Is ChatGPT GDPR-safe with client data? How we set it up

Short answer: yes, you can use ChatGPT with client data in a GDPR-safe way, but only if the safety sits in the setup and not in people's good intentions. For an advice practice or an AI accounting firm setup, that means: your own business account, read-only connections to your systems, EU hosting for anything in between, no client content in logs, a data processing agreement, and a person who approves before anything leaves the building. In Dutch this whole question usually comes in as "is ChatGPT AVG-proof?". Same question, same answer.

One thing first. I'm not a lawyer. This note is about design: what I build, and why. Whether your specific setup is compliant is a question for your privacy officer or legal advisor. What I can tell you is what makes their job a lot easier.

The real risk is the upload button

Most firms I talk to are already using ChatGPT with client data. They just don't call it that.

Rein Teune at Landsman & Teune (a mortgage advice practice plus a bookkeeping office in Kampen) is a good example. Before we started, he wrote advice letters like this:

1. Download the rate offer from the dossier.
2. Download the purchase contract, the financing overview and the application.
3. Upload them one by one into ChatGPT.
4. Prompt, check, copy, paste.

It worked. But every upload was a document full of personal data going in by hand, with nobody deciding what exactly went in, and no record of it.

So the honest starting point isn't "should we allow AI?". It's "it's already happening, so can we make it controlled?". A connector that fetches only what's needed is often safer than a person dragging PDFs into a chat window.

How we set up ChatGPT with client data, GDPR-safe

Landsman & Teune works in Accelerate (advice), Nmbrs Accounting (around 246 active client administrations) and a shared Outlook inbox. I built a connector so their own ChatGPT Business can read those three systems. Here is what makes it hold up on privacy, piece by piece.

• Their own business account. All the AI work happens in their ChatGPT Business account, under their own contract with OpenAI. A business account where data isn't used for training is step one. My server makes no AI calls at all. It fetches, converts to text and hands it over.
• Read-only everywhere. The connector can search and read dossiers in Accelerate and books in Nmbrs. It can't change a dossier, can't book anything.
• No send button. In Outlook it can read and prepare a draft. It cannot send. That permission simply isn't granted, so it doesn't exist.
• Only what's needed. Outlook access is limited to two shared inboxes, not every mailbox in the firm.
• EU hosting. The connector runs on a server in Germany.
• No content in logs. The log records who used which tool on which dossier, and whether it worked. Never the content. It's deleted after 30 days.
• Text recognition on the same server. Most client documents are scans. Instead of sending them to yet another service for OCR, that runs on the same box. A payslip takes about 10 seconds.
• A data processing agreement. One agreement covers both companies, with a single sub-processor (the hosting party, in the EU). OpenAI, Microsoft, Accelerate and Nmbrs are processors the firm already had a direct relationship with.

None of these are clever on their own. Together they mean the rules are enforced by the software. That's the bit I spent the most time on.

Human approval before anything leaves

Reading is one thing. The moment AI output goes to a client, you want a person in the loop. Not as a policy line, but as a hard gate.

For the advice letters, Rein and ChatGPT had already written their own method and Word template. I built the service around it. ChatGPT prepares the input, the service renders the letter in their template and runs their own checks, and then a person approves it on a separate page.

That approval is tied to a fingerprint of the exact input and the exact document. Change one number and it's a new version that needs a new approval. An approval filled in by the AI gets refused. A letter without amounts is blocked.

The human releases the letter. Never the AI.

What you can connect (and how to do this yourself)

The pattern isn't specific to mortgages. It works through MCP, an open standard that ChatGPT, Claude and others support, so you're not locked into one AI vendor.

What firms like this typically connect:

• Practice and CRM systems: Accelerate, AdminPulse, Simplicate, HubSpot, SuperOffice.
• Accounting and payroll: Nmbrs, Twinfield, Exact Online, AFAS, Snelstart.
• Mail and documents: Outlook, SharePoint, Google Drive, scanned PDFs via text recognition.

If you want to try this yourself, this is the checklist I'd use:

1. Start with a business account, not someone's personal ChatGPT login.
2. Default to read-only. Add writing later, and only behind a human approval.
3. Scope access per mailbox, per administration, per user. Not "everything".
4. Keep the middle layer in the EU and know exactly who your sub-processors are.
5. Log actions, not content. You want an audit trail without building a second copy of your client data.
6. Ask your vendor, not ChatGPT, what their API can do. When we started I told Rein Accelerate had no API to read dossiers, because the public docs suggested so. I asked Accelerate directly and there was a full API behind the licence. I corrected myself in writing.
7. Put it on paper in a data processing agreement, and run it past whoever does privacy for you.

When this is not the right approach

Honestly, sometimes it isn't.

• If nobody in the firm uses AI yet, start with the business account and some sensible habits. A connector solves a problem you don't have.
• If you need the AI to make decisions on its own, this setup won't do that, on purpose. Everything that leaves goes past a person.
• If your systems have no API, you're stuck with exports or uploads. Ask the vendor before you give up, though.
• If you're looking for a legal stamp, I can't give you one. I can build something your privacy officer will find easy to sign off on. That's different.

If you're curious how the full build looks, read the Landsman & Teune case, or browse other work.

And if you're wondering what this would look like with your own systems, grab 30 minutes with me. No pitch, just a look at what you've got.

Is it GDPR-compliant to use ChatGPT with client data?

It can be, if you design for it: a business account where data isn't used for training, a connector that only reads what's needed, EU hosting, no content in logs and a data processing agreement. I'm not a lawyer, so have your privacy advisor check your specific setup. The point is that the limits are enforced by software, not by good intentions.

Is ChatGPT AVG-proof for an accounting firm or advice practice?

AVG is simply the Dutch name for GDPR, so the answer is the same. A controlled connector with read-only access and human approval is usually safer than staff uploading PDFs into a chat by hand.

Can you connect ChatGPT to Accelerate or Nmbrs?

Yes. Accelerate has an API behind the licence, and Nmbrs Accounting has one too. With a connector, ChatGPT Business can search and read dossiers, mortgage details, documents, balances, transactions and the document archive. Nmbrs limits calls per day, so a good connector caches what it can.

Can the AI send emails or change client files?

Not in the setup we built. It can prepare a draft, but sending and writing are technically impossible because those permissions are never granted. Every advice letter needs a person's approval before it leaves.

Does this work with Claude or Microsoft Copilot instead of ChatGPT?

Yes. The connector uses MCP, an open standard that ChatGPT, Claude and others support. The privacy rules live in the connector, so they stay the same whichever AI you use.