What AI in marketing genuinely does — and where it only pretends

In two years AI has gone from a toy to the tool half of an agency's routine runs on. It has also picked up expectations it cannot carry. Here is an honest look at where it actually saves money.

Where AI beats a person

Volume and repetition. Going through a hundred competitor listings, reconciling a ten-thousand-row export, checking three hundred pages for broken markup — tasks where a human makes mistakes out of fatigue and a model does not. That is not "smarter", it is more durable.

Speed of a first version. A landing page draft, twenty headline options, an article outline — what used to take a day takes an hour. The operative word is "draft": it always needs editing, but editing is faster than starting from nothing.

Monitoring. Checking daily what moved in the search results, in competitor pricing, in reviews — work a person abandons in week two. A model does it every day at the same hour and does not forget.

Research with sources. Gathering industry data and bringing back links rather than a retelling. Demanding the source is essential here: without one, any number from a model is an invention with a probability you do not control.

Where it reliably fails

Numbers from memory. If the model did not visit a source but "recalled" one, it will produce a plausible number. Plausible, not correct. It is the most expensive way to be wrong, because the error is invisible.

Decisions. What to launch, what to switch off, where to move budget — a model will confidently propose an option, but it does not carry the outcome and does not know your economics. The decision belongs to the person with money at stake.

Positioning. AI is excellent at phrasing what has already been worked out and poor at working out what has not. The offer, the differentiation, the answer to "why you" — that is the job of someone who knows the market.

Brand voice without examples. With no samples of your voice, a model writes in a smooth average. That is exactly the "generated text" people recognise from the first line.

Why "generate content" is a bad brief

Not because the text will be weak, but because the task is framed wrongly. The value is not in getting a text — it is in getting a repeatable process: the same job, done equally well the hundredth time.

For us that looks like this:

• A skill built for the task. Not "write an article" but a written protocol: where the data comes from, in what order, what gets verified, what a finished result looks like.
• A checking loop. The task runs through "did it → checked it → redid it" until it clears quality control. The checker is also a model, but working from someone else's checklist.
• Connectors to real data. The model works with exports from ad accounts and live search results, not with what it "remembers" about your industry.
• Only what has been tested. What reaches a client is what we ran on our own projects first.

How this changes the economics

Simple arithmetic. The volume of work an agency staffs with a department of five is covered by a team of six with a named owner for every area. Not because people work three times faster — because the routine that ate three quarters of the time moved to the models.

The freed-up time goes where machines do not help: strategy, the conversation with the client, reading the numbers.

The practical takeaway: when choosing an agency that uses AI, do not ask "which models do you use". Ask "what exactly does a human check before it reaches me". The second answer tells you far more about quality.