Shooting Nothing: AI Video at Campaign Speed

The interesting thing about generative video is not that it replaces a shoot. For a hero film with a face the audience has to trust, it mostly does not. The interesting thing is that it collapses the cost of the twentieth version of a film to roughly the cost of the first — and modern campaigns do not need one film, they need twenty.

01The real bottleneck was never the shoot day

Count what a single campaign actually ships: a 16:9 for broadcast and YouTube, a 9:16 for Reels, Shorts and Stories, a 1:1 for feed, three cutdown lengths for paid, and all of that again in every language the plan runs in. A shoot produces the raw material for one of those properly and the rest as compromises cropped out of it. A landscape film cropped to vertical is not a vertical film, and it performs like what it is.

02Where generation genuinely wins

• Variants of a film that already works — same idea, twenty executions, tested against each other.
• Shots that do not exist and cannot be sourced: a product in a location nobody is flying to.
• Language versions where the voice and the on-screen type have to change together.
• Motion, type and transitions built to a brand system rather than to a template's defaults.

03Where it does not

Anything with a human face carrying the emotional weight of the film. Anything where a specific real place or a real product has to be unambiguously itself. And anything where the client's legal team needs to know exactly where every frame came from. Being straight about that list is what makes the rest of the offer credible.

04Board before you generate

The fastest way to waste a week is to prompt without a plan. Generation time is cheap per clip and expensive in aggregate, and footage produced without a board is footage nobody can cut. The order that works is the order that always worked: script, storyboard, sign-off, then production. The only thing that changed is what production costs.

05Post is still post

This is the part that separates usable from uncanny, and it is the part most demonstrations skip. Grade, clean-up, sound design and the frame-by-frame fixes for the details the model got wrong are all still real work done by real people. A generated film that has not been finished looks generated. A finished one mostly does not — and on a nine-second vertical cutdown, nobody is looking for the seams anyway.