AI in Customer Service: EU AI Act Compliance
Using AI in customer service — chatbots, AI agents, voice bots or AI-drafted replies — is usually not high-risk under the EU AI Act. But it is squarely covered by Article 50's transparency rule, which has applied since 2 August 2026: a customer must be told they are talking to an AI unless it is obvious. A support bot only becomes high-risk when it decides something Annex III lists, such as credit or insurance eligibility.
The one rule every chatbot must follow
Article 50(1): Before any interaction, inform the user they are talking to an AI. Not in the terms of service. Before the interaction. Explicitly.
Covered customer service AI products
Customer service chatbots on websites and apps
Article 50(1) — disclose AI nature
AI voice assistants in call centres
Article 50(1) + (3) if emotion recognition used
AI-generated email or ticket responses
Article 50(2) — label generated content
AI support agents using a human persona
Article 50(1) — no false impression of human interaction
AI sentiment analysis on customer contacts
Article 50(3) — disclose emotion recognition
Foundation model APIs provided to third-party chatbot builders
Articles 51–55 — GPAI provider obligations
Key obligations
Disclose AI nature before interaction
Any chatbot or AI assistant that interacts with natural persons must proactively disclose that it is an AI before the interaction begins. This must be explicit, clear, and not dependent on the user reading terms of service. Implement a visible 'You are chatting with an AI assistant' notice at the start of every conversation session.
Label AI-generated content
If your customer service AI generates text, images, audio, or video content that users receive — beyond simple conversational responses — that content must be marked as AI-generated in a machine-readable format. This applies to AI-generated email responses, AI-drafted support documents, or AI-produced explanatory content delivered to customers.
Emotion recognition disclosure
If your customer service AI uses emotion recognition or biometric categorisation — for example, detecting customer sentiment from voice or text to route calls or adjust responses — you must disclose this to the individuals being assessed. This applies to call centre AI that infers emotional state from voice patterns.
General-purpose AI model obligations
If your AI product is itself a general-purpose AI model accessed by third parties, you may have obligations as a GPAI model provider — including publishing technical documentation, a model card, and a summary of training data. These obligations apply to foundation model providers, not to downstream chatbot operators. If you are the chatbot operator, these apply to your model supplier.
Common compliance gaps in customer service AI
No disclosure at session start
The most common gap: the chatbot does not display any AI disclosure before the conversation begins. Adding terms of service that mention AI does not satisfy Article 50 — the disclosure must precede the interaction and be explicit.
AI impersonating a named human agent
Some customer service deployments give the AI a human name and persona ('Hi, I'm Sarah from support'). Article 50(1) prohibits deceiving users into thinking they are interacting with a human when they are interacting with an AI. A human name is not inherently prohibited, but using it to create a false impression of human interaction is.
No machine-readable AI labels on generated content
Article 50(2) requires machine-readable labelling on AI-generated content. Purely human-readable notices ('This content was generated by AI') may not satisfy the technical requirements once the European Commission publishes implementing standards. Prepare to support the C2PA standard or equivalent technical watermarking.
Assuming the GPAI provider handles all compliance
Using a foundation model from a third-party provider does not transfer your Article 50 obligations to that provider. You are responsible for the interface your users interact with. The model provider is responsible for the model. Your disclosure obligations exist regardless of which underlying model powers your chatbot.
Get your Article 50 compliance checklist
Nytivo generates Article 50 compliance checklists and technical documentation for chatbot and AI assistant providers.
EU AI Act for Customer Service AI — Frequently Asked Questions
Does the EU AI Act apply to our customer service chatbot?
Yes, but typically not as a high-risk system. Article 50(1) of the EU AI Act requires that any AI system intended to interact with natural persons disclose — before the interaction begins — that the person is interacting with an AI, unless this is obvious from context. A customer service chatbot is almost certainly covered by this transparency obligation. The more important question is whether your chatbot is also powered by a general-purpose AI (GPAI) model, which triggers additional obligations on the model provider under Articles 51–55.
What exactly does Article 50 require us to display to users?
Article 50(1) requires that users be informed they are interacting with an AI system before any interaction takes place. The disclosure must be explicit — not buried in a terms of service, not a small tooltip, not implied by the interface. The regulation does not prescribe the exact format, but it must be given in a way that is clear and comprehensible to the user. For a chat interface, a message at the start of every session stating that the user is talking to an AI assistant is the safest implementation. The disclosure must also be machine-readable to support regulatory compliance checks.
Our chatbot uses GPT-4 or Claude as its backbone. Does that change our obligations?
Yes and no. Using a third-party GPAI model (from OpenAI, Anthropic, or similar) means the model provider has obligations under Articles 51–55 — including publishing a model card, a summary of training data used, and compliance with the GPAI Code of Practice. As a deployer of that model, your Article 50 obligations remain unchanged — you still need to disclose the AI nature of the interaction to your users. You cannot delegate your Article 50 compliance to the model provider. You are responsible for the interface your users interact with.
Does Article 50 apply if our chatbot mostly handles simple FAQ queries?
Yes. Article 50(1) does not contain a de minimis exception based on the complexity of the interaction. The obligation applies to any AI system intended to interact with natural persons, regardless of the sophistication of the conversation. A chatbot that only answers product questions still needs to disclose that it is an AI. The only exception in Article 50(1) is where it is obvious from context — for example, a clearly labelled 'AI assistant' interface with explicit signposting throughout.
When does Article 50 for chatbots come into force?
Article 50 transparency obligations have applied since 2 August 2026; the Digital Omnibus (Regulation (EU) 2026/1744) did not move that date. Article 50(1) is a duty on the provider of the chatbot: it must be designed so people are told they are interacting with an AI system, unless that is obvious. If you build your own support bot on a model API, you are usually its provider. Obligations for the underlying general-purpose AI model (Articles 53–55) have applied to the model provider since 2 August 2025. Disclosure is usually a UI change, so there is no reason to wait.