When a company starts using an AI chat, the questions come up quickly: what is stored, where, by whom and for how long? The answer is not the same for every service, and that is often what makes the question hard. This article gives a practical working order for sorting it out. It is not legal advice. Which requirements apply to your business, and how they should be met, should be assessed with qualified advice, for example from a data protection officer or a lawyer.
Start from the flow, not from assumptions
A chat can involve different kinds of information depending on channel, settings and providers. So do not assume your service works like someone else's. Sketch the path a piece of information takes, from the visitor typing to where it may leave the system. Try to answer four questions:
- What is submitted? Messages, contact details in a booking, technical information.
- What does the system create? Conversation logs, summaries, booking records.
- Where is it passed on? Calendar, customer records, email, other services.
- Which providers are involved, and where is the data processed?
You find the answers in your own configuration and in the providers' documentation, not in a general description. Which connections exist depends on platform, plan and setup, so start with the integrations and channels you actually plan to use.
Define purpose and minimise
For each piece of information you collect, you should be able to say why you need it. If the purpose is to answer a question about opening hours, no personal data is needed at all. If it is a booking, you need what the booking requires, but rarely more. Write the purposes down and instruct the agent not to ask for what is not needed. Ask it also not to encourage visitors to share sensitive information in free text. What the agent answers from is governed by the knowledge you add, so keep it free of personal data.
Decide who is responsible and who may see
Name a person responsible for privacy matters around the chat, and decide which roles may read conversations. Access should be limited to those who need it for their work. Include in the assessment who on the provider's side can reach data, for example during support, and on what terms. Where the agent hands over to staff, the information goes with it, so the handover is part of the mapping too.
Retention: a general number of months does not fit everyone
There is no figure that applies to every business. How long a conversation should be kept depends on why you keep it, which legal basis you have, which other obligations exist and how sensitive the information is. The method is the same every time:
- Write down why the conversation is kept.
- Decide when that purpose is fulfilled.
- Decide what happens then: deletion or anonymisation.
- Check that the settings in each system actually follow the decision.
- Review the decisions at regular intervals, especially when you change how you use the chat.
Have a qualified adviser assess the legal basis and which periods are reasonable for you. Write the decision down so you can show how you reasoned.
Review agreements and sub-processors
If a provider processes personal data on your behalf, a written data processing agreement is generally required. Request it in advance, and read in particular what it says about sub-processors, transfers outside the EEA, access, help with data subjects' requests and deletion when the agreement ends. Find out also how you will be told if the provider changes a sub-processor. More questions to ask are in the vendor selection checklist, and terms such as data processor are explained in the glossary.
Routines for data subjects' requests
People can request access to their data, correction or deletion. Have a routine that answers:
- Who receives the request, and within what time must it be handled internally?
- How do you identify the person?
- In which systems does the data exist: the chat platform, the calendar, the customer records, email?
- How do you delete or correct in each system, and how do you check it was done?
Try the routine with a test person before a real request arrives. You will find out where the data actually lives, and it is rarely exactly where you thought.
Keep the documentation alive
Documentation is only useful if it is accurate. When you add a channel, change provider or change what the agent asks for, the mapping should be updated. Tie it to your regular routines for maintaining the agent's knowledge base and to the tests before launch, see testing the AI agent before launch. A closer look at common wrong assumptions is in AI and GDPR: common misunderstandings.
If you want to discuss what your setup could look like, you are welcome to contact us. Remember that the assessment of what the law requires of you specifically should always be made with qualified advice.
