Most of what is written about generative production is an argument for using it. Here is the other half, which is what clients actually need before they commit a budget.
01 — The hero product
If the product is the subject and the audience knows it well, it gets built or shot. A trainer with the wrong number of eyelets, a helmet with a vent that does not exist, a bottle with a label that almost reads — these fail at exactly the moment the audience is paying most attention. CGI is not slower here. It is the fast route, because it is right the first time.
02 — Anything with legal or factual weight
Technical products, safety equipment, medical, automotive detail, anything that appears in a claim. An approximated component is a liability, not an image.
03 — Typography and brand assets
Type gets set. Logos get placed. Every time a generated frame carries almost-correct lettering, it becomes a retouching job that costs more than compositing clean type would have.
Use AI where variation is valuable. Build where exactness is non-negotiable.
04 — Identity decisions
A mark, a visual language, a positioning. Generative tools are exceptional at producing many plausible options and structurally incapable of preferring one. That preference is the entire job. Where they do earn their place is immediately afterwards: showing a decided identity in forty environments in an afternoon.
The pattern
Generation is strongest where variation has value and weakest where precision is the point. Most campaigns contain both, which is why the useful question is never should we use AI. It is which parts of this specific job belong on which side of that line — and that judgement comes from having produced the work before the tools existed.