A chatbot on a website feels like a small addition, a widget you paste in before lunch. Legally, it is several things at once: a new way of collecting personal data, a new source of statements made in your name, and, since August 2026, a system that triggers transparency obligations under the AI Act.
None of this is particularly difficult to handle, but everything is easier to address before launch than after the first complaint.
Tell People They Are Talking to a Machine
Article 50 of the AI Act, applicable since 2 August 2026, requires people interacting with an AI system to be informed that they are talking to AI, unless this is obvious from the circumstances.
A widget with a friendly name and a human photo does not necessarily make this obvious.
What you should do:
Identify the chatbot as AI in the widget label.
Include a clear disclosure in the first message.
For voice agents, include the disclosure in the greeting.
Make it easy for users to request human assistance.
For example:
I am an AI assistant for this business. A person can step in at any time.
Treat the Chat as Personal Data
Conversations contain names, email addresses, order numbers and, because people type what they like, sometimes health or financial details you never asked for.
That makes chat logs personal data, and the usual data protection obligations follow.
A lawful basis: For customer support, this may be contractual necessity or legitimate interests rather than consent, depending on the circumstances.
A privacy notice: Explain the chatbot's purposes, the vendor, relevant AI model providers, recipients, and any international data transfers.
A data processing agreement: Put the appropriate agreement in place with the vendor and review its subprocessor list.
A retention rule: Define how long conversations are stored and implement a deletion routine.
A process for data requests: Make sure applicable access and deletion requests can be handled for chatbot logs.
Data minimisation by design: Do not ask for more information than the task requires, and tell users not to enter unnecessary sensitive details.
Cookies and Scripts
Many chat widgets load third-party scripts that set cookies or access identifiers on a visitor's device.
In France, the CNIL generally expects consent before non-essential trackers are used, subject to applicable exceptions. Refusing must be as easy as accepting.
If the widget loads before the visitor has made the required choice, you may have a compliance problem, whatever the vendor says.
What you should do:
Load the widget only after the required consent has been obtained.
Choose a solution that does not use non-essential trackers where appropriate.
Verify what scripts and cookies the widget actually loads.
Remember that the CNIL recommends renewing consent choices at appropriate intervals, commonly every six months.
Decide What the Bot Is Allowed to Say
Whatever the chatbot tells customers about prices, delivery, refunds or conditions may be understood as coming from your business.
Limit it to a knowledge base you control, make it show or link its sources for factual answers, and have it decline topics outside its scope instead of improvising.
Set clear boundaries:
Use approved, up-to-date business information.
Prevent the chatbot from inventing prices, discounts, or refund promises.
Make it acknowledge uncertainty when it cannot verify an answer.
Restrict legal, medical, and financial advice where appropriate.
Escalate quickly when a customer is upset, asks to make a complaint, mentions a deadline, or repeats a question without getting a satisfactory answer.
The chatbot should know when to stop answering and involve a person.
Make the Handoff Real
A visible route to a human, with a stated response time, is both good service and a safeguard.
Put a contact option in the widget.
Explain when users can expect a response.
Make sure someone is actually monitoring incoming requests during the hours you promise.
Test the handoff before launch.
A human support option is only useful if it works when customers need it.
Check Accessibility
If your site sells to consumers in the EU, the European Accessibility Act may cover the e-commerce service, subject to its scope and applicable exemptions, including the microenterprise exemption.
The chat widget is part of that service.
Before launch, check that it:
Works with a keyboard.
Supports screen readers.
Has sufficient colour contrast.
Does not trap keyboard focus.
Provides accessible labels and controls.
Test the actual widget rather than trusting a vendor's claim of compliance.
Test Before You Go Live
Write twenty real questions from your own inbox and run them through the bot, including French-language questions where relevant, poor spelling, and off-topic requests.
Then try to break it.
Ask it to ignore its instructions.
Ask it to reveal its configuration.
Ask it to promise an unauthorised refund.
Test whether it invents answers when information is missing.
Check whether it escalates questions it cannot safely answer.
Fix what fails, and run the test set again after every significant change to the knowledge base or chatbot configuration.
During the first month, review transcripts every week, using appropriate access controls, and keep a list of questions the chatbot could not answer.
The Launch Checklist
Before going live, confirm that:
AI disclosure appears in the widget label and the first message.
The privacy notice is updated, the data processing agreement is in place where required, and subprocessors have been reviewed.
A retention period is defined and deletion is scheduled.
Required consent is obtained before non-essential tracking scripts load.
The knowledge base is limited to approved information and sources.
Human handoff is visible and monitored.
Accessibility has been tested.
The twenty-question test has passed and weekly reviews are scheduled.
The Same List Applies to Voice AI
The checklist works for phone agents too.
Mirage Cloud's receptionist agent, Xavier, answers calls for businesses, and the points above apply directly:
The AI disclosure belongs in the greeting where required.
Call recordings and transcripts containing personal data need an appropriate retention and deletion policy.
Transferring a caller to a person should be straightforward.
The provider's data processing arrangements and subprocessors should be reviewed.
Mirage Cloud publishes a full subprocessor annex that identifies its voice provider, so businesses can review the data-processing arrangements using documentation rather than relying solely on marketing claims.
The principle is the same for both chat and voice: put the safeguards in place before launch, not after the first complaint.
Top comments (0)