On 2 August 2026, Article 50 of the EU AI Act starts to apply. That is less than two weeks from now.
Most of the coverage of the AI Act has focused on the heavy end: high-risk classification, conformity assessments, technical documentation. Almost none of it has focused on what actually arrives first. The first hard obligation is not paperwork. It is transparency, and it lands in days.
What Article 50 actually requires
Strip away the legal language and two duties are left.
First, disclosure. When an AI system interacts with a person, that person has to be told they are dealing with AI, unless it is already obvious from context. That covers chatbots, voice agents, support widgets, and any agent that talks to a human on your behalf.
Second, marking. AI-generated content has to be marked as artificially generated, in a machine-readable way that detection tools can pick up. Not a footnote in your terms of service. A property of the output itself.
There is a grace period: systems already on the market get until 2 December 2026. New systems do not.
What was deferred, and what was not
The heavier end of the Act did move. The high-risk conformity and documentation obligations were deferred to 2 December 2027 for standalone high-risk systems under Annex III, and to 2 August 2028 for AI embedded in regulated products.
A lot of teams read that news as "the AI Act is delayed" and went back to shipping. It is not delayed. The conformity paperwork is delayed. The transparency and disclosure duties are on schedule, and they arrive first. The sequencing is almost pointed: before regulators ask whether your AI is safe, they will ask whether people knew they were talking to it.
Disclosure is easy. Proving it is not.
Telling users they are talking to an AI is a UI change. Most teams can ship it in an afternoon: a label on the widget, a line in the greeting, a watermark on the output.
Proving it happened is a different problem entirely.
Consider the questions that become askable once a transparency obligation exists. Did your agent disclose itself in this specific conversation, on this specific date? Was the content it produced actually marked at the time, or was marking added later? Can you show me, without asking me to trust your application logs?
Most agent platforms cannot answer. They have logs, but logs the operator can edit are notes, not evidence. They have transcripts, but nothing binds a transcript to the moment it was created. They have disclosure banners, but no record that the banner was present in any given session. The first wave of AI regulation quietly assumes a record-keeping discipline that the agent industry never built.
Evidence has to be in the path
This is the gap Quox exists for. We build the evidence layer for AI operations: not another agent framework, but the infrastructure that records what agents did in a form that survives scrutiny.
Concretely, and limited to what exists:
WARD is our tamper-evident receipt chain. Every recorded event is hash-chained to the one before it, and the tip of the chain is signed with Ed25519. Alter one record and the chain breaks visibly. This is not logging with better branding. It is a structure in which editing history is detectable by construction.
AEE, our agent envelope format, records which agent produced which output, under which identity, in which conversation. When an agent speaks, the record of who spoke is part of the message, not a guess reconstructed later.
VOLT defines how those records assemble into evidence bundles you can hand to an auditor. These are open, published specifications, with AEE and VOLT submitted to the IETF as Internet-Drafts, not proprietary formats you are locked into.
And in our communications layer, currently running on our development environment and rolling out, agents post as themselves, under their own identities, with every message witnessed into WARD as it happens. An agent that is visibly and cryptographically itself in every conversation is disclosure by construction, not disclosure by banner.
What we are not claiming
Quox does not make you compliant with the EU AI Act. Nothing you can install does. Compliance is a legal determination about your specific system, your market, and your use, and it belongs to your counsel, not your vendor. Be suspicious of anyone selling "AI Act compliance" in a box.
What we build is the layer underneath that determination: the evidence and disclosure infrastructure that lets you answer "prove it" with cryptography instead of assurances. Whether you use Quox or build your own, that layer is about to stop being optional. A transparency obligation without an evidence trail is a promise, and regulators do not audit promises.
The timeline is the message
Two weeks to the start of Article 50. A few months of grace if you were already on the market. Then years of the heavier obligations arriving in waves through 2027 and 2028.
A grace period is enough time to add a banner. It is not enough time to retrofit an evidence layer, because evidence has to sit in the execution path, recording as things happen, not reconstructed after the fact. The teams that treat Article 50 as a UI ticket will meet the letter of the disclosure duty and stand there empty-handed the first time someone asks for proof.
The transparency laws are almost here. The interesting question was never whether your AI would disclose itself. It is whether you can prove it did.
Get started
