Everyone Heard About the Delay. The Deadline Is Still Sunday.

In May, Brussels handed AI operators what sounded like a reprieve. The Digital Omnibus agreement — provisionally reached on May 6, 2026 and confirmed by member states a week later — postponed the EU AI Act's high-risk compliance deadlines. Annex III stand-alone systems moved from August 2, 2026 to December 2, 2027. High-risk AI embedded in regulated products slid to August 2, 2028 (Gibson Dunn).

Most teams heard "delay," closed the tab, and reallocated the budget.

Here's what didn't move: this Sunday, August 2, 2026, the Act's Article 50 transparency obligations become applicable. Violations carry fines of up to €15 million or 3% of total worldwide annual turnover, whichever is higher (Jones Walker). The delay everyone celebrated was for the compliance program most operators hadn't started. The deadline nobody talks about is for the systems you're running right now.

What actually lands on Sunday

Article 50 is short, and it is aimed squarely at systems that interact with people and generate content — which is to say, agents. From August 2:

• Interaction disclosure. People must be told they're interacting with an AI system, unless it's already obvious to a reasonably well-informed person. Your support agent, your intake bot, your outbound SDR agent: disclosed, every channel.
• Synthetic content marking. Systems generating audio, image, video, or text must mark outputs as artificially generated, in a machine-readable way, to the extent technically feasible.
• Deepfake and public-interest text disclosure. AI-generated or manipulated media must be labeled, and AI-written text published on matters of public interest must be disclosed unless it went through human editorial control.
• Emotion recognition and biometric categorization notices. If you run these, the people exposed to them must be informed.

There is one grace period worth knowing: providers of content-generating systems already on the market before August 2 have until December 2, 2026 to implement technical marking. It applies to the marking machinery — not to the disclosure duties, and not to anything you ship after Sunday.

The awkward part: you can't disclose for agents you can't list

Transparency obligations attach per system, per interaction. That's a problem when you don't have a reliable roster of systems. An appliedAI analysis of 106 enterprise AI systems found that over half of organizations lack systematic AI inventories, and 40% of the systems studied couldn't be clearly classified under the Act's risk tiers (Cloud Security Alliance).

Read those two numbers together and the delay looks less comforting. The postponement bought time on high-risk conformity assessments — the heavyweight paperwork. It bought zero time on knowing what agents you run, which of them talk to humans or generate content, and whether each one discloses. That's not a 2027 problem. That's a Sunday problem.

Enforcement won't be a dawn raid. It'll be a documentation request.

Nobody's regulator is kicking down doors Monday morning. What week one actually looks like is quieter and more dangerous: a complaint from a user who wasn't told they were talking to a bot, a competitor's report, a market-surveillance authority asking you to demonstrate compliance. The omnibus also strengthened the enforcement side while it softened the timeline — the AI Office now holds investigative powers, inspection authority, and its own fining capability (Gibson Dunn).

When the request comes, "our agents disclose" is a claim. What the authority wants is evidence: this agent, this interaction, this disclosure, this timestamp. If your answer lives in a config file that changed four times since the interaction in question, you don't have an answer. You have a story.

What this maps to in VeriSwarm

We built for exactly this shape of problem — obligations that attach to every interaction, and auditors who want receipts.

Gate gives you the roster. Every agent registered, scored, and visible — including the ones a product team spun up in March and forgot. An inventory isn't a spreadsheet someone updates quarterly; it's the always-on layer your agents already pass through. Gate is free, which means the "we don't know what we're running" problem costs exactly nothing to fix before Sunday.

Guard sits in the output path. The Guard Proxy intercepts what your agents actually say and generate — which is where marking and disclosure either happen or don't. Policy rules can enforce that synthetic content carries its marking and that interactive sessions carry their disclosure, at the proxy, regardless of what the agent's own code remembered to do. The same layer that tokenizes PII before it leaks is the right place to guarantee a label before it ships.

Vault is the receipts. Every disclosure event, every marking decision, every policy enforcement lands in an immutable, hash-chained audit ledger. When the documentation request arrives, you export the chain and let the math vouch for you. Cryptographic evidence that the control existed and ran — not a declarative promise that it probably did.

The takeaway

The omnibus delay was real, and if you build Annex III systems, you got eighteen months of breathing room on conformity assessments. But the transparency layer of the AI Act arrives on schedule, this weekend, with a €15 million ceiling attached — and it targets the most ordinary things agents do: talking to people and generating content.

Don't confuse the deadline that moved with the one that didn't.

If your agents will be talking to Europeans on Monday, start with Gate — the inventory is free, and the disclosure evidence is a proxy away.