Before an AI tool reads your customer emails, invoices or internal documents, you should know where that data goes and who can see it. The good news is that the questions are not technical. Any supplier should be able to answer them in plain language.

This article is general information, not legal advice. For decisions about your specific situation, talk to a lawyer or a data protection specialist.

Where is the data stored and processed?

AI tools are rarely a single piece of software. A typical agent involves several services: somewhere the agent runs, an AI model provider that processes the text, and perhaps a database for searching your documents. Your data may pass through all of them.

Ask:

  • Which services are involved, and which company runs each one?
  • In which country or region is the data processed and stored? Is it within the EU or the EEA?
  • How long is it kept? Is text deleted after processing, or stored in logs? For how long?

Processing outside the EU and EEA is not automatically forbidden, but it needs a legal basis and safeguards. That is one more thing to check, so it is simpler when everything stays inside the EU where possible.

Ask for the answers in writing. A verbal "it's all secure" is not enough when a customer or employee later asks where their data is. A written description of the services and their locations also helps when you update your privacy notice.

Is your data used to train AI models?

This is the question many people worry about most, and rightly so. If your data were used to train a public model, fragments of it could in theory surface elsewhere.

The answer depends on the specific service and plan. Consumer versions of AI tools and business or API versions often have different terms. Many providers' business terms state that customer data is not used for training by default, but you should not assume. Ask the supplier which service and plan they use, and ask to see the relevant terms.

A practical rule: do not paste customer or confidential data into free consumer AI tools unless you have checked their terms.

Who has access?

Access rights matter at two levels.

The agent's access to your systems. An agent should only be able to read and do what its task requires. If it sorts the shared inbox, it does not need access to your HR folder. Ask for access to be set with the least rights needed, and to be easy to revoke.

People's access to the data. Who at the supplier can see your data, and why? Who in your own company can see the agent's logs? Logs can contain the same personal data as the original emails.

When the cooperation ends. Agree in advance how access is removed and what happens to the data and logs the supplier holds. They should be returned or deleted, and it is worth asking for confirmation of the deletion.

GDPR basics in plain terms

If the agent handles personal data, such as names, email addresses or anything else that identifies a person, the EU General Data Protection Regulation applies. The core ideas are straightforward:

  • You remain responsible. Your company is usually the controller: you decide why and how personal data is used. The AI supplier is usually a processor acting on your instructions.
  • Have a data processing agreement. When a processor handles personal data for you, the GDPR requires a written agreement setting out what they may do with it. Ask for one before any personal data is shared.
  • You need a lawful reason for processing the data. For many business tasks this already exists, for example handling a customer's order. Using the data for something new should be checked.
  • Use only what is needed. If the task works without personal data, leave it out. If it needs some, use as little as possible.
  • Know the sub-processors. The supplier should tell you which other companies, such as the AI model provider or the hosting service, also process the data.
  • People keep their rights. Customers and employees can ask what data you hold about them and ask for it to be corrected or deleted. You need to be able to answer, including for data in the agent's logs.
  • Update your privacy notice if you start processing personal data in a new way.

A short checklist

Before any AI tool touches customer or internal data, you should be able to tick these off:

  1. We know which services the data passes through, and who runs them.
  2. We know where the data is processed and stored, and for how long.
  3. We have checked, in writing, whether our data is used to train AI models.
  4. The agent has only the access its task needs, and we can revoke it.
  5. We know who can see the data and the logs, on both sides.
  6. We have a data processing agreement with the supplier.
  7. We know the sub-processors.
  8. We use only the personal data the task needs.
  9. We can answer a person who asks what data we hold about them.
  10. Our privacy notice reflects the new processing.

Keep the first project low-risk

The easiest way to stay safe is to start with a task that handles little or no sensitive data. Internal search over your own instructions, or sorting supplier invoices, carries far less risk than processing health information or HR records. How to choose your first AI project explains how to factor risk into the choice.

In short

Data safety with AI is less about technology and more about asking the right questions before you start. Where does the data go, is it used for training, who can see it, and is there an agreement in place. A supplier who answers these clearly and in writing is one you can work with.