The character outlives the campaign. Can AI keep up?

The most expensive mistake a brand can make with an AI-generated character might not happen in its first commercial. It might happen in the second.

The first film looks wonderful. The character has a face, a voice, a way of moving. Audiences recognise it. Then a new brief arrives, another production partner joins, the tools change and the character returns looking almost, but not quite, like itself. Its expression is different. Its humour has disappeared. Nobody can explain which version is definitive.

We talk endlessly about how quickly generative AI can create a new image. We talk far less about the creative responsibility of making sure that image still means something when it comes back.

The challenge is no longer just making a character. It is keeping a character.

For agencies and brands considering AI-native production, I think the real test is moving from a compelling debut to a recognisable second appearance. The challenge is no longer just making a character. It is keeping a character.

A mascot is not a moodboard

This distinction matters because character-based advertising has always gained power from familiarity. A distinctive face or behaviour can become shorthand for a brand. The IPA’s work on “fluent devices” underlines the value of recurring characters and motifs for recognition and emotional response. That is an argument for continuity, not merely repetition.

Now consider Ziggo’s Octopus, introduced in July through a campaign from Brothers And Sisters and Not Just Any. The character was developed using an AI-powered production process and launched not only in a television film but also in digital out-of-home and print. Crucially, the team described the octopus as a potential long-term piece of brand intellectual property, capable of returning in future work.

An octopus intended to represent a brand for years needs rules...

That ambition changes the brief. An octopus made for one 20-second commercial needs to work beautifully once. An octopus intended to represent a brand for years needs rules: what makes its movement unmistakable, which expressions fit its personality, what it should never do, and how it behaves when the medium changes. The first project delivers a character. The next project tests whether a character has really been created.

The identity lives in the performance

Swissquote offers a different version of the same problem. Its familiar orange bull was reimagined by Made by Humans, Cavalcade and director James Lovick for four generative-AI-led films. As shots reported, the production team worked to establish a consistent visual and behavioural language across settings and storylines.

That word, behavioural, is more interesting to me than photorealistic. A character can keep the same colour and silhouette while becoming unrecognisable in the details: how it enters a room, how quickly it reacts, whether it plays a joke or takes one seriously. Viewers may not articulate those differences, but they can feel them.

Petbarn’s Petey and Bernadette make the point from another direction. Working with Howatson+Company, the Australian retailer brought its dog and cat characters into a series of short films using a bespoke mix of 3D modelling, 2D animation, VFX and generative tools. The technology is hybrid, but the creative requirement is wonderfully old-fashioned: the animals must remain themselves from story to story.

And this is not an exclusively synthetic issue. Zapier brought its recurring human character Al back from its Super Bowl advertising into another workplace film, created with Strange Family. The performance is the connective tissue. Whatever technology makes the next instalment possible, its job is to preserve that recognition rather than dilute it.

Every new production inherits the last one

I would argue that brands now need something closer to a working character bible than an expanding folder of prompts. Not a static PDF that nobody opens, but a shared creative reference that travels with the IP: approved design and proportions; emotional range; body language; voice and performance boundaries; narrative history; the ways a character can change; and the details that must remain fixed.

It should also record the less glamorous details: who approved each version, where source artwork lives, which collaborators contributed, what likeness or voice permissions exist, and which uses are actually authorised. The World Federation of Advertisers has issued guidance on intellectual-property risks associated with generative AI in marketing. That matters especially when a character is being developed for life beyond a single piece of content.

In practical terms, the bible becomes a test. A producer can ask not only, “Does this frame look good?” but, “Would this character behave this way?” A marketer can ask not only, “Can we make ten versions?” but, “Do all ten still belong to the same world?” And a legal or licensing team can see what is documented instead of reconstructing the history after a campaign has already travelled.