9 Nov 2026 · 4 min read

AI Agents in Advertising: Live, Piloting, or Just Talking?

When a vendor says its advertising platform uses AI agents, that sentence can mean three different things, and buyers rarely learn which one they are looking at. An AI agent, here, is software that takes a multi-step action, such as adjusting a bid or finalising a deal, with limited human input at each step, rather than a human approving every move or a fixed rule firing on one trigger. Some platforms have agents like this live today, spending budget inside limits a human set in advance. Others have the same capability running in closed, invitation-only pilots. A third group has announced a direction or demoed a keynote, with no agent a buyer can point to in production. Sorting a claim into the right bucket matters: it determines who is answerable when the agent gets something wrong.

What separates an agent from automation with new branding?

Automated bidding has existed for over a decade: a human sets a target cost per outcome, and the system adjusts one bid variable within that target using a model trained on past performance. That is automation, not an agent, because the system has a single lever and no choice over which action to take next. An agent, by contrast, chooses among several possible actions and carries one out, such as confirming a line in a deal, without a human approving that action at the moment it happens.

Bucket one: agents live and transacting under oversight

This bucket covers agents carrying out real actions in production, inside boundaries a human set beforehand and can still see into. Amazon Ads Agent, the unified AI-driven platform Amazon moved to general availability through 2026, is built so the agent handles bidding and placement while the advertiser keeps top-line budget and brand decisions, a split Amazon Ads vice-president Kelly MacLean described to AdExchanger in October 2026 as making programmatic simple without taking control from marketers. Oversight sits at the boundary, not inside every bid.

Bucket two: agents running in limited pilot

Here the vendor itself has labelled the capability as not yet generally available. The Trade Desk's Koa Agents, introduced in August 2026, carry the company's own disclosure that several new product highlights mentioned are currently in closed beta and initial capabilities may vary. OpenAI is similarly direct, describing its May 2026 expansion of ChatGPT advertising as the next step in its ChatGPT ads pilot. A pilot is real, but not accessible to every buyer, and its path to general availability is usually unwritten.

Bucket three: roadmap talk, with no live agent yet

This bucket is the loudest and largest by volume of coverage: protocol announcements, future-tense keynote language, and position pieces describing where agentic advertising is heading, without a shipped agent a named customer is using to transact. Google's Marketing Live event in May 2026 introduced a broad set of agent concepts across its advertising products, and the coverage that followed spent as much time asking how marketers would verify those agents as discussing the capabilities themselves. A buyer asking to see this running against their account today deserves a straight answer, and sometimes that answer is not yet.

Why the bucket matters more than the word agent

Meta updated its commercial terms, effective 20 November 2026, placing responsibility for outcomes from its AI-driven ad tools more squarely on the advertiser, a change covered by Inc. and MediaPost in October 2026. Whatever view a reader takes of that change, it shows the point plainly: once an agent is live and transacting, who answers for its mistakes becomes a contract term with real money attached. A pilot carries lower stakes because fewer advertisers are exposed. Roadmap talk carries none at all. The word agent tells a buyer almost nothing about the risk they are taking on. The bucket tells them everything.

Three questions that sort any vendor's claim correctly

  • Is there a named, shipped agent a current customer can point to, running against their own account today, or only a roadmap reference?

  • At what point does a human check the agent's action: before every transaction, within a boundary agreed in advance, or not at all?

  • Who is named, in the contract, as answerable if the agent's action produces a measurable loss?

Two current efforts address parts of that list. The IAB Tech Lab's AAMP framework, Agentic Advertising Management Protocols, announced in early 2026, addresses how an agent's identity is checked before a transaction goes ahead. AdCP, the Ad Context Protocol governed by AgenticAdvertising.org, now at version 3.2 as of October 2026, addresses how agents exchange product and proposal information. Neither protocol settles who holds approval rights once an agent has acted, a commercial question each platform answers on its own terms. Alkimi's marketplace answers it with a DealSheet: a record written at the point of deal, timestamped and held independently by both agents, with a human given approval rights at deal level rather than every bid. That is one answer, not a substitute for asking the question of every platform under evaluation.

All articles

Speak to the team

Request a demo

You don't need to plug in your agent, the demo walks you through the steps.