EU AI Act Article 50: The Chatbot and Voice Agent Rules That Still Hit August 2, 2026

If you run a chatbot or voice agent that talks to anyone in the EU, you have until August 2, 2026 to make sure those users know they're talking to an AI. That's the eu ai act chatbot disclosure requirements deadline under Article 50, and contrary to most of the 2026 coverage you've been forwarded, it did not slip with the rest of the Act.

Most client alerts you're reading right now lead with the Digital Omnibus delays and bury the part that affects you. So let's invert it. The high-risk stuff moved. The transparency stuff didn't. If you're shipping a chat widget, a voice IVR replacement, or a sales SDR agent that calls EU prospects, you're in scope, and you have roughly two months from today (June 2026) to ship the disclosure layer.

TL;DR

• Article 50 transparency duties take effect August 2, 2026 — the Digital Omnibus did not postpone them.

• High-risk AI Act obligations slipped to December 2, 2027 and August 2, 2028, which is what's confusing operators.

• Chatbots and voice agents must disclose AI status to users unless it's obvious from context — a narrow exemption.

• Synthetic content watermarking grace period ends December 2, 2026, four months after the disclosure rule.

• US businesses serving EU users are in scope. Geography of the user, not the company, is what triggers Article 50.

1. What the Digital Omnibus actually delayed — and what it didn't

The EU Digital Omnibus, agreed in late 2025, postponed the bulk of the high-risk AI Act deadlines. Per Gibson Dunn's November 2025 analysis, high-risk system obligations now apply from December 2, 2027, with the full sweep landing August 2, 2028.

That's the headline most operators saw. Here's what got cut from the headline: Article 50 transparency duties were not part of the postponement. The chatbot disclosure rule, the voice-agent disclosure rule, and the deepfake labeling rule all keep their original August 2, 2026 start date. The synthetic-content watermarking grace period was preserved, ending December 2, 2026 — that's four months after the core disclosure rule activates.

So the operator read is simple. If your product touches EU users through conversation or generated media, your calendar has two dates on it: August 2, 2026 for the disclosure layer, and December 2, 2026 for watermarking on any synthetic content you produce. The 2027 and 2028 dates are a different conversation — those are for high-risk classifications most chat and voice deployments will never trigger.

2. Article 50 in plain English: who must disclose, and when 'obvious from context' applies

Article 50 of the AI Act says providers of AI systems intended to interact directly with natural persons must design and operate them so that those persons are informed they're interacting with an AI — unless this is obvious from the standpoint of a reasonably well-informed person, taking into account the circumstances and context of use.

The operator translation: if a reasonable user might think they're talking to a human, you have to tell them they're not.

The obvious from context carve-out is narrower than most agencies are selling it. A chat widget on your pricing page that opens with "Hi, how can I help?" is not obvious. A widget labeled "AI Assistant" with a robot icon and a one-line preamble — that's obvious. The difference is whether the disclosure is baked into the interface, not whether the user could have figured it out.

There's a law-enforcement exemption for authorized criminal-offense detection and investigation. Almost nobody reading this qualifies. Assume you don't.

Article 50 scope test. The disclosure layer is required from August 2, 2026.

3. Chat widget disclosure: placement, wording, accessibility

The rule doesn't prescribe exact wording. That's good and bad. Good because you have room to match your brand voice. Bad because "we'll figure it out" is how teams ship something that fails an audit.

Here's the operator-grade pattern we use for client builds, and what we'd ship Monday if it were our widget:

Placement. The disclosure lives in three places: the widget header (persistent label), the first message (explicit sentence), and the bot's identity card if a user clicks to expand it. Persistent label, opening line, profile — that's the belt-and-suspenders read.

Header label. "AI Assistant" or "AI Chat" next to the bot name. Not "Assistant." Not "Help." The word AI has to appear in the chrome.

Opening message — copy-paste version:

Hi, I'm Ada, an AI assistant for [Company]. I can help with [scope]. For anything I can't handle, I'll connect you to a human teammate. How can I help?

That sentence does four things: names the AI, names the company, scopes the capability, and signals the human escalation path. It also tees up the agent governance framework most teams skip — what the bot does, what it doesn't, who picks up when it can't.

Accessibility. Screen readers need the disclosure too. The persistent label must be in the accessible name of the widget, not just the visual chrome. If your widget renders "AI Assistant" as a background image, you've failed the rule for blind users — and probably your local accessibility law on top of it.

4. Voice agent disclosure: the first-15-seconds script

Voice is where most teams will get this wrong. Phone calls don't have a header label. The disclosure has to happen in the audio, and it has to happen early.

The ai voice agent disclosure law under Article 50 doesn't specify a timestamp. Our operating standard, based on what's clearly defensible: the AI disclosure happens in the first utterance, before any substantive question is asked, and within the first 15 seconds of the call.

Inbound voice agent — opening script:

Thanks for calling [Company]. This is Ada, an AI voice assistant. I can handle [scope] — for anything else, I'll transfer you to a teammate. To get started, can you tell me your account number?

