AI video is a pipeline, not a button

The most common mistake in generative video is treating generation as the whole job. Someone types a prompt, gets four seconds of something impressive, and then discovers there is no way to make a second shot that matches it.

Generation is one step out of six. The other five are what make the output usable in an ad account.

The six steps

• Angle map — decide what the ad is arguing before anything is generated
• Reference build — assemble a locked kit of brand plates, colours and product geometry
• Generation — produce plates against that kit, not against a blank prompt
• Composite — put the real product back into the generated environment
• Grade and sound — the step that most obviously separates amateur from shipped
• Version — cut every winner into hook, length and market variants

Steps two and four are where the actual craft lives. A generated environment with your real product composited into it reads as production. A fully generated product reads as a video game, and customers notice immediately.

Where AI genuinely wins

B-roll, environments, transitions, seasonal variants, language localisation and volume. These are the places where a shoot day was always a bad use of money, and where the tooling is now clearly good enough.

Where it does not

A real customer's face saying a real thing. Hands doing something dextrous. Anything where the claim depends on the viewer believing a person did it. We still point a camera at those, and we always will.

The model is the cheapest part of the process. The kit you feed it is the expensive part, and it is the part that compounds.

Build the kit once — brand plates, product geometry, type system, voice profile — and every subsequent generation inherits it. That is the actual unlock. Not the model. The kit.