Article 50 of the EU AI Act took effect on 2 August 2026, requiring anyone publishing synthetic images, video, audio or text to disclose that AI was involved. Two weeks in, the practical consequence is not showing up as enforcement. It is showing up in delivery contracts.

According to Numonic, the London company building provenance infrastructure for AI-generated creative work, advertisers, broadcasters and games publishers have begun asking their creative suppliers for an audit trail of the models, versions and settings behind delivered work. Most cannot produce one.

The company has completed a second close on its pre-seed round, adding Josh Stellick, a founding member of the US e-discovery and information governance company Lighthouse, as an angel investor alongside existing backer Fuel Ventures. Numonic is putting the money into the record-keeping layer that both the commercial requests and the new European rules assume already exists. Stellick’s background sits close to the problem: e-discovery exists because organisations are periodically required to produce a defensible record of what they hold, on demand, to a standard somebody else sets.

The requests are specific. Not a general assurance that AI was used responsibly, but a per-asset record: which model produced this frame, which version, which settings, and what happened to it between generation and delivery. Every one of these industries already works this way for licensed material. An agency can produce the licence for a stock image, a broadcaster can trace chain of title, a games publisher can account for every third-party asset in a build. Generated work is the gap in an otherwise well-documented supply chain.

“The studios and agencies we work with deliver into advertising, film and television, and games, and those contracts are now asking for an audit trail,” said Casey Milone, founder of Numonic. “Show us every model that went into this work. Most of them cannot answer that. The record is either lost or sitting in a spreadsheet that nobody can search and nobody would want to stake a delivery on.”

The reason is structural rather than careless. A generated asset carries its production record at the moment it is made, and then loses it. Upscaling, editing, versioning, compositing and export each strip or fail to carry the information forward, so by the time a file reaches a client it is usually just an image. The knowledge of how it was made survives, if at all, in a prompt log, a chat thread or a spreadsheet maintained by whoever happened to be at the machine.

Article 50’s scope is what makes this urgent outside the EU. The obligation follows the audience rather than the address on the letterhead, so a London agency whose campaign runs in Berlin is inside the rules, as is a series licensed to a European platform or a game selling on EU storefronts. Britain has no equivalent law of its own. There is no horizontal AI statute, the King’s Speech in May set out 37 bills without proposing one, and the government’s taskforce on labelling AI-generated content is not expected to publish even an interim report until the autumn. A further EU date follows on 2 December, when the grace period ends for marking obligations on AI systems that were already on the market in August.

“A taskforce reporting in the autumn is not much use to an agency that shipped work into Europe on the third of August,” Milone said. “But honestly, nobody is going to rebuild their pipeline because of a taskforce. They will rebuild it because a client asked a question they could not answer.”

“Every industry eventually gets asked to prove what it holds,” said Stellick. “The ones that cope are the ones that kept the record as they went. Creative work is next.”

Numonic captures assets alongside the provenance of how they were produced, holding the model, version, settings and workflow with the work itself rather than in a spreadsheet or a chat log. It writes that record into each asset as embedded IPTC metadata, so the record travels with the file, and keeps the authoritative copy in the library, where it survives the exports and re-encodes that strip embedded data in ordinary use. The company is deliberate about the limits of that. Numonic does not make anyone compliant, and embedding a record after the fact is not the same duty as the marking obligation the Act places on the tools that generate content. It addresses the prior question, which is whether a business can say what it made and how.