A Practical Framework for Applying AI in Advertising — Beyond Generating More Ad Copy

Many advertisers I speak with are using AI to do what they already did, only faster. More ad variations. More creative iterations. More headlines tested per hour. This is not wrong, and it has real value — but it is largely an existing workflow accelerated rather than a new one designed. The deeper leverage in applying AI to advertising, in my view, is less about producing more of the same and more about rethinking where human judgment is spent.

I find it useful to think about AI in advertising across three distinct layers, because conflating them is where a lot of implementations lose their way: generation, judgment, and orchestration.

Generation is the obvious layer and the most widely adopted. This is AI producing assets — copy, images, variations, descriptions. The value is real, and it meaningfully lowers the cost of creating creative. But it is also where the field is crowding fastest, which means it tends to confer the least lasting advantage. If most teams can generate a large volume of variants, then having a large volume of variants stops being the point, and the constraint moves to the next layer.

Judgment is where I find AI more interesting and less understood. This is using models not to make things, but to reason about things. A concrete example: feeding a model the open-ended responses from a customer survey, the reviews on your product, and the support tickets nobody has time to read in full, and asking it to surface the recurring themes and the language customers actually use. The output is not an ad. It is a set of hypotheses about what an audience cares about and how it describes that — the raw material for positioning you might otherwise be guessing at. Used this way, AI behaves more like a research instrument than a copywriter, and I think that is a higher-leverage use.

Orchestration is the layer I see fewest teams using well so far, and the one I expect to matter most over the coming years. This is AI not as a tool you invoke once, but as a system that runs a loop: monitoring performance, forming a view on why something is underperforming, proposing or making an adjustment, and observing the result. An agent that reallocates budget across campaigns as signals change, or flags creative showing fatigue, operates in the gaps between human attention. The work was always there; people simply could not do it continuously.

If I were implementing this today, I would resist starting with generation, precisely because it is the easiest and least differentiating. I would begin with judgment — pointing a model at the qualitative data a team is already accumulating but rarely reads in depth, and using what comes back to sharpen its hypotheses. Then I would automate one narrow, well-instrumented orchestration loop: something with a clear metric and a limited blast radius, learning from how it behaves before widening its mandate. Generation I would treat as the increasingly commoditised capability it is — useful and cheap, but not where the contest is decided.

There is a risk worth naming plainly. AI systems optimise relentlessly toward whatever they are measured against, which makes them very good at improving a proxy and largely indifferent to whether that proxy still means what we assume. A system optimising for clicks may find the clicks that convert least well. One pointed purely at short-term return can quietly under-invest in the brand equity that does not appear in this quarter's numbers. The more autonomy these systems are given, the more carefully the objective has to be chosen, because they will pursue it more literally and tirelessly than any person would. In that sense, the hardest part of applying AI in advertising is less technical than it is about deciding what is genuinely worth optimising for.

That is the point I keep returning to. As generation becomes inexpensive and orchestration becomes more automated, the scarce resource in advertising shifts from production toward judgment and taste — knowing what is worth making and what is worth measuring. The teams that do best, I suspect, will not be the ones with the most AI, but the ones clearest about what they want from it. So the question I would put to any team adopting these tools is a simple one: once the machine can do the work, do you actually know what good looks like?

The above reflects my personal views only and is intended for informational and discussion purposes. It does not represent the position of any employer or organisation.