AI and GDPR: when ChatGPT and Claude are an option for client data

A GDPR-compliant use of ChatGPT or Claude for client work can be possible if you use a plan with a data processing agreement (DPA) and no training on your data, and enter only what the task needs. You also need to be able to explain it to your clients. Free and personal plans are the weak spot: keep personal data of clients and staff out of them, or make it anonymous first.

This guide is general information and does not replace legal advice, because the assessment depends on your individual case. Ask a data protection or tax professional about your own case.

Vendor terms (as of 08/10/2026) change often, so check the pages named under each point before you rely on them.

When is client data in an AI chat an option?

It can be possible if you meet GDPR requirements. An AI chat is a service that processes personal data for you. Names, email addresses, invoice amounts linked to a person and project notes about a person all count as personal data (Art. 4 (1) GDPR). Four things decide:

1. A legal basis. For client work this can usually be performance of the contract or your legitimate interest (Art. 6 (1) (b) and (f) GDPR). Special categories, such as health data, need more (Art. 9).
2. A processor contract. If a provider handles personal data on your behalf, Art. 28 GDPR requires a written agreement, the DPA. Without one you should not enter personal data.
3. Data minimisation. Enter what the task needs and nothing else (Art. 5 (1) (c) GDPR). "Write a friendly reminder for an invoice of EUR 2,400, due 15 September" does not need the client's name or IBAN.
4. Transparency. Your privacy notice should say that you use AI tools and which kind of data goes there (Art. 13 GDPR).

Two cases are simpler. Text without personal data, such as a template or your own price list, is an option for any plan. And with your own data, the decision is yours.

If you are a tax adviser, lawyer or in another profession bound by secrecy, also check § 203 StGB (DE). Under it, you bring in helpers only as far as necessary and bind them to secrecy. In Austria your professional law sets the duty, for example § 9 RAO for lawyers and § 80 WTBG 2017 for tax advisers. Ask your professional body before you use AI on client files.

Which plans offer a DPA and no training on your data?

The short version, as of 08/10/2026:

What the sources say:

• OpenAI: it says that data from ChatGPT business plans and the API is not used for training unless you opt in. For personal plans, you can switch off "Improve the model for everyone" under Settings > Data Controls; it then applies to the whole account. Temporary Chats are deleted after 30 days and not used for training. For businesses in the EEA, the DPA names OpenAI Ireland Ltd as the contracting party.
• Anthropic: training on Free, Pro and Max chats depends on a setting; with it on, data is kept for five years, otherwise 30 days (announcement of 28/08/2025). Team, Enterprise and the API run on the Commercial Terms, which exclude training on customer content, and the DPA, including Standard Contractual Clauses, is part of them automatically. You do not sign it separately.

Two points to keep in mind. First, "no training" is not "no storage": providers may keep data for a limited time for safety and abuse checks. Second, check the plan behind your login. Many freelancers use a private Plus or Pro account for client work out of habit. If that is you, switch to a business plan or turn the training setting off and stay anonymous.

What about data going to the US?

Both providers are US companies. A transfer outside the EU needs a legal basis (Art. 44 to 46 GDPR). The usual ones are the EU-US Data Privacy Framework or Standard Contractual Clauses in the DPA. The Framework's adequacy decision is in force. The General Court rejected a challenge on 03/09/2025; the appeal (Case C-703/25 P) was still pending at the Court of Justice when we checked. If the Framework fell, the Standard Contractual Clauses in the DPA could be the fallback basis. You can also choose storage in Europe where a provider offers it: OpenAI does for new Enterprise and Edu customers and for new API projects, as per its announcement.

What do Germany and Austria add to the GDPR?

The GDPR applies in both countries in the same way. The national laws add details.

Germany:

• The BDSG supplements the GDPR. Under § 38 BDSG you usually need a data protection officer only once at least 20 people are constantly busy with automated processing of personal data, or if your processing needs a data protection impact assessment; Art. 37 GDPR can also apply. A two-person studio usually does not.
• The state data protection authorities supervise you. The joint body of German authorities (DSK) published the paper "Künstliche Intelligenz und Datenschutz" on 06/05/2024. It is a guide for companies that use AI; it is not binding, but it shows what regulators look at. One example: you should check whether the AI application was trained lawfully.

