AI World-Building for Agencies: Brand Worlds That Outlive the Campaign

In March, Campaign US gave its Agency of the Year award in a category that did not previously exist: AI Creative Studio. It went to BarkleyOKRP, whose CEO had run an agency-wide AI upskilling programme and whose SVP of creative and AI had a job title that would have read as a typo three years earlier.

Look past the award at the work, though, because the work is where the actual lesson is.

One of the campaigns was for Slice. It was not a spot. It was a 1990s radio station: forty original songs, synthetic DJ hosts with personalities, a full month of live FM broadcast in Los Angeles. Built on Gemini, Imagen, and Veo. Reported results: 120 million impressions at a CPM 60% more efficient than benchmark.

Another, for the Salvation Army, generated ads automatically triggered by thrift-store item sales, driving store visits at 82% below category cost benchmarks.

Notice the shape of the first one especially. The deliverable was not an asset. It was a world with rules and inhabitants, and a month of material came out of it.

That is the shift. Not faster ads. A different unit of work.

The pitch room changed first

The oldest version of this argument is about pitches, and it is the easiest one to see.

A mood board describes a world. It says: imagine a city like this, a character like her, a palette in this direction. Everyone in the room nods and privately imagines something slightly different, which is why the first round of comments is always about things nobody said out loud.

A living world shows it. The client meets the campaign character from every angle. Walks through the location. Sees the palette applied to three scenarios they had not asked about. The imagining stops being a group hallucination and starts being a shared reference.

The commercial consequence is what matters: approval cycles shorten and revision rounds drop, because the ambiguity that generates revisions has been removed before the meeting rather than discovered after it.

The economics of building once

Here is where the model breaks from traditional production economics in a way worth being precise about.

Traditional campaign work is linear in assets. The hero film costs what it costs. The social cuts cost more. The Q4 refresh costs again, and it costs more than it should, because half the budget goes to re-establishing what the Q1 work already established: casting continuity, location matching, grade matching, brand-guideline enforcement across three vendors.

World-building work is front-loaded then marginal. You spend real effort once, defining the characters, locations, palette, and rules with enough precision that they are enforceable. Every subsequent piece of material inherits that definition. The Q4 refresh does not re-establish anything. It shoots a new scene in an existing world.

The asset stops being the file. The asset is the world, and the files come out of it all year.

That has a direct consequence for how agencies charge. A file is a deliverable, billed once and depreciating from the moment it ships. A brand world is a retained asset with a maintenance relationship attached. Agencies moving first on this are treating it accordingly, and it changes what a retainer is for.

What "on-model forever" actually requires

It is easy to say a brand world stays consistent. Making it true requires solving the same problem filmmakers hit, which is that models do not remember.

Every major video model now ships strong reference features. Seedance 2.5 accepts up to thirty image references. Kling 3.0 offers subject reference and multi-shot continuity. Veo 3.1 does reference-to-video. All genuinely useful, and all of them operate within a single generation. Close the session and the model retains nothing.

For a campaign running from Q1 to Q4, across multiple teams and probably multiple vendors, that is the whole problem. The brand character will drift, not because anyone was careless, but because six people re-supplied references six slightly different ways over nine months.

What closes it, practically:

Character bibles with multi-view sheets. One canvas, several angles, defined lighting conditions. Not one flattering hero image, which gives the model too much freedom everywhere it cannot see.

Numerically locked palettes. Hex values, named light sources, specified grade. "Warm and premium" drifts across a quarter. Precise values drift far less.

A written world bible that outranks any brief. When a new team joins in Q3, the bible is what they inherit. If the canon only exists in the head of the creative director who built it, the world ends when they roll off the account.

A continuity layer that enforces it automatically. This is the structural version. On Magic Lantern, the AI Showrunner holds the world's canon and applies it to every scene generated inside that world, across sessions, teams, and quarters. The consistency stops depending on whether everyone remembered to attach the right reference folder.

Settle the rights questions before the pitch, not after

This is the part agencies skip and then regret. The 2026 legal position is clear enough to plan around, and it has moved.

Copyright. The US Copyright Office position is unchanged: purely AI-generated works are not registrable, human contributions are registrable case by case, and prompts alone are insufficient authorship. In March 2026 the Supreme Court declined to take up the question, which leaves that position standing. Practically: a brand world generated entirely by prompt may not be protectable as such. The human creative direction, arrangement, and selection are where the registrable authorship lives, and it is worth documenting that work as you do it rather than reconstructing it later.

Performer likeness. SAG-AFTRA's 2025 Commercials Contract added digital-replica consent provisions and a standardised Digital Replica Rider. The 2026 TV/Theatrical agreement carries a set of AI provisions including terms covering synthetic performers. If your brand world has a recurring human-looking character, decide early whether that character is a synthetic original or a digital replica of a real performer, because the contractual paths are completely different.

Training provenance. Netflix paid $587 million for InterPositive, a company whose models train only on consented, production-owned footage. That design choice was not decoration; it is a meaningful part of what made the company acquirable. The same logic scales down. Clients in regulated categories will ask what your world was trained on, and "the internet" is an answer that ends conversations.

Ownership of the world itself. If the world is a retained asset, the contract should say who owns it, what happens at the end of the relationship, and whether the client can take it to another agency. This is a new question and most master service agreements do not currently answer it.

Where to start

If you want to run this on one account before committing:

Pick a brand with a recurring character or a distinctive place, because those are where continuity pays fastest. Build the world properly: multi-view character sheet, location plates, numeric palette, written rules. Produce the pitch material inside it. Then, critically, produce the second wave inside it too, three months later, and measure what you did not have to redo.

That second wave is the whole argument. Everything before it is a nice demo.

What agencies are asking

Are agencies actually using AI world-building for client work?

Yes. Campaign US created an AI Creative Studio Agency of the Year category for 2026, awarded to BarkleyOKRP in March. Campaigns like Slice's "The Fizz FM," a fully built 1990s radio world with original music and synthetic hosts that ran as a live LA broadcast, reported 120 million impressions at a 60% more efficient CPM.

What is a brand world?

A persistent universe for a brand, containing defined characters, locations, palette, and visual rules. Instead of producing campaign assets one at a time, teams define the world once and generate every subsequent piece of material inside it, so continuity is inherited rather than re-established.

Can we own AI-generated brand assets?

Partially, and it needs care. The US Copyright Office holds that purely AI-generated works are not registrable and prompts alone do not constitute authorship; the Supreme Court declined to revisit this in March 2026. Human creative direction, selection, and arrangement are where registrable authorship sits, so document that contribution as you work.

How do we keep a brand character on-model across a year?

Build a multi-view character sheet rather than a single hero image, lock the palette to specific values, maintain a written world bible that any new team inherits, and use a continuity layer that enforces canon automatically rather than relying on each person attaching the right references.

What about performer likeness and union rules?

SAG-AFTRA's 2025 Commercials Contract introduced digital-replica consent terms and a standardised Digital Replica Rider, and the 2026 TV/Theatrical agreement includes provisions covering synthetic performers. Decide early whether a recurring character is a synthetic original or a replica of a real performer, as the contractual requirements differ substantially.

Build the world once. Shoot inside it all year.

Magic Lantern is an AI cinematic storytelling platform used by creative agencies and brand studios alongside independent filmmakers. The AI Showrunner keeps every asset on-model for as long as the campaign runs.

Build yours at magiclantern.io.