What AI Still Cannot Do in Marketing

We sell an AI platform, so this is an odd thing for us to publish. It is also the more useful list. Everybody writing about AI in marketing is describing the frontier of what it can do, which moves and is therefore hard to plan against. The things it is structurally bad at have barely shifted in three years, and knowing them is what stops you handing over the one part of the job that was actually keeping the business alive.

Two reasons, and neither is modesty.

The first is that overselling produces churn. A business that hands over the wrong work, gets a bad result and concludes AI does not work is worse for us than a business that never tried, because it has an expensive story to tell other people.

The second is that the limits are the design brief. Everything about how our product works, particularly the part where it asks before it sends, exists because of the fifth item on this list. If we pretended the limits were not there we could not explain the product.

Ask a model which of three campaigns to run next quarter and it will answer. The answer will be articulate, structured, and produced without any knowledge of the things that decide it: that one customer segment is about to churn for reasons nobody has written down, that the founder has a personal relationship in one of those markets, that cash is tight until March.

Strategy is mostly constraint knowledge, and constraints live in people's heads. A model reasons from what it has been given. In a small business, the most important inputs have never been given to anyone, including sometimes the person making the decision.

This sounds obvious and is routinely forgotten in practice, because the output reads as though it knows things.

A model writing to your customer does not know that this customer complained last year, that their contract renews in six weeks, that they mentioned a redundancy round, or that the last person who emailed them got a sharp reply. It knows what is in the record. In most small businesses, the majority of what matters about a relationship is not in the record.

The consequence is specific and it is the most common way AI outreach embarrasses a company: a technically appropriate message sent at a humanly inappropriate moment. Nothing in the text is wrong. The problem is context nobody wrote down.

A model can produce competent work at volume. It cannot reliably tell you whether what it produced is any good, because good is not a property it can measure.

The reason is structural. Models are trained toward the centre of what exists, so asked for a post about a topic, they produce something near the average of everything written about that topic. Average is a perfectly good target for a booking confirmation and a fatal one for anything meant to distinguish you.

This is also why so much AI marketing content reads identically across companies. It is not that everyone is prompting badly. Everyone is drawing from the same distribution, so everyone lands in the same place, and the more people do it the more the centre reinforces itself.

Most small business marketing is not persuasion at scale. It is a small number of relationships, maintained over years, that produce referrals and repeat work.

An agent can remember every interaction, which is more than the human can. What it cannot do is be the relationship. When someone chooses your firm over a cheaper one because of how you handled a bad situation three years ago, the thing they are choosing is a person having behaved a particular way. That cannot be delegated, and attempting it damages the asset.

This is the failure mode with the longest tail. Nobody tells you they received a generated note that felt hollow. They simply weight your next message slightly lower, permanently.

This is the most important item and the least discussed, because it is the one that does not improve with model quality in the way people assume.

A junior employee who is unsure sounds unsure. That signal is doing enormous work in any organisation: it tells you where to look. A model produces an invented statistic with precisely the fluency it produces a correct one. The confidence carries no information.

Everything about how a sane AI workflow is built follows from this. It is why a human approves anything that leaves the building, and it is why the approval is not a formality. If the model could tell you when it was unsure, the gate could be conditional. It cannot, so it is not.

The useful test is not how important the task is or how senior the person doing it. It is this: can you verify the output faster than you could have produced it?

Hand over drafting from a brief, summarising a long thread, reformatting, first-pass research you will check, chasing people who have not replied, scheduling, and monitoring for changes. All of these are checkable at a glance.

Keep deciding what to work on, anything where the relationship is the product, final approval on anything customer- facing, and any claim you would be embarrassed to have invented.

The edge cases are where judgement is required, and the honest guide is the cost of being wrong. A generated internal summary that is slightly off wastes ten minutes. A generated statistic in a published article that is invented costs credibility you spent years building, and you will not find out from the person who noticed.

The list of what AI does well is in AI agents for small business and the one-person marketing piece. Both are worth reading against this one, because the useful position is not enthusiasm or scepticism. It is knowing precisely which half of your work is which.