How AI changed the creative testing loop

For most of the history of paid social, the expensive part of a creative test was the creative. A shoot, an edit, a few rounds of feedback. By the time an ad went live, a lot was riding on it. So brands tested cautiously: a handful of concepts, a long wait for a clear read, and a strong temptation to call a winner too early.

AI tools have broken that equation. A new hook, headline, voiceover, cut-down or background can now be produced in minutes. That doesn’t make creative free, but it makes it cheap enough that the constraint moves somewhere else. Most of the gains we see come from noticing where it moved.

The bottleneck moved

When production stops being the slow part, three other things become the limit:

• Ideas worth testing. Volume is easy. Distinct, well-reasoned ideas are not.
• Budget to get a read. Every ad still needs enough spend to tell you something. More ads with the same budget means thinner data per ad.
• The discipline to learn. A team that launches fifty ads a week and doesn’t write down what it learned is just generating noise faster.

So the job changed. We spend less time making and more time deciding what deserves to be made, and what a result actually means.

Concepts, not variations

The trap with cheap production is volume for its own sake. Twenty versions of the same ad with different background colours will tell you almost nothing. We split every test into two layers.

Concepts are distinct reasons to buy: a problem, a promise, an audience, a moment in someone’s day. A new concept is a genuine bet, and it’s where big swings come from.

Variations are different hooks, formats, lengths and edits of a concept that is already working. They’re where efficiency comes from: keeping a winning idea fresh and finding the version that travels furthest.

AI is most useful in the second layer, where it multiplies a proven idea quickly. In the first layer it’s a sparring partner. It’s good for pressure-testing angles, mining reviews for language customers actually use, and drafting first passes. It doesn’t replace the judgement of someone who understands the customer.

A loop, not a launch

Our creative work runs on a weekly loop rather than a campaign calendar:

1. Read. What did last week’s ads tell us? Not just which won, but why we think they won.
2. Brief. Turn those reads into a small number of new concepts and a larger number of variations on what’s working.
3. Produce. Creators, editors and AI tools each take the parts they’re best at.
4. Launch. New ads go into a testing structure with enough budget to get a signal, separate from the ads already scaling.
5. Record. Every result goes into a shared log: the hypothesis, the outcome and the learning.

The log is the most underrated part. After a few months it becomes a map of what this brand’s customers respond to: which problems, which proof, which faces, which first three seconds. That map is worth more than any single winning ad.

Where people still matter

Some things we won’t hand to a model. Taste is one: knowing when an ad is on-brand and when it’s just loud. Real customers are another. We use AI to produce and adapt creative, but never to invent customers, testimonials or results. People can tell, platforms are increasingly strict about it, and a brand built on fake proof doesn’t last.

Creators matter more than ever, too. A real person using a product on camera carries a kind of trust that generated footage doesn’t. The best work we see combines both: creators supply the authenticity, and AI multiplies it into the many cuts, hooks and formats that paid media needs.

What this means for a brand

If your creative process still treats every ad as a small production, you’re likely testing too little and learning too slowly. The fix isn’t to buy a tool and generate a hundred ads. It’s to rebuild the loop: fewer, sharper concepts; many fast variations; honest reads; and a record of what you learned. Speed matters, but only when it’s pointed at learning.