Generative AI in advertising produces content: copy, images, video, audio and variations of each. Agentic AI in advertising takes actions with consequences, such as committing budget, agreeing deal terms or changing a live campaign, within a mandate a person has set. The difference is not the model underneath, because many agents use a generative model to reason. The difference is what the system is permitted to do before a human looks at its work, and that decides how it has to be governed.

TL;DR. Generative AI makes things, and a person usually reviews the output before it has any effect in the world. Agentic AI does things, and its actions take effect when they execute, so the controls have to be set in advance through a mandate, spend thresholds and a record of what was agreed. The two combine in a modern campaign: generative tools produce creative variants, and agents buy the media those variants run on. Confusing the two leads to the wrong governance: approval after the fact works for a headline, but not for a signed deal.

What is generative AI in advertising?

Generative AI in advertising is software that creates new content from a prompt or a set of inputs. Typical uses include drafting ad copy, producing image and video variants for different formats and localising creative into other languages.

The defining feature is that the output is an artefact. A headline, an image or a script sits there until someone uses it. The generation step commits no money and agrees nothing with anyone. The main risks concern the content: accuracy, brand fit, rights in source material and claims a regulator would object to.

What is agentic AI in advertising?

Agentic AI in advertising is software that works from a mandate set by a human principal, takes decisions inside that mandate on its own, passes decisions outside it back to a person, and leaves an auditable record. A buying agent might discover inventory, compare seller proposals, negotiate a price, execute a deal and pace spend across a flight. A selling agent might package a publisher's inventory, respond to buyer requests and accept, counter or decline proposals.

The defining feature is consequence. When an agent accepts a deal, the commitment exists at that moment, and the counterparty acts on it.

What is the practical difference between the two?

A simple test separates them, which this article calls the consequence test: does the system's output change anything outside the system before a person has looked at it?

For generative AI the answer is usually no: forty headline variants change nothing until a marketer picks one. For agentic AI the answer is yes by design. An agent with a mandate to spend has authority to act without a person approving every step, which is the reason for using it.

A hypothetical example makes the line clear. A generative model drafts twelve video scripts for an autumn campaign; none of them exists outside the tool until the brand team approves two. An agent working to a £30,000 mandate agrees a £9,000 private marketplace deal with a publisher's selling agent; that commitment is live from the moment both agents accept it. Both systems may run on similar models. Only one of them has spent money.

Why does agentic AI need different governance from generative AI?

Generative AI is governed mainly after the output exists. Review, brand approval, legal sign-off on claims and rights checks all happen between creation and use. That works because the content is inert until someone releases it.

Agentic AI has to be governed mainly before it acts, because there is no gap between decision and effect. Four controls matter, and all are set in advance. The mandate defines what the agent may buy or sell, at what price, against which audiences, in which environments and through which deal types. A spend threshold holds any commitment above a set value until a named person signs it off. A default for the unfamiliar sends a new seller, data source or deal type back to human review. And a record captures what was agreed, by which agent, for which principal, and who approved it.

Either way, the deploying organisation stays accountable. Neither a generative model nor an agent is a legal entity.

How do generative and agentic AI work together in a campaign?

The two are complementary, and most real workflows use both. In a hypothetical campaign, a generative tool produces copy and image variants for each format, and people approve the set. A buying agent then takes the brief and mandate, finds inventory, negotiates deals and distributes the approved creative to the sellers it bought from. Mid-flight, the agent shifts spend between deals within its limits, while the generative tool drafts fresh variants that go back through approval.

The standards work reflects this overlap. AAMP, IAB Tech Lab's Agentic Advertising Management Protocols, focuses on how agents communicate and transact: its components include agentic-direct for agent orchestration, reference buyer and seller agents, and OpenProposal for how publishers describe products and packages in RFP responses. AdCP, the Ad Context Protocol from AgenticAdvertising.org, is a separate protocol that includes creative tasks such as build_creative and sync_creatives alongside buying tasks, so generated assets can move through the same agent workflow as the media buy.

There is also overlap inside the agent itself. Many agents use a large language model to read a brief, interpret a seller's offer or draft a negotiation message. That makes the model generative and the system agentic. The governance follows the system's authority, not the model's architecture.

What goes wrong when the two are confused?

Two failures are common. The first is treating an agent like a content tool: letting it act, then reviewing what it did. By the time a person reads the log, the deals are agreed and the counterparty holds its own version of the terms. If that version differs, the dispute starts from two records rather than one.

The second is labelling a generative tool as an agent. A system that drafts a media plan for a person to execute is not acting on anyone's behalf. Buyers evaluating a vendor's "AI agent" can ask three questions: what is it authorised to commit to, what happens when a decision exceeds that authority, and what record of the agreed terms does the counterparty hold?

Alkimi, a neutral agentic advertising marketplace working within IAB Tech Lab's AAMP framework, addresses that last question directly: buyer and seller agents agree terms on a DealSheet that both parties hold, and people keep approval rights over each deal.

Frequently asked questions

What is the difference between generative AI and agentic AI in advertising?

Generative AI produces content such as copy, images and video, which a person usually reviews before use. Agentic AI takes actions with consequences, such as committing budget or agreeing deal terms, within a mandate set in advance.

Can an AI agent use generative AI?

Yes. Many agents use a large language model to interpret briefs, read seller offers or draft messages. The system is agentic because of what it is authorised to do, not because of the model it runs on.

Is a tool that writes media plans an agent?

Not on its own. If a person reviews the plan and executes it, the tool is generative. It becomes agentic when it is authorised to act on the plan, such as booking inventory, within set limits.

Who is accountable for the output of generative or agentic AI?

The organisation that deployed it. Neither system is a legal entity, so responsibility stays with the advertiser, agency or publisher.

Which standards cover agentic AI in advertising?

IAB Tech Lab's AAMP framework covers agent communication, reference buyer and seller agents, agent discovery and OpenProposal. AdCP, published by AgenticAdvertising.org, is a separate protocol defining tasks across the campaign lifecycle.

Generative AI vs Agentic AI in Advertising: What Is the Difference?