What to automate in marketing, and what you never should
Automation is good at repetition and bad at judgement. Most teams get that split backwards, then wonder why the output reads like nobody wrote it.
Every marketing team is now running some version of the same experiment: point a model at the work and see what it can take over. The results are wildly uneven, and the reason is rarely the tool.
It is that most teams automate the wrong half. They hand over the thinking and keep the typing, when the useful split is the exact opposite.
What follows is the test we use to decide, and what it costs when the decision goes the other way.
The split that decides everything
There are two kinds of work inside any marketing operation, and they behave completely differently under automation.
The first kind is repetition. Same input shape, same output shape, done often, judged right or wrong by a rule anyone can state. Resizing, reformatting, transcribing, tagging, checking a page against a checklist.
The second kind is judgement. It is deciding what is worth saying, what to leave out, and what a specific reader will find credible. There is no rule to state, which is why it is hard to hand over.
Automation is extremely good at the first and structurally unable to do the second. Every disappointing rollout we have seen is a team that applied it to the wrong one.
Where automation genuinely earns its keep
The unglamorous work is where the return is. It is also where nobody minds, because nobody enjoyed doing it.
- Turning one asset into its five required formats.
- Checking published pages against a standards list nobody remembers in full.
- Pulling the recurring objections out of a pile of call transcripts.
- Drafting the first pass of something whose shape is already decided.
Automation is good at repetition and bad at judgement. Most teams get that split backwards.
Notice the last one. A first draft from a specified brief is repetition. A first draft from a topic is judgement, and that is where output starts reading like it was assembled rather than written.
The work that cannot be handed over
Some decisions are the product. Handing them over does not save time, it removes the thing customers were responding to.
The position is the obvious one. So is the decision about what you are refusing to claim, which is invisible in the output and load-bearing in the strategy.
Less obvious: deciding whether a piece is true enough to publish. A model will produce a confident sentence about a number it has no basis for. Somebody has to be accountable for every figure carrying its source, and that accountability cannot sit with a tool.
A test for any task you are considering
Before automating something, answer two questions about it. The answers put the task in one of four boxes.
| Rule is statable | Rule is not statable | |
|---|---|---|
| Done often | Automate it fully | Automate the draft, review every one |
| Done rarely | Not worth the setup | Keep it human, always |
Most teams try to run the bottom right box through a tool because it is the box that takes the most time. That is precisely why it takes the most time.
What goes wrong when judgement is automated
The failure is not dramatic. Nothing breaks, which is what makes it expensive.
Output volume goes up and everything reads slightly generic. It passes review, because there is nothing identifiably wrong with it. Six months later the blog is bigger and no more persuasive, and nobody can point at the week it went wrong.
The second failure is quieter still. The team stops practising judgement. Deciding what matters is a skill that degrades when it is not used, and a year of not using it is hard to reverse.
Keeping a human in the loop without slowing down
"Human in the loop" usually means somebody rubber stamps a queue, which is not review. Review is somebody with authority saying no often enough that the yes means something.
Two habits make that practical. First, the reviewer reads for the argument, not the prose, so review is fast. Second, the reviewer is allowed to kill things. A loop with no rejections is a loop with no signal, and it is worth measuring rejections for exactly that reason.
The version of this that lasts
The teams getting real leverage are not the ones automating the most. They are the ones who wrote down what only they can decide, automated everything around it, and left that core alone deliberately.
That list is usually short. The position, what you refuse to say, whether a claim is defensible, and which objection matters most this quarter. Everything else is logistics, and logistics is what machines are for.