AI Marketing Agents: Which Half of the Job Can They Do?

An AI marketing agent is software that takes a marketing goal, works out the steps, and carries them out: research a topic, draft the asset, schedule it, report on it. The category has moved fast and the honest assessment is lopsided. Generating things is close to solved. Deciding what is worth generating, getting it in front of the right people, and knowing whether it worked are not, and those were always the hard parts.

Which means the risk with these tools is not that they produce bad work. It is that they produce a great deal of adequate work, very cheaply, and you end up with more published material and no more customers.

Here is what matters most:

• Drafting and repurposing genuinely work. Targeting, distribution and attribution do not, yet.
• The failure mode is volume without effect: more assets, same pipeline, and now a maintenance burden.
• Tie any agent to one measurable outcome before you turn it on, or you cannot tell if it helped.
• Search engines penalise scaled content produced without added value, so volume alone carries real risk.
• Keep a person between the draft and publication. That single gate is what separates useful from embarrassing.

The half that works

Drafting inside a known format. Give an agent a clear brief, a defined structure and real source material, and it produces a solid first draft. This is real time saved, and it works because a human edits before anything ships.

Repurposing. Turning one substantial asset into the six shorter things it should have become. Genuinely mechanical work, low risk, and the source material is already yours.

Research and gathering. Assembling what competitors have published on a topic, what questions people ask, what the existing material misses. Useful, and the output is a brief for a person rather than a decision.

Reporting assembly. Pulling numbers from several places into a consistent weekly summary. This is not really an agent job, it is ordinary automation, and it is one of the highest-value unglamorous wins available.

The half that does not

Deciding what to make. This requires knowing which customers are worth winning, what they actually search, and what your firm can credibly say that a competitor cannot. An agent will happily generate a plausible content calendar full of topics nobody searches for. We have direct experience of this: our own blog ran to over 150 posts and 86% of them earned no clicks at all, because the topics were chosen by plausibility rather than by checking real demand.

Distribution. Publishing is not distribution. An agent can schedule a post. It cannot get anybody to read it, and for most firms the constraint was always attention rather than production.

Attribution. Knowing which asset produced revenue. This is hard with humans and no easier with agents, and an agent reporting engagement metrics can give you a confident picture of activity that has no relationship to pipeline.

Anything where being wrong is public. A confidently incorrect claim about a regulation, a competitor or a price, published under your name, costs more than the drafting saved.

The trap: volume without effect

This is the specific thing to protect against, because it is the default outcome.

These tools reduce the cost of producing an asset to nearly nothing. That changes the economics of publishing, and the instinctive response is to publish more. What happens next is that you have four times the assets, the same pipeline, and a much larger surface area to keep accurate as facts change.

Worth being clear on the search consequence too, because it is commonly misunderstood. Search engines do not penalise content for being AI-assisted. They penalise scaled content produced without adding value, which is a different and narrower thing. A well-researched piece drafted with assistance is fine. Forty thin pieces on topics nobody searches is the pattern that gets treated as abuse, and it is exactly what these tools make easy.

So the constraint to impose on yourself is not a volume target. It is that every asset has to answer a question somebody is actually asking, with something only your firm can say.

How to scope one so it is measurable

Pick one outcome, not a volume. Not "publish twelve posts a month." Something like "reduce the time from brief to publishable draft below two hours" or "cover the twenty questions our sales calls keep repeating." Volume targets produce volume. Outcome targets produce outcomes.

Put the human gate in writing. Somebody named reads every asset before it publishes. This is the control that prevents the expensive category of mistake, and it is the first thing dropped when the volume target arrives.

Verify demand before generating. Check that a topic has real search volume and a page that could plausibly rank, before an agent writes about it. Our own corpus is the cautionary example: the traffic concentrated in five posts out of 152, and the other 147 were written without that check.

Measure the outcome, not the activity. Assets published, words generated and engagement rate are activity. Enquiries, calls booked and deals are outcomes. An agent reporting the first set can look successful indefinitely.

What it costs

The build is small if you are scoping properly: connecting the agent to your source material, defining the brief format, building the review step. Days to a couple of weeks for one job.

Running cost is per use and it scales with how much you generate, which is the opposite of how people budget for it. An agent that reasons through eight steps per asset costs roughly eight times one that takes a single step. Ask for cost per finished asset at your expected volume rather than a monthly fee.

Then size the saving honestly. Count assets per month, the hours each takes now from brief to publishable, and a fully loaded hourly cost. Twelve assets a month at five hours each is sixty hours; at forty dollars an hour, twenty-four hundred a month. If the agent halves the drafting time but the editing time stays, you are saving a portion of that, not all of it. The editing rarely shrinks as much as the pitch implies.

FAQ

Will AI-generated marketing content hurt our search rankings?

Not for being AI-assisted. Search guidance targets scaled content produced without adding value, which is about thin material published at volume rather than about the tool used to write it. The practical risk is that these agents make thin material cheap, so the same firm that would have published two researched pieces publishes twenty unresearched ones. A well-sourced piece that answers a real question is fine regardless of how it was drafted.

Can an AI marketing agent replace our agency or marketer?

It can replace a portion of the drafting, which is usually a minority of what either actually does. What it cannot replace is deciding which customers to pursue, which topics have real demand, what your firm can credibly claim, and how to get the material in front of anyone. Firms that cut the strategic role and keep the generation tend to produce more material with less effect.

How do we stop it publishing something wrong?

One named person reads everything before it goes out, and that is written into the process rather than assumed. Additionally, require that any factual claim, statistic or regulatory reference in a draft carries a source the reviewer can check, because unsourced specifics are where these tools are least reliable and where the damage is largest.

What is a realistic first project?

Repurposing, almost always. Take material you already own and have already verified, and use an agent to turn it into the shorter formats it should have become. The source is trustworthy, the risk is low, the time saving is real, and it teaches your team how to brief and review before anything customer-facing depends on it.

Key takeaways

• AI marketing agents are reliable at drafting and repurposing, and unreliable at targeting, distribution and attribution.
• The common failure is producing more assets with no change in pipeline, plus a larger maintenance burden.
• Search guidance penalises scaled content without added value, not content that was AI-assisted.
• Verify that a topic has real search demand before generating anything about it.
• Scope the agent against one measurable outcome rather than a publishing volume target.

If you want a straight read on whether your content problem is a production problem or a targeting problem, that is a short conversation. In our own case it turned out to be targeting.

Book a free strategy call

Related reading: AI agents for business · why AI sales agents fail · what is workflow automation

Services: business process automation consulting · AI agent development