Agent, rule and assistant
Three things get sold under the same word. They behave differently when something unexpected happens, which is the only test that matters.
| Trigger | Decides | Fails by | |
|---|---|---|---|
| Rule | A condition you wrote | Nothing. It executes | Firing correctly at the wrong moment |
| Assistant | You asking | Nothing. It suggests | Being ignored |
| Agent | Its own reading of the account | What to do and whether to do it | Acting on a conclusion that was wrong |
A rule that pauses any keyword above £50 cost with no conversion will pause your best brand term on the morning your checkout breaks. It is working exactly as written. An agent is supposed to notice that conversions fell across every campaign at once, conclude that the problem is not the keyword, and escalate instead of pausing. Whether it actually does that is the question to put to a vendor.
What an agent does in a marketing account
Four steps, in order, every time:
- Read. Pull the current state across the accounts it has access to, not one platform at a time.
- Judge. Decide what is worth acting on, and what is noise, a seasonal pattern or a symptom of something upstream.
- Act. Make the change, inside spend and scope limits a person set in advance.
- Record. Write down what it did, the data it was looking at, and why, in a form somebody can argue with afterwards.
Step four is the one that gets dropped, and it is the one that makes the rest safe. A system that acts and cannot explain itself is not an agent, it is an outage waiting for an audience.
Why the term appeared when it did
Marketing automated execution years ago. Bid management, email sends, budget pacing and reporting have been automatic for a decade. What stayed manual was judgement: deciding which of nine hundred findings is worth acting on this week, and in what order.
Large language models made that judgement cheap enough to apply continuously rather than in a monthly review. "Agentic" is the label for the software that followed. The underlying shift is that the scarce resource moved from execution to attention, which is the same argument behind autonomous performance marketing as a category.
What agentic marketing is not
- Not a chatbot on a dashboard. Answering questions about data is an assistant. It changes nothing.
- Not automation renamed. If the vendor can hand you the list of conditions it fires on, it is a rule engine, and a good one is still useful. It is just not this.
- Not unsupervised. An agent with no limits is not more advanced, it is less finished. Guardrails are the product, not a restriction on it.
- Not a replacement for the marketer. It removes the labour of finding and executing. Deciding what the account is for stays where it was.
How to evaluate a claim
Ask the vendor to describe a case where the agent decided not to act, and why. A rule engine has no answer to this, because a rule that does not fire has no reasoning behind it. An agent should be able to tell you that it saw a 40% drop in conversions, checked whether it was isolated to one campaign, found it was site wide, and escalated rather than optimising into a broken checkout.
If the demo only ever shows the system doing things, you are watching automation with better copywriting.