On August 2, AI-Generated Content Gets a Legal Label. Your Client Documents Are Already Wearing It.

Three weeks from now, on August 2, 2026, Article 50 of the EU AI Act comes into force — the transparency layer of the world's first comprehensive AI law. From that date, chatbots have to identify themselves as machines, deepfakes have to be labelled, AI-generated text published on matters of public interest has to be disclosed, and — the clause that matters most for anyone who sends documents for a living — providers of generative AI systems have to mark their outputs as artificially generated in a machine-readable format. Penalties run to €15 million or 3% of worldwide turnover. The European Commission published its draft guidelines on May 8, a voluntary Code of Practice on June 10, and the May 2026 AI Omnibus agreement gave systems already on the market a grace period to December 2 for the marking requirement — which is regulatory language for this is happening, the only question is the month.

Most coverage of the deadline is about chatbots and deepfakes. The part almost nobody in client-facing work has thought through is quieter: the mark travels with the content. And your proposals, decks, and board packs are the content.

The label doesn't ask whether you have a duty. It ships in the file.

Read Article 50 as counsel and you might exhale: the deployer-side disclosure duty for AI-generated text is narrow — it covers text "published with the purpose of informing the public on matters of public interest." A proposal to one client, a board pack to twelve directors, an investor update to a cap table — most of it arguably sits outside that clause. If the legal question is "must I stamp DISCLOSURE on slide one?", the answer for most client documents is probably no.

But that's the wrong question, because the marking obligation in Article 50(2) doesn't sit with you. It sits with the provider — the model and tool vendors — and they comply at the point of generation, not the point of use. The infrastructure is already live ahead of the deadline: as of 2026, the largest commercial image and video models label their output by default, C2PA Content Credentials verification has shipped into Adobe workflows, OpenAI outputs, and Google's Gemini, Search, and Chrome surfaces, and TikTok alone has labelled over 1.3 billion videos with provenance credentials at consumer scale. Google Photos now shows provenance data to anyone who taps the details of an image.

Which means the practical effect of August 2 is not a duty you must perform. It's a property your documents acquire: detectability by default. The AI-generated hero image on your cover slide, the synthetic product render on page seven, the generated chart — each arrives carrying a machine-readable receipt, written by a vendor complying with a law that fines 3% of global turnover for forgetting. Your reader's browser, their document platform, and — increasingly — their own screening AI can read that receipt without asking your permission.

Your buyer's tools will read the mark before your buyer does

This lands on an evaluation process that already went machine-first. We've written before that 82% of VC firms screen deals with AI and that Gartner projects 90% of B2B buying will be AI-agent intermediated by 2028 — the deck's first reader is increasingly software. Provenance metadata is exactly the kind of structured, verifiable signal that machine readers are built to consume. A buying agent that today checks whether your numbers reconcile across pages will, trivially, also surface "7 of 22 assets in this document are marked AI-generated."

And the human reading over the agent's shoulder has a documented reflex about that. Buyers are roughly four times more likely to trust a brand less when they spot synthetic content passed off as authentic, and 81% of B2B buyers now require brand trust before they'll even take a sales conversation. Note the precise shape of the risk: the damage isn't in the AI use — with Gartner projecting 70% of enterprise presentations AI-generated by 2026, your client's own decks are made the same way. The damage is in the mismatch: content that presents as authored and verified, revealed by its own metadata to be unreviewed machine output. The label doesn't punish using AI. It punishes pretending you didn't — and after August 2, the pretence has a technical failure mode.

The regulation's escape hatch is a named human who reviewed it

Here is the detail worth building a workflow around. Article 50(4) contains an explicit exception: the disclosure obligation for AI-generated text lifts where the content "has undergone a process of human review or editorial control and a natural or legal person holds editorial responsibility" for its publication.

Sit with what the regulators just encoded. Faced with the question "when does AI-generated content stop being a transparency risk?", the EU's answer was not "when the model is better" or "when the prompt was careful." It was: when a human reviewed it and someone put their name on it. That is the same conclusion every agency, consultancy, and IR team that ships high-stakes documents has always operated on — the work leaves the building when a senior has read every page and is willing to own it. The AI Act didn't invent a new standard for trustworthy content. It wrote the old one into law and aimed it at machine output.

Which reframes the August 2 deadline entirely. The firms scrambling are the ones whose honest answer to "did a human review this?" is sometimes, informally, when there was time. The firms with nothing to fear are the ones for whom review-and-sign-off is the workflow, not the aspiration — because their compliance posture and their quality posture are the same sentence.

What to have in place before the label arrives

Know which of your assets carry marks. Every image, video, and text block generated by a commercial model after August 2 (December 2 at the outside) should be presumed to carry machine-readable provenance. If you don't know which parts of your standard proposal template are synthetic, your reader's tooling soon will. Inventory it now, on your schedule, rather than during a procurement call on theirs.

Move generation somewhere it's grounded. The trust penalty attaches to content that's synthetic and unmoored — the stock-average claim, the invented statistic, the render of a product that doesn't exist. Generation grounded in your firm's own knowledge — your real case studies, the client's actual brief, numbers that trace to a source — produces documents whose provenance mark is a manufacturing detail, not a confession. In Lurio, that grounding is the default: the deck is drafted from the knowledge in your workspace, so every claim has somewhere to point.

Make the review the record. The Article 50 exception asks for "a process of human review or editorial control" — a process, not a vibe. A review layer that checks every page before it ships — Data Integrity on whether the numbers reconcile, Brand Compliance on whether the work is unmistakably yours, Audience Fit on whether it lands for this reader — with critiques cited back to your own knowledge, followed by a human sign-off, is that process, on the record. When the AI schedule in the next MSA asks "who reviews AI output on our account?", the answer is a name and a workflow, not a pause.

Get your disclosure sentence ready. The strongest position on August 3 is not "we don't use AI" — nobody believes it and the metadata contradicts it. It's one true sentence: AI drafts our documents grounded in our own verified knowledge; every page is reviewed against that knowledge before a named person signs off and sends. That sentence satisfies the spirit of Article 50, the letter of its exception, and — not coincidentally — the client.

The label was always coming. The question was your relationship to it.

Every previous transparency wave worked the same way: nutrition labels didn't end processed food, they ended pretending; HTTPS warnings didn't end the web, they ended casual insecurity. Provenance marks won't end AI-made documents — 70% of enterprise presentations, remember. They'll end the free ride for unreviewed ones, by making the difference legible to every reader and every reader's machine.

That splits the market in two. For firms shipping raw model output with the serial numbers filed off, August 2 starts a countdown. For firms whose AI-assisted work is grounded, reviewed, and signed by a human who read every page, the label is the cheapest credibility they've ever acquired — a machine-readable receipt that says made with AI, checked by us, attached to work that survives the check.

The law just made your documents testify about how they were made. Ship documents with nothing to hide.

— The Lurio Team