Why AI slips exactly in marketing: the feedback arrives weeks too late
There's a course for everything. "AI COURSE FOR MARKETING!! Grab it before your competitors take it." I'm about to save you the money.
Models learn where the feedback is instant
Neural networks work wonderfully where the answer about the result arrives on the spot: you wrote, you saw, you corrected, you repeated. That's why the first to go down were copywriters, developers and designers: there the cycle "I did it - I checked it" closes in seconds, and the model manages to learn from its own misses on the fly. It is the same loop that makes a/b testing work at all, and the same loop marketing almost never has.
In marketing the feedback arrives weeks later
And now take marketing. The time to the result is measured not in seconds but in weeks, if not months. You did something, and the feedback is going to show up much later, if it arrives at all, and only provided you're smart and set up the analytics. First the warm-up, then the touchpoints, then the delayed conversions, and only a month later do you find out that you poured the budget down the drain back in the second week.
Here the neural network simply has nothing to learn from quickly. Or rather it does, only the signal arrives so late that by the time the "oops, it didn't work" appears the money is already spent, and the real kind, the audience has changed its patterns and the combination has burned to dust. That is also why marketing attribution is a whole discipline of its own here and a footnote everywhere else: the customer acquisition cost you're optimising against is a number you only see afterwards, and marketing roi is the last of the three to arrive. Doing marketing by eye, on vibes, like a landing page over a weekend, is the direct road to an empty ad panel and a very sad ROAS.
The unglamorous part: making the model actually work
I have spent an indecent amount of time fighting heavily with neural networks to get out of them not a stream of pretty water, but a fairly clean and defensible result: that the model picks up the tools properly and doesn't poke at them blindly, that it reads the context and doesn't invent it, that it doesn't produce hallucinations, doesn't burn budget and doesn't drown the report in filler. That is what people are actually buying when they buy ai agents - not the model, the plumbing around it.
It's considerably less glamorous than it sounds in hype posts about courses. It's meticulous and at times demented fiddling over every detail.
What it looks like when it works
Research on a competitor: 30 minutes. Market research: about 25. Evaluating an idea: something like an hour. And the depth can be turned according to the task: from several thousand keywords to hundreds of creatives or tens of thousands of mentions, and what comes out is not porridge but a normal summary with a breakdown.
When you have large volumes of research, that doesn't save hours: it saves weeks and thousands of dollars. That is what marketing research on autopilot was made for, and for the same reason competitor research stopped being a two-week project. This is the honest boundary of marketing automation: it collapses the gathering, it doesn't shorten the feedback loop.
Next: automating the start itself
The most interesting thing is that the launch of those studies can also be automated: so that they start when they should and without your participation. For now that's only in my wet architectural dreams, but the outline is visible. Continuous marketing experimentation without a human pressing the button is the only version of this that scales.
Conclusion
I'm with all my might in favour of neural networks in marketing and radically against firing by eye and calling it a strategy. The whole difference between "I asked the model and crossed my fingers" and "I forced the model to actually work" is, by the look of it, the main profession of the next couple of years.