Outbound voice agent — opening script:

Hi, this is Ada calling from [Company]. I'm an AI voice assistant following up on [reason]. Is now an okay time to talk for about two minutes?

The outbound version matters more than the inbound one. Cold-call AI agents are where regulators and journalists will look first, because that's where the public-trust failure mode lives. Get the disclosure into the first sentence, before you ask anything. Don't bury it after a question.

If you're stress-testing voice scripts before launch, the patterns in our voice agent testing guide cover the disclosure-audit beat specifically.

5. Synthetic content and the watermarking grace period ending December 2, 2026

Article 50 also covers AI-generated or AI-manipulated content. If your product generates synthetic audio, image, video, or text that could be mistaken for authentic, you have to mark it as artificially generated in a machine-readable format.

The watermarking obligation has a separate runway. Per Gibson Dunn's reading of the Omnibus, the grace period ends December 2, 2026 — four months after the disclosure rule.

For most chatbot and voice-agent operators, this is a downstream concern. You're not generating deepfake video. But if your stack includes AI-generated marketing creative, AI-voiced product demos, or anything that synthesizes a real person's voice or likeness, you're in scope and your December deadline is real.

6. Does this apply to a US business serving EU customers?

Short answer: yes.

The AI Act's territorial scope follows the user, not the provider. If your chat widget or voice agent is placed on the market in the EU, or its output is used in the EU, you're in scope. A US SaaS company with EU customers logging into its product and using its chatbot is in scope. A US e-commerce site whose checkout chat answers a buyer in Berlin is in scope.

The operator test we use: if a user in the EU can talk to your bot during the normal course of using your product, assume Article 50 applies and ship the disclosure. The cost of disclosing to everyone is a header label and a sentence. The cost of not disclosing to an EU user is a regulatory finding. The asymmetry isn't close.

This is also where the eu ai act august 2026 deadline catches US teams off guard. They read the Omnibus headlines, see "delayed to 2027," and stop reading. The disclosure layer is the part that didn't move.

7. A 7-item readiness checklist for August 2

What to do this month, in order:

1. Inventory every conversational surface. Web chat, in-app chat, SMS bot, voice IVR, outbound voice agent, embedded assistants. List them with owners.

2. Tag which surfaces serve EU users. If you can't prove a surface doesn't serve EU users, assume it does.

3. Ship the header label. "AI Assistant" or equivalent in the widget chrome, in the accessible name, on every in-scope surface.

4. Ship the opening message. Use the copy-paste pattern in Section 3. Get legal sign-off on the exact wording once; reuse everywhere.

5. Rewrite voice greetings. Inbound and outbound. AI disclosure in the first utterance, within 15 seconds. Re-record where needed.

6. Document the governance contract. What the bot decides, what it escalates, who gets paged. Same one-pager pattern from the agent governance framework. This is what an auditor asks for.

7. Calendar December 2, 2026 for the watermarking review on any synthetic content you ship.

Most of this is a two-week build for a team that already has the surfaces in production. If you're greenfield, wire the disclosure into the ai-automation plumbing from day one and you'll never have to retrofit.

FAQ

Did the EU Digital Omnibus delay Article 50?

No. The Digital Omnibus postponed high-risk AI Act obligations to December 2, 2027 and August 2, 2028, but Article 50 transparency duties for chatbots and voice agents take effect August 2, 2026 as originally scheduled. The synthetic-content watermarking grace period also kept its December 2, 2026 end date.

What exactly does 'obvious from context' mean for a chat widget?

It means a reasonably well-informed user would recognize they're talking to an AI without being told. In practice, that bar is high. A widget labeled "AI Assistant" with an explicit opening sentence clears it. A generic "Chat with us" button does not. Treat the exemption as narrow and disclose by default.

Are US companies serving EU customers subject to Article 50?

Yes. The AI Act applies based on where the system is used and where its output lands, not where the provider is headquartered. A US business whose chatbot or voice agent interacts with users located in the EU is in scope and must comply by August 2, 2026, the same as an EU-based provider.

Does Article 50 require a specific disclosure script?

No prescribed wording exists, which means operators choose the language. The requirement is that natural persons are informed they're interacting with an AI. Most legal readings expect the disclosure to be clear, prominent, and presented before substantive interaction begins — not buried in a footer or terms of service.

What's the difference between the August 2026 and December 2026 deadlines?

August 2, 2026 is when chatbot and voice-agent disclosure obligations under Article 50 begin. December 2, 2026 is when the grace period for synthetic-content watermarking ends, meaning AI-generated audio, image, video, or text must be marked in machine-readable format from that date forward. Two separate obligations, two separate runways.

What to ship this week

Pull every conversational surface you own into one spreadsheet, mark the ones that touch EU users, and write the opening sentence for each. That's a Monday morning. The rest of the work follows from there.

If you'd rather have the wiring done for you, say hi.