AI tools for teams: Claude, ChatGPT and the GDPR in practice

The short answer

Before your team uses Claude or ChatGPT at work, settle three things: who is the data controller and who is the processor, whether you have a data processing agreement with the provider, and where the data is stored. Then set up access rights and an internal rule for how staff pass work to each other, which is often forgotten.

This article is for team leads and administrators who decide how AI assistants will be used in the company. Below you will find the key terms in plain words, an eight-point checklist and the practical risk that arises not inside the assistant itself, but between colleagues.

Who is who: data controller and data processor

The General Data Protection Regulation (GDPR) sets out two main roles. The data controller decides why and how personal data is processed. The data processor processes data on the controller's behalf (GDPR Article 4, points 7 and 8). Between them there must be a contract that describes what the processor may do with the data (GDPR Article 28). That contract is the data processing agreement (DPA).

When a team works with an AI assistant, there are usually three parties:

• Your company. For the data of clients, staff and partners, it is usually the controller: you are the one who decides which data goes into the assistant and who it is passed to.
• The assistant provider: Anthropic (Claude) or OpenAI (ChatGPT). Its role and the terms that apply are set by your agreement with it and the plan you have chosen. So read the terms that apply to your plan, not the marketing page.
• Tools connected to the assistant, such as a handoff or document tool. Each has its own role and its own agreement.

Personal accounts deserve a separate mention. If an employee puts client data into their own personal assistant account, the company may have no agreement with the provider, no way to delete that data and no knowledge that it is there at all.

8 points worth checking

You can use this list with any AI assistant or any tool connected to one. Look for the answers in the provider's contract, the data processing agreement and the admin settings.

Accounts and agreement

The first two points shape all the others. As long as staff use personal accounts, the company controls neither the settings nor the data. Once accounts are managed centrally, you can decide on access, retention and what happens when someone leaves. In the data processing agreement, check who the sub-processors are and how you will hear about changes to them.

Data location and model training

Ask the provider, or find it in their documentation: where the data is physically stored and which safeguards apply if it is transferred outside the European Economic Area. Separately, check whether your plan uses data to train models. Settings can differ between plans, so rely only on the terms of your own plan.

Retention, access and logs

If a client asks you to delete their data, you need to know where it is and who can delete it. Shared projects and shared links are handy, but before using them, make sure every member should see everything in them. An audit log lets you answer "who saw this document, and when?" without guesswork.

Internal rule

One page is enough. It should cover which data must not go into the assistant (especially special categories, such as health data, GDPR Article 9), which data is allowed only in a company account, and how to hand work to a colleague. Put in only as much data as the task needs: that is the GDPR principle of data minimisation (Article 5(1)(c)).

Where the risk lies: handoffs between colleagues by email and chat

Even when the assistant is well configured, the weakest point can be the moment work changes hands. An employee has prepared a proposal for a client with Claude or ChatGPT, and now a colleague needs to take it over. Typically it looks like this:

• A copy in an email. An excerpt of the conversation with client data is pasted into an email. That data now sits in both the sender's and the recipient's mailboxes, often with no retention period at all.
• The wrong recipient. Address autocomplete or a group chat, and a contract ends up with someone it was never meant for. A sent email usually cannot be recalled.
• Files in several places. The same file is in the assistant conversation, in an email, in a chat app and on a colleague's laptop. When data has to be deleted, it is hard even to find every copy.
• Links to conversations. A shared link can show more than the recipient needs. Before sharing, check exactly who will see the conversation and whether it contains data they do not need.
• No record. A month later, nobody can say who received the client's document, or when. If the client asks or an incident happens, the answer has to be dug out of mailboxes.
• A leaver. The context stays in their personal conversations with the assistant. The company cannot see it, but the data remains there.

Every extra channel is one more place to protect, and later to clean up. A practical fix: agree on one channel for handoffs, and make sure you know three things about it: who can see the work that was handed over, how long it is kept, and where the record of who received what is.

