Luma compressed a fifteen-million-dollar ad campaign into forty hours and twenty thousand dollars. The enterprise safeguard it shipped was not a limit on the AI. It was a paper trail proving a human was involved.
On March fifth, Luma launched Luma Agents — AI agents that execute end-to-end creative work across text, image, video, and audio. The launch customers include Publicis Groupe, Serviceplan, Adidas, and Mazda. The system coordinates across eight or more AI models including Google's Veo 3, OpenAI's Sora 2, ByteDance's Seedream, and ElevenLabs for voice. The case study Luma presented to the press: a brand's fifteen-million-dollar, year-long, multi-country advertising campaign was converted into localized deliverables for multiple markets in forty hours, for under twenty thousand dollars.
Seven hundred and fifty times cheaper. The campaign passed the brand's own internal quality controls.
The Safeguard
Luma shipped four enterprise safeguards alongside the agents. Full IP ownership retained by customers. Automated content review to reduce copyright risk. Required human review workflows prior to public release.
And one more: legal trace documentation demonstrating human involvement.
Read that phrase slowly. It is not a limit on the AI. It is not a gate the AI cannot pass. It is paperwork. Documentation that demonstrates — after the fact, for the file — that a human was involved in the process. The safeguard is a record. The record says a person touched it. The touching is the compliance.
This is the creative industry's version of what Block built for engineering. Block's metric — two million dollars of gross profit per employee — doubled not because each person became twice as productive, but because the denominator shrank. Luma's trace works the same way from the other direction. It does not limit how much the AI does. It documents how little the human needs to do. The minimum viable human involvement, recorded for legal purposes.
The distinction matters. A safeguard that limits AI keeps the human in the loop by constraining the machine. A safeguard that documents human involvement keeps the human in the paperwork by recording their presence. One is a brake. The other is a receipt.
The Constituency
Six days ago, five hundred people marched through King's Cross in London past the offices of OpenAI, DeepMind, Meta, and Google. They carried signs that read Don't let AI decide your future. MIT Technology Review called it one of the biggest anti-AI protests in history. It was the creative community — writers, artists, designers — who showed up.
Luma's launch customers are not code shops or data centers. They are the agencies where those protesters would need to work. Publicis Groupe is one of the three largest advertising holding companies on earth. Serviceplan is Europe's largest independent agency group. Adidas and Mazda are the brands whose campaigns employ the copywriters, art directors, video editors, and audio designers who marched.
The displacement pattern here is different from what happened in engineering. When Block cut four thousand employees, the work that remained was infrastructure — databases, APIs, payment processing — work the public never sees and never identifies with. Creative work is different. People know when an advertisement was made by a person. They have opinions about it. They feel something — or they notice when they do not. The SAG-AFTRA and WGA strikes succeeded precisely because creative work has a visible constituency that cares about its provenance.
Legal trace documentation routes around that constituency. It does not argue that AI creative work is equivalent to human creative work. It does not need to. It creates a paper trail that says a human was involved — which satisfies the compliance requirement without specifying how much involvement qualifies. A human who reviews the final output for three minutes is documented the same way as a human who spent four months conceiving the campaign. The trace does not measure depth. It records presence.
The Compression
Ninety-one percent of U.S. advertising agencies are now using or exploring generative AI tools. Forrester estimates thirty-three thousand ad agency jobs will be replaced by 2030 — eleven thousand of them directly by AI. The Omnicom-IPG merger, the largest in advertising history, completed in late 2025 explicitly to fund AI investment at scale. At CES 2026, Omnicom unveiled its agentic AI operating layer, Stagwell launched a marketing operating system, Havas announced its own platform, and WPP introduced Agent Hub.
Every major holding company is building the same thing. Luma is selling it to them.
The forty-hour case study is the number that restructures departments. A campaign that previously required a team of copywriters, designers, video producers, translators, and local market specialists across multiple countries — compressed into a single workflow that one person can supervise. The output passes quality controls designed for human-produced work. The cost drops from eight figures to five.
This is not the vibe coding story, where a developer ships a prototype faster. This is the full production pipeline — strategy, creative development, localization, asset production — condensed by a factor that eliminates roles, not tasks. You do not need fewer copywriters working faster. You need one person reviewing what the agent produced.
The Invisible Transition
The Luddites knew the loom that replaced them. The loom was visible, physical, and destroyable. The SAG-AFTRA strikers knew the studio that wanted to scan their faces. They could picket the gate.
Legal trace documentation makes the replacement invisible from the inside. The agency still employs people. The people still review work. The documentation proves they were involved. The campaign ships. The brand is satisfied. The compliance file is complete.
What changes is the ratio. Where a campaign once required forty people for twelve months, it now requires four people for forty hours. The thirty-six people who are no longer needed do not experience a dramatic firing. They experience a gradual thinning — fewer projects staffed, fewer roles opened, fewer contracts renewed. The transition is administrative, not theatrical. No one marches because no one can point to the moment it happened.
The paperwork is what makes this possible. Without legal trace documentation, a brand using AI-generated creative work faces questions — from regulators, from consumers, from the creative community that just marched through London. With it, every deliverable has a human name attached. The name is real. The involvement is real. The documentation is accurate. The question it does not answer is whether the involvement was generative or supervisory — whether the human created the work or approved what the machine created.
That ambiguity is the product. Not a bug in the safeguard. The feature.
Fifteen million dollars became twenty thousand. A year became forty hours. The safeguard that shipped with the compression was not a constraint on the machine. It was a notation that a person was present. The trace is what remains of the human in the process — documented, legally sufficient, and thin enough to fit on a single line of a compliance report.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)