The August 2 Deadline Is Not a Disclosure Problem. It Is a Documentation Problem.

By Andreas Moekesch

On August 2, 2026, Article 50 of the EU AI Act becomes enforceable. From that day, any AI system that interacts directly with people in the EU must make sure users know they are dealing with a machine, no later than the first interaction. The disclosure cannot hide in the terms and conditions. It has to be perceivable in the conversation itself.

The penalty for getting it wrong runs up to 15 million euros or 3 percent of worldwide annual turnover, whichever is higher.

If you have heard that the EU delayed the AI Act, you heard half the story. The Digital Omnibus, signed on July 8, 2026, postponed the high-risk obligations to December 2027 and beyond. The transparency rules in Article 50 were left exactly where they were. This deadline stands.

Most coverage treats this as a UX story: add the sentence, ship it, move on. And the sentence genuinely is the easy part. A competent team can add "You are chatting with an AI assistant" to a chat window in less time than it takes to read the regulation.

The uncomfortable question comes one step earlier. Add it where, exactly?

You cannot disclose an AI touchpoint you have not found.

Where the chatbots actually live

In 2023, the answer was simple. There was one chat widget, one vendor, one contract, and everyone knew about it.

In 2026, AI talks to your customers from more places than any single person in your company can name. The widget on the website, yes. But also the voice bot in the support line, the assistant inside the mobile app, the AI drafting replies in the shared inbox, the scheduling agent in the booking flow, and the AI features your vendors quietly switched on inside tools your teams use every day. Some of these were bought deliberately. Some arrived in a product update nobody read. Marketing bought one, support bought another, and someone in operations enabled a beta because it looked useful, which it was.

This pattern should feel familiar. It is the same way companies lose track of their processes: knowledge scattered across teams, no single owner, no map. Enterprises used to meet their undocumented processes for the first time during an audit. Many are about to meet their undocumented AI the same way.

How the legal duties split between you and your vendors depends on your setup, and that is a question for counsel. Knowing where those duties could land is a question only your own documentation can answer, and no law firm can build that list for you.

A practical inventory, before the deadline

The good news: a usable first inventory takes about a week, not a quarter, if you work by channel instead of by system.

Start with channels, not tools. List every way a customer or user can interact with your organization: web, app, phone, email, chat, messaging platforms, kiosks. For each channel, ask the owning team exactly one question: what answers automatically? You are not asking for an architecture diagram. You are asking who, or what, replies when a human writes or calls.

Then pull the vendor angle. Go through the tools your customer-facing teams use and check which ones have AI features enabled, because the touchpoint you will miss is the one a vendor added for you. Your CRM, your help desk, your booking system, and your phone platform have all shipped AI capabilities in the past two years, and many are on by default.

Finally, write it down as a living document with a named owner per touchpoint: where it is, what it does, whether a person could mistake it for a human, and what the disclosure currently says. That last column will be empty more often than anyone expects. Now it is a work list, sorted by risk, with a deadline attached.

The deadline behind the deadline

Article 50 is not the last word from Brussels, and Brussels is not the last regulator. More obligations phase in over the coming years, and every one of them makes the same silent assumption: that you know where AI acts on your behalf.

This is the part I find genuinely useful about the deadline. Regulators are mandating something companies should have wanted anyway: an accurate map of how their business actually runs, including the parts software now handles. Organizations that treat August 2 as a one-off sentence insertion will repeat this scramble for every future deadline. Organizations that treat it as the week they finally documented their customer-facing processes get to reuse that work for every deadline after, and for every automation decision in between.

The fastest way to build that map is not a spreadsheet that starts aging the day it is finished. Map the customer-facing processes themselves, and the AI touchpoints fall out as a byproduct, alongside the handoffs, the owners, and the gaps. That is exactly the work Vevos AI was built to make fast: describe a process in your own words, or hand over the documents and recordings you already have, and it becomes living process documentation you can keep current long after this particular deadline has passed.

August 2 is days away. The sentence takes an afternoon. Start with the map on vevos.ai.