What we mean by agentic marketing, and what we don't
AI-assisted research. This note was prepared with the assistance of AI tools. It may contain errors and is not professional advice. See our Legal Disclosures.
Language models have moved from writing drafts to taking actions, and marketing teams are starting to hand them real work. The word used for this shift, "agentic," now gets attached to chatbots, scheduled automations, and almost anything with a model inside.
A word that covers everything stops being useful. When the term is vague, teams buy the wrong tools, set the wrong expectations, and judge results against the wrong standard. This post sets out the definition we use in our research and the line we draw around it.
A working definition
A marketing agent is software that pursues a goal by deciding which actions to take, taking them through tools, and adjusting based on what happens. Three parts of that definition carry the weight.
- A goal, not a script: the agent is told what outcome to achieve, such as "produce a briefing on these 20 accounts," rather than every step to follow.
- Tool use: the agent can act. It can search, read a CRM record, draft into a document, or query analytics.
- A feedback loop: the agent observes the results of its actions and decides what to do next, including when to stop or ask for help.
"Agentic marketing" then means marketing work in which some of the planning and execution is done by systems like this, under human direction. The human direction is part of the definition. We do not use the term for systems that run without anyone accountable for them.
Our post on the anatomy of a marketing agent breaks these parts down further, adding memory and guardrails.
What it is not
Several common systems fall outside the definition. The distinction is useful in practice, because each one carries different risks and needs different controls.
- Prompt templates: generating copy from a fixed prompt is useful, but nothing is being decided.
- Rules-based workflows: "if a lead fills out this form, send this email" is automation. It is valuable and often the right answer, but it does not reason.
- Chat interfaces: a chatbot that answers questions but cannot take actions or pursue a goal across steps is an interface.
None of these are worse than agents. They are different tools. In many workflows, a simple automation beats an agent because it is cheaper, faster, and fully predictable.
A system can also contain more than one of these. A chat interface might sit in front of an agent, or an agent might call a rules-based workflow as one of its tools. The label belongs to the part that makes decisions.
A worked example
Consider a hypothetical B2B software team that wants to prepare sales reps for discovery calls. It could build the same feature three ways.
The first version is a template. A rep pastes in an account name, and a fixed prompt produces a one-page summary from whatever the model already knows. Nothing is looked up, and the output has the same structure every time.
The second version is an automation. When a meeting is booked, a workflow pulls the account's CRM record and recent web visits, fills a template, and emails it to the rep. It is reliable and cheap, and it does exactly the same steps for every account.
The third version is an agent. It is told to prepare the rep for the call. It reads the CRM record, notices the account had a support escalation last month, searches the support system for details, and decides to lead the brief with that issue.
For a different account with no history, it might instead research the company's recent hiring. Its plan changes with the situation.
All three can be useful. Only the third needs agent-level governance, because only the third makes choices nobody wrote down in advance.
Why the distinction matters
Agents introduce a specific kind of uncertainty: they choose. They can choose well in situations nobody anticipated, and they can choose badly in ways nobody anticipated. Managing that uncertainty is the core discipline of agentic marketing.
That discipline requires controls simple automations do not need. We suggest four as a minimum for any system that meets the definition.
- Tool boundaries: clear limits on which tools an agent may use and what each tool can change.
- Decision logs: a record of what the agent decided and the inputs it used.
- Matched review: review points set by the consequence of a mistake, not applied uniformly.
- Judgment testing: evaluation that tests decisions, not just output format.
Our post on tool design for agents covers the first control in detail. The autonomy ladder describes how review should loosen as an agent earns trust.
When a team says it is doing agentic marketing but has none of these controls, it is usually running prompt templates with extra steps. That can still be worth doing. It just needs to be measured as what it is.
A simple test
We use one question to sort systems quickly. If the situation changes, does the system change its plan?
If yes, you are dealing with an agent and should govern it like one. If no, you are dealing with automation or a template, and you can manage it with the tools you already have.
The test is useful in buying decisions too. Asking a vendor to show what their system does when an input is missing or unexpected reveals more than a feature list. Our list of questions to ask agentic marketing vendors builds on this.
What to measure
Once systems are labeled, measurement can match the label. For templates and automations, the usual measures apply: output quality, error rate, and throughput.
For agents, we suggest adding measures of judgment. Track how often the agent's chosen plan differs from the default path, and sample those cases for review. Track escalations, both how often the agent asks for help and how often it should have asked but did not. These numbers show whether the agent's decisions are adding value or adding risk.
Where the line is moving
The boundary between automation and agent is not fixed. Workflows are gaining model-driven steps, and agents are being wrapped in stricter rules. We expect many production systems to be hybrids, with agent judgment confined to a few decision points inside a predictable workflow.
We are now studying how to describe those hybrids precisely, so teams can govern each decision point at the right level rather than labeling a whole system one way. We suspect the useful unit is the individual decision, not the system, and we are testing that idea across several common marketing workflows.
How to apply this: inventory your current AI tools and label each one as a template, an automation, an interface, or an agent, using the test of whether it changes its plan when the situation changes. Apply agent-level governance only where the system makes decisions.
Related research
The anatomy of a marketing agent: goal, tools, memory, and guardrails
Every dependable marketing agent we study has the same four parts. Get each one right and most failure modes become easier to prevent and easier to diagnose.
Brand voice as a specification: turning style guides into agent instructions
Most brand voice guides describe a feeling. Agents need a specification: observable behaviors, contrasting examples, clear boundaries, and tests.
How to design your first marketing agent
A practical walkthrough for scoping, building, and governing an agent that ships real work.
Put this research to work
Marketix turns these patterns into software: agents, automated workflows, and a hub that keeps people in control. Sign up for early access.
