AI in ad production: what we automate and what we still shoot
Generative tools now write, cut and resize faster than any team. That solves the cheap half of the problem. Here is where we use them and where a camera still has to leave the office.
What we hand to a machine
Resizing is the obvious one. One hero cut becomes nine formats for feed, story, reel and display, and nobody on the team should spend an afternoon on that. The same goes for subtitle burn-in, rough cuts from a long interview, and the twentieth variation of a headline that only differs in the call to action.
Static product shots on a plain background are close behind. If the object is the whole picture and the light is even, a generated backdrop is cheaper and faster than a studio day, and no one can tell.
What still needs a camera
Anything with your actual venue, your actual crowd or your actual crew. A festival ad works because the person watching recognises the place. A generated crowd does not carry that, and the audience notices long before it can say why.
Faces are the second line we do not cross. Talent has a right to their likeness, a rider often says so explicitly, and the reputational cost of a synthetic artist portrait is far higher than a shoot day.
The part that actually decides performance
The tooling debate distracts from the thing that moves numbers. What a campaign lives on is the first two seconds, a clear offer and a checkout that reports back correctly. A generated background does not fix a broken purchase event, and a beautifully shot spot does not either.
We treat AI as a production discount, not a strategy. It buys us more variants to test in the same budget. What those variants say still comes from the brief.
How we brief it
Every generated asset gets the same review as a shot one: does it match the brand, is it honest about the event, would we defend it to the client. Assets that only exist because they were free to make do not go live.
What it changed about the schedule
The bottleneck moved. It used to be production - waiting on a cut, waiting on a resize. Now it is review. When a team can generate forty variants before lunch, somebody has to decide which four are worth spending media on, and that decision needs the same person who wrote the brief.
We plan for that explicitly. Fewer, better-defined variants beat a flood, because a flood produces decision fatigue and then the safest option wins by default. Safe options do not sell tickets.
Version control matters more than it used to
When assets are cheap to make, they are also cheap to lose track of. We keep one naming convention and one source of truth for what is live, or the client sees a version we killed two weeks ago.
The same applies to rights. Every asset gets a note on what it was made from and whether it can run beyond the campaign. Generated material with unclear provenance is a liability the moment a brand wants to reuse it.
Do you use AI images in client campaigns?
For backgrounds, resizing and variation work, yes. Not for crowds, venues or people - if a viewer should recognise the place or the artist, we shoot it.
Does AI production make campaigns cheaper?
It lowers the cost per variant, which means more can be tested on the same budget. It does not lower the cost of strategy, tracking or media.