Austria:

• The DSG supplements the GDPR. Section 1 DSG makes data protection a fundamental right: everyone may demand secrecy of data about them, where they have a worthy interest.
• The Datenschutzbehörde supervises you. Complaints go there (§ 24 DSG).

For a small business, the practical result is the same in both: DPA, minimisation, transparency. Fines for breaches are set in the GDPR itself.

What do GoBD (DE) and BAO (AT) mean for AI-made records?

The AI does not change the rules for your books. It only changes who types.

• A draft is not a record. What the AI writes is a draft until you issue it. After that it counts as an accounting document.
• Issued means unchanged. You must not change an issued invoice so that its original content is lost (§ 146 (4) AO in Germany, § 131 (1) BAO in Austria). If the AI or you made a mistake, you cancel and re-issue. See Cancelling or correcting an invoice.
• The chat is not your archive. Keep the final invoice or receipt, not the conversation. Retention is eight years for invoices in Germany (§ 147 AO), seven in Austria (§ 132 BAO). Details in How long to keep invoices and receipts.
• Document how you use AI. The GoBD expect your process to be traceable. For a small business, a short note can be enough: which tasks the AI drafts, who checks, where the final document is saved. See GoBD. No AI tool is "GoBD certified", because there is no such certificate.
• You stay responsible. Whether VAT is right, whether a client is in default, whether a receipt is deductible: you or your tax adviser decide, not the AI.

What does AI Act Art. 4 ask of a two-person studio?

Art. 4 of the AI Act covers "AI literacy". Providers and deployers of AI systems must take measures so that their staff, and others who use AI for them, have enough understanding of it. You are a deployer as soon as you use an AI tool in your business.

What changed: Regulation (EU) 2026/1744, the "Digital Omnibus on AI", was published on 24/07/2026 and has been in force since 27/07/2026. It replaced Art. 4. The old text said you must "ensure" a sufficient level. The new text asks you to take measures that support the development of AI literacy, and says that this does not mean guaranteeing a specific level for any person. The Commission and the member states must support small and medium-sized businesses. The old duty had applied since 02/02/2025.

For a studio of two, it can look like this:

• one hour to read the vendor's data pages and this guide;
• a one-page rule, see below;
• a short note when someone new joins or you start a new tool.

Keep the rule and the date you went through it. It can serve as cheap proof that you took measures.

What should a one-page AI rule for your team say?

Copy it, adjust it, and put it where everyone sees it.

**AI rule of [Company], valid from [date]**

1. We use only these AI tools for work: [e.g. ChatGPT Business, Claude Team]. Private accounts are not used for client work.
2. Personal data of clients, staff and applicants goes only into tools with a signed or accepted DPA. In other tools we use made-up or anonymised data.
3. We never enter passwords, bank logins, health data or anything that is under a secrecy duty.
4. We enter only what the task needs. Names and IBANs are left out when they are not needed.
5. AI output is a draft. A person checks every figure, date and legal statement before it is sent or issued.
6. Issued invoices are never edited. Mistakes are cancelled and re-issued.
7. We save the final document in [place], not the chat.
8. Before we start a new AI tool, [name] checks training, DPA and storage location and notes the date.
9. If something goes wrong (wrong recipient, data in the wrong tool), tell [name] at once.
10. [Name] reviews this rule every [6] months and when a tool changes its terms. Everyone has read it: [date].

How does Backbone handle your data?

AI prepares, you approve. Backbone holds your receipts and invoices and lets ChatGPT or Claude work on them through the AI connector. You sign in with OAuth, so no password or key is pasted into the AI app. You approve what the AI may use, and the AI works with your role: before it sends or deletes anything, it asks you first; bank details change only in the app. The workspace owner can switch AI connections off in Settings > Integrations, and revoke a connected app there at any time.

The connector does not replace the points above. What the AI reads travels to your AI provider, so use a plan with a DPA for it, too. Setup: ChatGPT, Claude. More on the idea: What is MCP?