Roughly one in five employees at Luma AI is a working creative. An artist, a filmmaker, a designer. The company is valued at close to $4 billion, and it sells AI video models.
That staffing choice tells you more about where creative work is going than any model benchmark.
The pitch agencies did not want to hear
A piece in Forbes this week laid out the model. Luma embeds its own artists and filmmakers with brand and studio teams. It calls them forward-deployed creatives, borrowing the idea from enterprise AI labs that place engineers directly with customers.
CEO Amit Jain's message to brands is blunt. You know your brand best, so do not outsource it. Build a team in-house and work with Luma to produce at global scale. He says some large legacy brands are already making "an end run around their agencies" and asking Luma directly to help them build a team.
Luma's partners include Coca-Cola, Adidas, Visa, Mazda and Lufthansa. It works with agencies too, including Publicis and Serviceplan, but selectively. Jain's view is that agencies "can slow you down."
A second company, the same shape
Infinite Studios, co-founded by Adrienne Lahens, who helped build TikTok's Creator Marketplace, runs a similar play. Its forward-deployed creatives are "a hybrid of a creative and an engineer." They learn an organisation's goals and how it measures results, then tailor the tools to fit.
One number from Lahens stuck with me. A typical creator video goes through about five brand team reviews, often by hand and over email. She estimates roughly a third of those videos never go live because they need small fixes, a logo correction, a more visible label.
A third of paid creative work dies in the approval process. Not because it was bad. Because the fix took longer than the window.
That is not a creativity problem. It is a systems problem, and systems problems are where AI tools win first.
What is actually being unbundled
I run an agency, so I read this piece with more than academic interest. My honest take is that the threat is real but narrower than the headline suggests.
What gets unbundled is production. Making thirty variations of the same asset, resizing for every placement, adapting a hero visual for twelve markets.
That work used to justify retainers and headcount. Jain says it directly: humans are not good at producing the same thing thirty times. Machines are.
What does not get unbundled is judgment. Which idea is worth producing at all. Which of the thirty variations is on brand. How the creative connects to the media plan and the commercial target.
Lahens herself warns that pressing a button for cheap outputs is not efficiency. Efficiency comes from building the right infrastructure.
Building the right infrastructure is exactly what a good agency should sell. The problem is that many agencies still price by the hour of production, which is the part that just got cheap.
The agency that survives looks different
I wrote about a related shift in three signals the expert builds the tool. The people who understand the craft are now building the systems, not just using them. Luma and Infinite are that pattern at company scale.
For agencies, the move is to become the forward-deployed team. Sit inside the client's workflow. Build the prompt libraries, the brand guardrails, the approval logic and the measurement loop. Then hand the client a system their in-house team can run, and stay on as the people who keep improving it.
That is what we have been rebuilding toward at difrnt. Fewer hours sold for production, more work on the architecture that makes in-house teams faster. It is a harder sale, because the client has to understand they are buying a capability rather than a deliverable. It is also the only version of the agency model I believe holds up over the next five years.
Lahens has a line that sums it up: campaigns are dead, and brands need to think as a content engine and a modular storytelling system. An engine needs engineers. Someone has to design it.
What CMOs should do with this
If you are a CMO, the useful question is not "agency or in-house." It is which parts of the creative process need deep brand knowledge, and which parts are repetitive production.
Keep the first part close. That means in-house people who live with the brand every day, or an agency team embedded deeply enough to behave like one.
Automate the second part, and be suspicious of anyone, agency or vendor, still billing it by the hour. If a partner cannot show you how AI tools changed their cost structure in the last year, you are paying for the old one.
Then fix the approval chain. If a third of your creator content dies in review, the cheapest win in your entire marketing budget is not a new tool. It is fewer reviewers with clearer rules.
The question is no longer who makes the asset. It is who designs the system that makes them.