How Colle handles this

Colle lets you hand work to a colleague straight from your AI assistant: you tell Claude or ChatGPT "send this to Anna", and Anna gets the summary, facts and files in her own assistant. From a data protection point of view, it works like this:

• Roles. For your team's content (summaries, facts, files and cases), Colle is the data processor under a data processing agreement (DPA); for the website, accounts and billing, it is the data controller. The DPA is available on request: email michael@trycolle.com with the subject "Colle DPA request".
• Assistant providers. Anthropic and OpenAI are not Colle's sub-processors. Your conversation with the assistant is processed by its provider under your agreement with them.
• Data location. Data is stored in the EU, in Frankfurt, on Fly.io infrastructure (Fly.io is a US company). Disks are encrypted, data travels over TLS and backups are made daily.
• Model training. Colle does not use personal data or customer data to train AI models, and it does not sell personal data.
• Who can see what. Work sent to a person is seen only by the sender and the recipients. Work sent to a space notifies all its members except the sender, and one of them picks it up; cases in closed spaces are seen only by members. Personal data is not shared with people who do not have access.
• Company isolation. Each company's data is kept apart by row-level security on every database query.
• Personal data in facts is encrypted with a separate key for each person. When a person is erased, their key is destroyed.
• Audit log. Entries are linked in a hash chain, so a changed or removed entry shows up. Admins can export the log.
• Retention and deletion. Customer data is deleted within 30 days of the end of the subscription, and from backups within a further 14 days; security logs are kept for 30 days. An admin can export or delete the company's data at any time.
• Breaches. If a personal data breach happens, Colle notifies your organisation without undue delay, because for your content it is the processor; you, as the controller, decide whether to notify the supervisory authority within 72 hours (GDPR Articles 33–34). For its own website, accounts and billing, where Colle is the controller, it notifies the authority itself.
• Protection against prompt injection. Text received from a colleague is marked as untrusted and scanned.

Colle does not replace your agreement with the assistant provider: what happens in a conversation with Claude or ChatGPT is governed by that agreement. Colle is responsible for what is passed between colleagues. The full description is on the security page. For a wider look at the same questions, read whether it is GDPR-safe to pass client data between AI assistants.

Supervisory authority contacts

In Lithuania, the supervisory authority for personal data protection is the State Data Protection Inspectorate (Valstybinė duomenų apsaugos inspekcija, VDAI). If your company is based in another EU country, contact the data protection authority of your own country instead. For example, a controller notifies the authority of a personal data breach within 72 hours of becoming aware of it, unless the breach is unlikely to put people's rights and freedoms at risk (GDPR Article 33).

The Lithuanian inspectorate's details:

• Address: L. Sapiegos g. 17, 10312 Vilnius
• Website: vdai.lrv.lt
• Email: ada@ada.lt

Contacts checked on 15 October 2026. Before getting in touch, check them on the inspectorate's website.

Frequently asked questions

Can we put client data into Claude or ChatGPT?

It depends on your plan, your agreement with the provider and the data itself. If you use a company-managed account with a data processing agreement, know where the data is stored and put in only what the task needs, the position is far clearer than with a personal account. For special category data, talk to a lawyer.

Do we need an agreement with every tool connected to the assistant?

If a tool processes personal data on your behalf, you need a data processing agreement with it. For Colle, you can get one on request: email michael@trycolle.com with the subject "Colle DPA request".

Can Colle see our conversations with the assistant?

Colle stores what is handed over: summaries, facts, files and cases. Your conversation with Claude or ChatGPT is processed by the assistant provider under your agreement with them; Anthropic and OpenAI are not Colle's sub-processors.

What happens to our data if we stop using Colle?

Customer data is deleted within 30 days of the end of the subscription, and from backups within a further 14 days. Before that, an admin can export the company's data.

How much does Colle cost?

€15 per active person per month, or €12.50 per month when paid annually, excluding VAT. The first 7 days are free, and no card is needed. More on the pricing page.