How to brief an AI film studio

Two sentences from people who do this work explain most of what goes wrong between a brand and an AI studio. Tim Katz, co-founder of Knucklehead: "changing a shot late is really hard." Chris Hewitt, creative director at Airhead: leave details unspecified and the model fills the gap with "the AI slop we see all over social media." Both quotes are from the same interview. Under-specify and you get slop. Specify late and it costs. A brief that works has to clear both.

That is why an AI brief is not a production brief with a different supplier name on it. In live-action, late changes are expensive but normal, and you buy flexibility with a bigger crew day. In AI work the decision points sit earlier and are harder to reverse, because a change to a character or a look does not re-light a set, it re-generates everything built on top of it.

Decide one thing before you write anything

Is there a character that has to stay itself?

That single question moves the schedule by an order of magnitude, and the public examples make the point better than any estimate we could give. Svedka's Super Bowl LX spot took about four months, most of it on one character. Kalshi's NBA Finals spot took three days, one operator, no character to hold, with 300 to 400 generations producing 15 usable clips. The Kalshi production account.

Made in

3 days

Crew

One

Generations

300–400

Usable clips

15

Both are finished, aired work. The difference between four months and three days is almost entirely whether something in frame had to remain identical from shot to shot. If your concept needs a recurring character, say so in the brief and budget for it as a build phase, separate from the films it appears in. If it does not, say that too, because it frees the studio to work much faster.

Specify the things that must be true

The useful division is not detailed versus loose. It is fixed versus discovered. Fix everything that has to be identical across the campaign or that someone will sign off against:

• The character, product or logo, with real reference material. Photographs, CAD, packaging, a 3D file if one exists.
• The look, as references rather than adjectives. "Warm", "premium" and "cinematic" are not instructions.
• The full deliverable list, including cutdowns, aspect ratios, language versions and any market variants. Not "plus socials later".
• Non-negotiables: legal copy, product accuracy, brand colour, anything regulated.
• Where it runs, which determines the disclosure obligation and the format spec.

Dave Clark of Promise framed the limit of prompting directly: "You can't prompt your way to the ends." The hard parts, in his account, are holding one look across a long runtime and communicating a vision to a large team. Forbes, July 2026. Both of those are brief problems before they are craft problems. A brief that defines the look precisely enough for forty people to recognise a wrong shot is doing real work.

Leave the method open

The thing most clients over-specify is the thing the model is actually good at choosing. Camera movement, blocking and lens character all get decided by the model unless directed. Paul Trillo made this argument back in 2023, and we cite it as the origin of the idea rather than as current testimony about today's tools. His account of that early work.

“A brief that defines the look precisely enough for forty people to recognise a wrong shot is doing real work.”

In practice a shot described as "slow push in on her face, 85mm, shallow" narrows the output to whatever the model does when pushed into a corner. A shot described by what has to be felt, with the character and look locked by reference, leaves room for the better version to appear. This is practitioner advice rather than a published finding, and it has a limit: it only holds when the fixed list above is genuinely fixed. Vague look plus vague camera is how you get Hewitt's slop.

References and IP

Jason Zada of Secret Level put the responsibility where it belongs: "If you're prompting 'Star Wars,' it's your fault, not the model's fault." The WrapWrap interview. Reference boards full of other people's films and other brands' campaigns are standard practice in advertising and a liability here, because the model will honour them more literally than a DOP would. Send references for light, texture, pace and performance. Strip out anything whose value is that it is a recognisable property.

Versioning is where the economics are

AI production does not reduce cost evenly. It reduces the cost of the second, tenth and hundredth variant. Copacino Fujikado delivered a campaign that needed 26 models across several locations for 26% of live-shoot cost. Digiday's reporting on agency pricing.

Read that as a guide to which briefs suit this. Many deliverables, many variants, many markets: the saving is real. One hero film with a bespoke character: the honest answer is that you are buying something you could not shoot, not a discount. Write the brief accordingly, and ask for the variant count to be priced separately from the build.

Disclosure goes in the brief, not in trafficking

EU AI Act Article 50 has been in force since 2 August 2026 and puts the duty on the advertiser and agency, not the platform. TikTok mandates an advertiser disclaimer. Meta and Google mostly apply labels automatically, which is not the same as you having complied. The IAB framework advises burning a visible disclosure into the creative where you cannot rely on platform rendering. EU guidance summary; the IAB framework.

A burnt-in disclosure is a design decision. It affects framing, end-card layout and safe areas. Decide it at brief stage, because adding it after the grade means re-opening finished shots, which is exactly the late change Katz warned about.

Name the model, for clearance reasons

Indemnification is not a property of AI generally. It exists at Google Veo on the enterprise API and GA versions only, at Runway Enterprise, at Adobe Firefly for the native model only, and at Luma Enterprise. It does not exist at Seedance or BytePlus, consumer Kling, Wan or Moonvalley.

"Made in Firefly" tells a legal team nothing, because Firefly hosts partner models on different terms. The distinction explained. If clearance matters to your risk committee, the production spec names models and versions, and the brief says which tier of protection is required. That constrains the toolset, and it should be a conscious trade rather than a surprise at legal review.

One more clause worth adding. The US Supreme Court declined Thaler v. Perlmutter on 2 March 2026, leaving purely AI-generated output without US copyright protection. Only the human creative contribution is protectable. Ask for human authorship to be documented as part of the deliverable: direction, edit decisions, design, written material, performance capture. It costs nothing at brief stage and it is the only record you will have later.

If you take one page from this, make it the first page of the brief: the character decision, the full deliverable list, the markets it runs in, and the clearance tier. Everything else in an AI brief can be discussed in the kickoff. Those four cannot.