October 31, 2026Enterprise AI · Compliance & Governance5 min read

The First Sentence Matters: EU AI Act Article 50 for Customer-Facing AI Agents

Published by:Intelligent Research

Articles

2 Aug 2026Article 50 transparency obligations in force

2 Dec 2026Grace period deadline for marking AI content

€15M / 3%Maximum fine or total worldwide turnover

4Core transparency duties under Article 50

Most of the conversation about the EU AI Act has focused on what was delayed. The AI Omnibus, which entered into force on 27 July 2026 as Regulation (EU) 2026/1744, pushed the bulk of the high-risk obligations out to 2 December 2027 for stand-alone Annex III systems and 2 August 2028 for systems embedded in regulated products (White & Case).

What did not move is just as important. The transparency rules in Article 50 have applied since 2 August 2026 (European Commission). For banks, telecoms, healthcare providers and utilities now putting voice and chat agents in front of customers, those rules are already live. And one more date is coming fast: 2 December 2026.

Why this matters now

Customer-facing agents have moved from pilots into daily operations. In October alone, Banner Health described ElevenLabs voice agents that answer patient calls and book appointments directly in its medical record system (ElevenLabs), and McKinsey told banking CIOs that "the copilot phase is ending" as agents start to run workflows end to end (McKinsey).

Every one of those agents begins with a first sentence. Under Article 50, what that sentence says is no longer only a design choice.

What Article 50 actually requires

The Commission's FAQ and its Article 50 guidelines of 20 July 2026 set out four separate duties. Three are directly relevant to most enterprise deployments:

Tell people they are talking to AI. Systems that interact directly with people, such as chatbots, AI agents and avatars, must inform them that they are dealing with an AI system, unless that is obvious. The notice must come at the start of the first interaction, in a clear and distinguishable way that meets accessibility requirements.
Mark generated content. Providers of systems that generate synthetic audio, image, video or text must mark outputs in a machine-readable way so they can be detected as AI-generated.
Label certain published content. Deployers must label deepfakes, and AI-generated text published to inform the public on matters of public interest, unless that text has gone through genuine human review or editorial control.

The fourth duty covers emotion recognition and biometric categorisation, where deployers must inform the people exposed to those systems.

Fines for breaching these obligations can reach 15 million euros or 3% of total worldwide turnover, and enforcement sits mainly with national market surveillance authorities (European Commission). The rules also reach providers outside the EU when their system's output is used in the EU, which matters for teams in Türkiye, the UK and elsewhere that serve European customers.

The "obvious" exception is narrower than it looks

Many teams assume their assistant is obviously a bot. The Commission says the exception should be read restrictively and judged from the point of view of a reasonably well-informed, observant average person. A natural-sounding voice agent on a bank's main service line is exactly the case where that assumption breaks down.

The guidelines are concrete about what good looks like (Commission guidelines, PDF):

In chat, a plain-language first-turn greeting or banner, ideally placed near the input field, with a persistent label where the context calls for it.
In voice and telephony, an explicit spoken statement at the beginning, with periodic reminders in longer calls, especially after an interruption or when the agent's role changes. Audio tones alone are not considered enough.
For an email generated by an AI agent and sent to a person, an AI label at the top.

The 2 December 2026 date

There is one limited grace period. AI systems placed on the market before 2 August 2026 have until 2 December 2026 to meet the machine-readable marking and detection obligation for generated content. Systems placed on the market after 2 August were already required to comply. The grace period does not extend the duty to tell people they are speaking to AI (European Commission).

If your organisation built or bought generative tools before August that produce customer letters, synthetic voice, images or summaries, this is a one-month window to confirm where marking happens and who is responsible for it.

A practical checklist for the next 30 days

1. Inventory every customer-facing agent and generator. List each channel (voice, chat, WhatsApp, email, documents), whether it talks to people directly, and whether you are the provider, the deployer or both.

2. Write the disclosure into the agent, not the terms and conditions. Treat the opening line, reminders and handover-to-human messages as versioned, tested configuration, in every language you serve.

3. Test disclosure like any other requirement. Add it to your evaluation suite so a prompt change cannot quietly remove it, and keep logs that show it was delivered.

4. Confirm marking with your vendors. Ask model and platform providers how outputs are marked, and whether they follow the Commission-endorsed Code of Practice on Transparency of AI-generated Content.

5. Define real human review. If you rely on the editorial exemption for published text, document who reviews it and on what substantive grounds. Spell-checking does not count.

Transparency as part of the architecture

Disclosure works best when it is built into how agents are designed, tested and audited, not bolted on after launch. At Intelligent, we treat it the same way we treat permissions and audit trails: part of the system from day one, owned by the client's team, and provable on the record.

If you are reviewing your customer-facing agents ahead of December, we are happy to compare notes. Book a meeting with our team at intell.web.app.