant Center to set up your new campaign.” AI Max for Search, which began rolling out in beta in May 2025, already acts without asking each time: it widens which searches your ads match, writes headlines and descriptions, and picks landing pages, while you keep brand, URL and location controls.

Meta’s direction was reported in June 2025, when the Wall Street Journal said the company aimed to fully automate ad creation and targeting by the end of 2026. Its agent features are still in testing. Work its agents stage through its ads MCP server lands as a draft in Ads Manager, and every action is gated behind advertiser-set permissions and explicit approval before anything goes live, Relevant Audience reported from Meta’s announcements.

Amazon has gone furthest on autonomy inside one campaign type. In Full-Funnel Campaigns buyers can’t pick individual channels or audiences, Digiday reported. The Trade Desk’s agents set up campaigns, build audiences and troubleshoot, and it uses AI to optimize frequency settings, with a preview of bulk changes before they apply. Yahoo DSP’s agents mostly propose and wait for approval, as our look at Yahoo DSP’s AI agents explains. The same pattern is spreading beyond media buying, into workplace tools such as Agentforce in Slack.

The protocols that let agents buy from each other

An agent buying across many sellers needs a common language. The Ad Context Protocol (AdCP), launched in October 2025 by Yahoo, PubMatic, Scope3, Optable, Swivel and Triton Digital with more than 20 companies signed on, is built on MCP. It gives a buyer’s agent a common way to discover audiences, negotiate media buys and run campaigns with sellers’ agents across platforms. Our explainer on AdCP covers how it works.

IAB Tech Lab runs a parallel track under the name AAMP. Its September 2026 release, AAMP 3.0, aims to “harden deterministic guardrails governing pricing, ad packages, orders and media buy details,” standardize how agents prove who they are and what they are authorized to do, and stop a retried request from creating a duplicate order. Those are the failures you would worry about if an agent were spending your money. Anthony Katsur, IAB Tech Lab’s chief executive, said in a January statement, as Marketing Dive reported, that practical use would require “years of market experimentation, standardization and alignment” and that the industry should expect “several false starts.”

Shopping protocols are a separate layer. Google’s Universal Commerce Protocol and OpenAI’s Agentic Commerce Protocol govern how AI assistants find products and check out, not how ads are bought. They matter to advertisers because they decide what an assistant knows about your products, which our guide to the product data AI shopping assistants read covers. They also squeeze the ad tech companies in the middle, as Criteo’s results show.

What still needs a person

Three things stay with you, whichever platform you use.

The objective. An agent optimizes what it is told to optimize, using the signal it is given. If your conversion tracking counts form fills that never become customers, an agent will find more people who fill in forms, faster than a person would. Fix the signal before you hand over the controls.

The limits. Budget ceilings, inventory you will and won’t buy, audiences to exclude, and creative that needs your approval are decisions about risk, and they belong to the person who carries it. Agents still get facts wrong. “Agentic AI still hallucinates,” Greg Langer of Havas Media Network said at a Digiday summit in May 2026, where agency executives described agents quoting the wrong CPMs.

The measurement. An agent reports on its own success, and the platform that sells the agent also reports the results. A holdout test, in which a random group sees no ads, is still the most direct way to know what the spend caused. Agencies are adding their own checks: Digiday reported agencies logging what their agents decide and, at PMG, capping how many AI tokens each user can spend in a day, because running agents can get expensive quickly.

Most companies haven’t written the rules yet. In an IAB Europe study of 50 respondents, 78% had a formal owner for AI governance, but only 48% had AI guidelines written for advertising and marketing.

What to ask before an agent spends your money

Start where a mistake is cheap: diagnostics, reporting and setup checks with approval on. Move to agents acting inside limits once you can measure the result. That order protects the account while the platforms work out their transparency, and it is the same discipline performance marketing applies to any new buying tool. If your product feeds and store will feed the agents, ecommerce marketing work comes first, and an online presence analysis shows what you are running and where before you hand any of it over.

What changes for media teams and agencies

The work of operating interfaces shrinks: trafficking, pacing checks, bulk edits and first-pass troubleshooting. The work of deciding and checking grows. Someone has to define the goal, set the limits, read the logs and run the tests, and that person needs to understand media well enough to spot an agent doing something plausible and wrong.

Agencies are building their own agents and writing rules for them. WPP Media formed an agentic standards group with Disney, Fox, NBCUniversal, Netflix and Paramount in June 2026 and is developing a buyer agent for video, while Omnicom has said its agent-to-agent tests aim at shortening the media supply chain. Gartner predicts that by 2028, 60% of brands will use agentic AI for one-to-one interactions across marketing, sales and support. Whether that happens or not, the buyer who knows what to ask the agent keeps control of the budget.

Frequently Asked Questions

What is agentic AI in advertising?

It is AI software that carries out advertising tasks in sequence rather than making one automated decision. An agent can take a goal, build a campaign, set budgets and audiences, write ads, spot delivery problems and fix them, using the same tools a person would. How much it does alone depends on the platform: some only recommend, some act after your approval, and some act within limits you set.

Which ad platforms use AI agents?

Google offers Ads Advisor and Ask Advisor, which apply changes with your approval, and AI Max for Search, which expands matching and writes ads on its own. Amazon renamed its buying platform Amazon Ads Agent in September 2026. Meta is testing agent actions behind advertiser approval. The Trade Desk, Yahoo DSP and PubMatic have added agents for setup, audiences, troubleshooting and agent-to-agent buying.

Will AI agents replace media buyers?

They are replacing much of the button-pushing: trafficking, pacing checks, bulk edits and routine troubleshooting. They are not replacing the decisions about goals, limits and measurement, and they still make mistakes, such as quoting wrong prices. The media buyer’s job moves toward setting objectives and guardrails, reading what the agent did, and testing whether it improved the results in your account.

What is the Ad Context Protocol?

The Ad Context Protocol, or AdCP, is an open standard launched in October 2025 by Yahoo, PubMatic, Scope3 and others. Built on the Model Context Protocol, it lets a buyer’s AI agent discover audiences, negotiate media buys and run campaigns with sellers’ agents across different platforms, rather than through each platform’s own interface. IAB Tech Lab runs a separate set of agentic standards called AAMP.

How do I control an AI agent that buys ads?

Set limits before it runs: spend ceilings, approved inventory, audience exclusions and creative approval. Keep approval on for budget, bid and targeting changes until the agent has proved itself, check that its actions are logged, and fix your conversion tracking so it optimizes toward real sales. Then judge it with a holdout test rather than relying on its own reports.