AI influencers for brands: what they cost, and where they break

A brand character does not fail at shot four. It fails at shot forty, when the jaw is a little wider than it was in the hero film, the hands have changed shape, and the walk cycle reads as someone else wearing the same face. Everything you need for a single beautiful frame of an AI influencer has been available for two years. Holding that frame steady across a campaign is a different problem, and it is the one that sets the budget.

So treat a brand character as a consistency problem before a design problem. The question is not whether you can generate a face the client likes. It is how many distinct shots, formats, lighting conditions and performances that face has to survive before the campaign ends, and what has to exist up front for it to survive them.

Reference capacity is the lever

The technical reason character work became viable for commercial campaigns is unglamorous: reference capacity jumped by an order of magnitude in 2026. Seedance 2.5 takes 50 multimodal references, including images, audio, 3D white models and style references, up from 12. Kling 4.0 takes 15 references and 10 keyframes. Pika H3 Flex takes 12. MiniMax H3 takes nine images plus three video and three audio. Luma UNI-1.1 takes nine.

Seedance 2.5

50 refs

Kling 4.0

15 refs

Pika H3 Flex

12 refs

That number matters more than any benchmark score, because consistency is bought with references. A character defined by one portrait will drift. A character defined by a built reference set — multiple angles, a 3D form, costume detail, a voice, a set of expressions — drifts much less, and drifts in ways you can correct.

Two capability notes worth having in the room when a client asks what is actually possible. Luma Ray3.2 is the only model publishing performance tracking of skeletal posture, gestures and facial expression for up to eight faces, which is the difference between a character standing still and a character doing something in a group. Seedance 2.5 leads the independent Artificial Analysis board on human anatomy and photorealism. Luma's announcement; the benchmark reporting.

What a character costs in time

The most useful public number on character work came out of the Super Bowl. Svedka's LX spot took about four months, and most of that was reconstructing one character — Fembot — and training the model on her. Sazerac's CMO was direct about it: the exercise was not about saving time or money. TechCrunch's account of the 2026 AI ads.

The Swissquote orange bull went the same way. It had to be rebuilt in 3D, with substantial R&D behind it, so that it could perform physically complex actions and stay itself across shots. David von Ritter of cavalcade put the standard plainly: "AI is like plastic surgery: it's good if you can't see it." The full interview, June 2026.

In both, the character absorbed the schedule. Price the character, not the shot count.

Where it breaks

This part is practitioner judgement rather than published finding. Take it as a working list from production, not a benchmark. The failure points cluster:

• Physical complexity. A character talking to camera is close to solved. A character lifting, throwing, climbing or interacting with a product in a way that needs real weight is where you end up rebuilding in 3D, as Swissquote did.
• Hands and contact. Anything where the character touches the product closely and the viewer is looking at the product. The brand asset and the weakest region of the frame are the same region.
• Extended runtime. Drift accumulates. A six-second cutdown hides it. A ninety-second film does not.
• Scene changes. New lighting, new environment and new wardrobe all pull on the same identity. Change one at a time and the reference set holds. Change three and the character becomes a lookalike.
• Second and third campaigns. A character that was never documented has to be rediscovered months later, usually by a different team, often on a newer model version.

Announcing a character is not the same as it working

Tilly Norwood is the reference point to keep in mind whenever a synthetic performer is pitched as a ready asset. The only 2026 output is a music video, "Take The Lead", released to YouTube around 10 March 2026, with vocals by Suno and a disclaimer stating it was "made by 18 real humans". A feature called "Misaligned" was announced on 6 July 2026 and is in development with no release date. Background here. A character with a press cycle and a character with a body of work are different things, and the gap between them is where client money goes.

The rules that now apply to characters

Union terms arrived in 2026 and they are specific about synthetic performers. SAG-AFTRA's four-year contract was ratified on 5 June 2026, with over 90% in favour on roughly 19% turnout. The operative AI term allows producers to use synthetic performers only where they bring "significant additional value" over a live actor or a digital capture of one. President Sean Astin said the deal "strengthens protections around artificial intelligence and digital identity." Fast Company's report. Earlier in 2026 the union was reported to be weighing a "Tilly tax" on digital performers. We could not verify that it survived into the final contract, so do not plan around it.

“A character with a press cycle and a character with a body of work are different things, and the gap between them is where client money goes.”

In the UK, Equity members voted on 18 December, 99% in favour of refusing to be digitally scanned on set, on 75% turnout of 7,746 eligible performers. The vote is indicative rather than binding. The core objection is performers' data being used to train AI without informed consent. General secretary Paul W Fleming called artificial intelligence "a generation-defining challenge". PACT made an improved offer on 27 January 2026, which Fleming said "don't cover everything we've asked for". There is no verified resolution as of November 2026. Screen Daily's coverage.

Disclosure applies directly. The IAB framework published 18 August 2026 recommends disclosure for fully AI-generated avatars mistakable for humans, and for digital twins of living people placed in fabricated scenarios. EU AI Act Article 50, in force since 2 August 2026, uses a synthetic influencer demonstrating a sponsored product as its own example of content requiring disclosure. The IAB framework; a summary of the EU guidance. If the character is the ad, the disclosure is not optional.

What audiences do with it

Disclosure does not buy goodwill on its own. Valentino's AI-generated DeVain handbag video was labelled and still drew comments calling it "disturbing," "lazy," and "AI slop". Getty's Dr Rebecca Swift made the distinction that matters: audiences accept AI in personal projects and expect more of luxury brands. The reporting.

DoubleVerify's study of 22,000 consumers, published 29 July 2026, puts numbers on both sides of it. 42% say low-quality or uncanny AI ads would negatively affect their opinion of a brand. 40% view polished, professional AI ads positively. The study. The split is not about AI. It is about execution quality, which brings it back to the consistency problem.

Own the reference set

A brand character is a reusable asset, and the thing that makes it reusable is not the finished film. It is the reference set that keeps the character stable. That set should belong to the client, written into the contract before work starts, so using the character again does not depend on staying with one studio.

Four lines to settle before the first frame is generated: who owns the reference set and in what format it is delivered; which model versions the character was built against; what the character is licensed to do and where; and who holds the disclosure obligation on each placement. Settle those and the character is an asset on the balance sheet. Skip them and it is a campaign that cannot be extended.