Marketing is the part of shipping a side project that most developers postpone until launch day, then do badly for a week and quietly abandon. The instinct now is to hand the whole thing to an AI agent. That works, but only for about half of it, and picking the wrong half is why people conclude the tools do not work.
Here is the split that keeps holding up in practice.
What Hands Off Cleanly
The tasks that automate well share one property. They repeat, and you can check the output afterward without much thought.
Site audits are the clearest case. Missing title tags, pages with no meta description, internal links pointing at 404s, images with no alt text, templates that got slow. This is mechanical work with a right answer, so an agent can run it nightly and hand you a diff instead of a report nobody opens.
Follow up sequences are the second. Most solo projects have exactly one email, the welcome message, because writing five more never reaches the top of the list. An agent drafting a five step sequence from your existing docs gets you something to edit, and editing is far faster than starting from an empty box.
Then the plumbing. Segmenting a list by behavior, routing a form submission to the right place, generating the same weekly numbers pull, turning one long piece into the short versions each platform expects. None of it is interesting, all of it compounds, and all of it is verifiable at a glance.
What Breaks When You Automate It
The failures are consistent, and they are not really failures of the model.
If your targeting is wrong, automation scales the wrong targeting. If your positioning is vague, you get fifteen well formatted versions of a vague message. If your offer is not compelling, an agent will happily produce a hundred variants of an offer nobody wants, and the volume makes the underlying problem harder to spot rather than easier.
That is the part worth internalizing. AI is a multiplier on what you already had, so when the output feels off, the fix is almost always upstream of the tool. Nobody enjoys hearing that the problem is the positioning, which is why we keep buying tools instead.
The other consistent break is anything with taste in the loop. Naming, the shape of a launch, deciding which objection to answer first. Those are judgment calls, and a judgment call handed to an agent comes back as an average of everything it has ever seen.
A Practical Order For A One Person Team
Automate in this order, and stop when the next thing needs your judgment.
First the audit, because it runs forever and needs no input from you. Second the follow ups, because a mediocre sequence you actually shipped beats the perfect one still in your head. Third the reporting, because it is the task most likely to silently stop happening the week you get busy.
Keep positioning, pricing and the offer itself in your own hands. Those three decide whether any of the automated work has something worth amplifying.
If you want the longer version with the specific jobs mapped out, we keep a breakdown of AI marketing work that splits the tasks into what hands off cleanly and what still needs a person.
The Takeaway
The question is never whether AI can do marketing. It is which half of your marketing is repetitive and checkable, and which half is a judgment call wearing a task's clothing. Sort the list that way first, and the tooling decision mostly makes itself.
Top comments (0)