An AI marketing agent is a system that holds delegated authority over a bounded marketing outcome: it decides its own next step inside that authority, and it stops at the edge of it. That is the whole definition. Everything contested sits in two words: delegated and bounded.
The phrase is currently being defined by the companies selling it, which is a poor way to arrive at a definition. This piece argues the word already had one.
Disclosure: this publication is produced with support from Docket, which sells software in this category. No product is named or recommended below, and the tests further down are deliberately ones a buyer can run against any vendor, including ours.
First, the two meanings you probably did not want
Search for a marketing agent and most of what you find is not about software. Two older meanings hold the ground.
The occupational one. A marketing agent is a job, an established occupational title with career guides on the major job boards, describing a person who promotes and sells on behalf of a business.
The classical one. In marketing's own reference vocabulary, an agent is an intermediary. The Universal Marketing Dictionary, published by the Marketing Accountability Standards Board with definitions drawn from the American Marketing Association's dictionary, gives it three ways: in sales, a person who acts as a representative of a firm or individual; in retailing, a business unit that negotiates purchases, sales, or both but does not take title to the goods in which it deals; in global marketing, a company or individual that represents a company in a particular market.
Read that retailing definition again, because it is doing more work than any vendor page on this subject. An agent transacts on your behalf. An agent does not own the goods. Authority is lent, not transferred.
The word was already precise
Nothing about the AI usage breaks that inheritance. An agent has always been the party you authorise to act in your place, within limits you set, while you remain the one who answers for the outcome. A commission agent could negotiate a sale and could not rewrite your prices. The authority was real and it was bounded.
So the interesting question about an AI marketing agent was never whether it uses a language model. It is what authority has been delegated, and what happens when the work reaches the edge of it.
The technical half of the definition
The engineering literature draws a line in the same place. Anthropic's guidance for building agents separates two architectures: workflows, where models and tools are orchestrated through predefined code paths, and agents, where the model directs its own process and tool use, retaining control over how it accomplishes the task.
That distinction is about who owns the control flow. In a workflow, you wrote the branches. In an agent, you wrote the goal and the guardrails, and the system chooses the route.
Put the two halves together and the definition holds up without any marketing in it: an agent is a system with delegated authority over an outcome, choosing its own steps inside that authority. The marketing part only names the outcome.
Four tests before you call it one
This is our framework, not anyone's finding. If a system cannot pass these four, the word is being borrowed rather than earned, and each is a question you can put to a vendor in a first call.
The rulebook. Is there a written, versioned statement of what it may do, what it may never do, and what evidence its output requires? Delegated authority that exists only in someone's prompt history is not delegated authority. It is a habit.
The approval gate. Is it precisely defined what a human signs off on, and what invalidates that sign-off? Approving a campaign means nothing. Approving an exact version means something.
The exception queue. Where does the system's uncertainty go? An agent that never escalates is not confident. It is unsupervised, and the boundary of its authority is decorative.
The kill switch. Who can stop it, and what conditions stop it automatically? A named person and written tripwires, not a thread where someone eventually notices.
Those four are the same artifacts that decide who actually owns an agent once it is running. That is not a coincidence. Ownership and definition are the same question asked at different times: the rulebook and the kill switch are how authority gets bounded in the first place, and how you can tell later whose authority it was.
What it is not
Not a chatbot with a longer prompt. A chat interface is a way of talking to a system. It says nothing about what the system may do without asking.
Not an assistant. An assistant produces an artifact and hands it to you. An agent carries an outcome from trigger to measurement. The tell is what happens when nobody opens the tab: an assistant waits, and an agent proceeds or escalates.
Not a workflow with a model dropped into a step. If every branch was written in advance, the model is a component and the workflow is the product. Useful. Just not an agent.
The boundary is genuinely blurry, and the blur favours workflows
Here is the complication, and it cuts against the fashionable answer. Most marketing systems worth building are workflows, and that is usually the correct engineering decision. Predefined paths are cheaper to reason about, easier to test, and fail in ways you can predict. Anthropic's own guidance treats agents as the option you reach for when the task genuinely requires open-ended decisions, not as the default.
Real systems also sit on a spectrum rather than in one camp. A pipeline can be nine deterministic steps with one genuinely delegated decision in the middle. Calling that an agent oversells it; calling it a workflow hides the one place where authority was actually handed over. The honest description names the delegated decision and the bound around it.
Which is why the useful question in a vendor conversation is never whether the thing is an agent. It is: which decision here is yours, which is mine, and what happens at the seam.
Why the definition is worth defending
Because the word is being applied to almost everything, and the deployment data suggests the reality is thinner. In Brinker and Riemersma's Martech for 2026, 90.3% of surveyed participants reported using AI agents somewhere in their stack, but only 23.3% had them in full production, integrated into core processes or customer-facing experiences. Most of what carries the label is an experiment wearing production vocabulary.
Definitions matter commercially here, not just semantically. If a vendor calls a workflow an agent, you will buy a workflow's capability and take an agent's risk: you will believe something is being handled that is in fact waiting for you. The four tests are cheap insurance against that trade.
FAQ
What is an AI marketing agent? A system with delegated authority over a bounded marketing outcome, which chooses its own steps within that authority and stops at its edge. In practice it needs a versioned rulebook, a defined approval gate, an exception queue, and a named kill switch.
What is the difference between an AI agent and an AI assistant? An assistant produces an artifact when asked. An agent carries an outcome and decides its own next step within granted limits. The test is what happens with nobody watching.
Is a marketing agent the same as a marketing agency? No, and the confusion is old. Classically an agent represents a principal without taking title to the goods; an agency is a firm you hire. In software, marketing agent now means a system holding delegated authority.
Is a chatbot a marketing agent? Only if it holds authority to act rather than only to answer. Most are conversational interfaces over retrieval, which is a legitimate thing to be and a different thing to be.
Do we need one? Only where the work genuinely requires open-ended decisions and you are willing to write down the authority you are delegating. Where the path is knowable in advance, a workflow is the better build and the cheaper failure.
How do we tell whether a vendor's agent is one? Ask for the rulebook, the approval gate, the exception queue and the kill switch. If the answers are a demo and an adjective, you are being sold a workflow with better branding.