Art. 50 EU AI Act: your assistant must disclose it is AI

I’m Denis Shokhirev, Agentic AI Systems Architect in Freiburg, Germany. I build and operate autonomous multi-agent systems for DACH B2B clients using Claude, Supabase, n8n, and self-hosted Postgres. Last week, a logistics client asked: “Are we legally required to tell users that our chatbot isn’t a person?”—not as a theoretical debate, but because a contract was at stake.

What Article 50 of the EU AI Act Actually Requires

Article 50 of the EU AI Act (CELEX:32024R1689) is explicit: whenever a user interacts with an AI system, the system must clearly inform them that they are not communicating with a person. For customer-facing chat, phone, or email assistants, this is not optional—it's the bare minimum for compliance. Non-compliance risks fines up to 7% of annual turnover (EU AI Act, Art. 99, 2024).

This is Not a "Paper" Problem

In production deployments, I’ve seen users routinely fail to distinguish whether they’re talking to a person or a bot. In one logistics pilot, 12% of users complained, believing a “human” was ignoring simple questions. This is not just a UI/UX concern—it’s a direct legal risk. If users are not clearly informed, the entire conversation can be ruled a violation of their right to accurate information under Article 50.

How to Implement Disclosure in Your Stack

Touchpoints: Where and How to Disclose

I enforce disclosure at two key touchpoints:

• At the start of the conversation (“This assistant is powered by artificial intelligence and is not a human.”)
• Whenever the user asks directly (“Are you a person?”—“No, I am an AI assistant designed to help with your product questions.”)

Sample Integration: n8n + Claude

// n8n workflow step example
const disclaimer = "Note: This assistant is AI, not a human.";
item.json.message = disclaimer + "\n\n" + item.json.message;
return item;

Automated Testing to Prevent Regressions

I automate checks to ensure the disclaimer is always present with a simple Python test:

import requests
def test_ai_disclaimer():
    resp = requests.post(
        "https://live.gerdennisai.com/api/chat",
        json={"message": "Hello"}
    )
    assert "assistant is AI" in resp.json()["answer"]

A test like this takes one minute to run and catches issues before they reach production.

Business Impact: Time, Cost, and Risk

Implementation Time

For a standard chat or email assistant, adding disclosure takes 1–2 hours of developer time. If the assistant is deeply integrated (e.g., via SAP APIs), expect 1–2 working days to align and test. For voice systems, confirming the message plays at the right time adds another 2–3 hours of QA.

The Cost of Getting it Wrong

Fines for non-compliance are up to €35 million or 7% of turnover (EU AI Act, 2024). In 2023, Italy’s Garante blocked ChatGPT for lack of transparency (Garante, 2023). Cyber insurance providers in Germany now include this risk in their calculations (Allianz Cyber Report, 2024).

When NOT to Use an AI Assistant

If your business relies on trust in a “real person” (e.g., a family-run notary), an AI assistant—even with upfront disclosure—may reduce client loyalty. In these cases, sticking with a traditional call center is the better answer.

Legal and Practical Details

Language Matters

Disclosure must be in the user’s language. A German user should see the disclaimer in German—otherwise, it may be considered invalid in a DACH audit.

Logging and Audit Trails

I log every instance of disclosure, including the exact text and timestamp. If a regulator audits, this proves the requirement was met at every contact point.

Ongoing Testing

Run automated tests monthly on all contact points. A single missed disclaimer is enough to trigger compliance issues and potential fines.

FAQ

How much does it cost to implement disclosure?

For a typical assistant, market rates in Germany (2024) are €200–500 if the code is well-structured. For complex integrations, costs can reach €1,500.

How long does it take?

For standard deployments, 1 working day from analysis to tests. Complex cases, 2–3 days.

What if disclosure fails?

Your company is liable. If a user is not properly informed, you risk fines and service suspension.

Does user data leave the company?

Depends on the architecture. If you use cloud APIs, data may go to the provider. Using self-hosted Postgres and Supabase keeps data local.

Who is responsible if something goes wrong?

The operator of the assistant (your company). Contractors are only liable per contract; final responsibility is on your business.

Have you already implemented disclosure on every user touchpoint, or do some “blind spots” remain? I offer a free 30-min AI compliance audit for DACH founders. DM me on LinkedIn or write to @ger_dennis_ai.