You can use ChatGPT and other AI tools in a technical practice without breaching the GDPR, but not with just any account or any document. The practical rule is this: text with no personal data (regulations, blank templates, generic drafts) can go into almost any tool; documents with client data only into a business plan that doesn’t train on your data and comes with a data processing agreement; and some information (ID numbers, bank details, health data, third parties’ confidential documents) shouldn’t go into a general-purpose tool at all.

This guide explains what counts as personal data in a project, how personal and business accounts differ, how to classify documents, and what to set up before you start.

This guide is for information only and isn’t legal advice. Providers change their terms often: check them before you decide.

What personal data looks like in a project

More than you might think. In an architecture practice, an engineering firm or a construction company, personal data includes, among other things:

  • The client’s or owner’s name, ID number, address, phone number and email.
  • Names of workers, subcontractors and consultants in minutes and site reports.
  • Scanned signatures.
  • Site photos in which a person can be recognised.
  • A property’s address or land registry reference when it can be linked to its owner.

There’s also a special category that needs more care: health data. It turns up more often than you’d expect, for example in a project to adapt a home for a disability, or in the fitness certificates of a subcontractor’s workers.

If a document contains any of this, the GDPR applies to whatever you do with it, including pasting it into a chat.

Personal account or business account: the key difference

The same tool behaves differently depending on the type of account.

  • Personal accounts (free or individually paid). In many of them, what you type can be used to train or improve the model unless you turn that off in the settings. There’s usually no data processing agreement with your company, and each employee decides for themselves what to upload.
  • Business plans (for example ChatGPT Business or Enterprise, Claude for teams, Gemini in Google Workspace, or Microsoft 365 Copilot with a work account). Generally, the provider commits not to train on your data, offers a data processing agreement, and the company controls who has access and what is kept.

If people in your practice use ChatGPT on personal accounts to write minutes that include clients’ names, that’s the first problem to fix, before anything else.

What the GDPR requires when you use AI with client data

When you upload your clients’ personal data to an AI provider, that provider processes the data on your company’s behalf. In GDPR terms it’s a processor, and that means three things:

  1. A data processing agreement (Article 28 GDPR). Business plans usually include one; accept it and keep a copy.
  2. Knowing where the data is processed. If the provider processes data outside the European Union, it needs a valid transfer mechanism, such as certification under the EU-US Data Privacy Framework or standard contractual clauses. Some providers also offer data hosting in Europe.
  3. Using only the data that’s needed. That’s the data minimisation principle: if writing the minutes doesn’t need the client’s ID number, it doesn’t get uploaded.

It’s also worth checking your contracts with your own clients: if you committed to keeping their documents confidential, uploading them to an outside tool may need their permission, even if they contain no personal data.

Traffic lights: what you can upload and what you can’t

A simple way to explain it to the team:

Examples Which tool
Green Regulations, technical articles, blank templates, your own text with no client data, an anonymised document Any, preferably with training turned off
Amber Minutes, reports, specifications and emails with names of clients or workers, project documentation Business plan only, with a data processing agreement and no training
Red ID numbers, bank details, health data, passwords, a client’s security drawings or documents, third-party documents under confidentiality Not in general-purpose tools. Only in a dedicated system, with permission and agreed safeguards

When in doubt, treat it as amber.

How to anonymise before uploading

Many tasks work just as well without personal data:

  • Replace names with roles: “the client”, “the contractor”, “house A”.
  • Remove what adds nothing: ID numbers, phone numbers, bank accounts, signatures.
  • Generalise the location: “detached house in an urban area” instead of the full address, if the address isn’t needed for the task.
  • Check metadata and photos. A PDF can carry the author’s name, and a photo can show faces or number plates.
  • Don’t upload whole folders. Upload only the document the task needs.

Anonymising by hand every time is tedious. If a task involving personal data comes up every week, it pays to automate it in a system built for that, rather than asking the team to clean documents by hand.

What to set up before you start

In the tool:

  • Turn off the use of your conversations for training if you use a personal account. In ChatGPT it’s in the data controls settings; temporary chats aren’t used for training either.
  • Check memory and history. Some tools remember information between conversations; decide whether you want that.
  • On business plans, review the admin console: who has access, how long data is kept, and which connectors (email, Drive, SharePoint) are switched on.
  • One account per person. Shared accounts make it impossible to know who uploaded what.

In the company, a one-page usage policy. It’s enough to answer these questions:

  1. Which AI tools are allowed, and with which account?
  2. Which tasks are they used for?
  3. Which data is never uploaded? (The red light.)
  4. Who reviews what the AI produces before it’s sent or signed?
  5. Who do people ask when in doubt?
  6. When is the policy reviewed? (Once a year is enough.)

This policy is also one of the AI literacy measures the EU AI Act asks of companies that use AI; we cover that in the EU AI Act for architecture practices and builders.

How we do it

In the tools we build and in NotaFiel, customer data is hosted in the European Union, never used to train AI models, and every provider is bound by a data processing agreement. You can see the list of providers, and what each is used for, in our Trust Centre. We sign a confidentiality agreement before we see any of a client’s documents.

We don’t mention this as a sales point, but because it’s the minimum you should ask of any AI provider, whether that’s us or anyone else.

Frequently asked questions

Is it illegal to use free ChatGPT in the practice?

Not in itself. The problem arises when clients’ personal data or confidential information is uploaded to an account that has no processing agreement and may use the data for training. For text without personal data, a free account with training turned off is reasonable.

Do I have to tell my clients I use AI?

If you process their personal data with an AI provider, that provider should be listed as a processor in your privacy information, just like your email or cloud storage provider. Being open with clients about how you work also tends to build trust.

What about site photos?

If people or number plates can be recognised in them, they contain personal data. Assessing progress on site almost never requires anyone to be visible: choose or crop photos before uploading them.

Are business plans completely safe?

They reduce the risk a lot, but don’t remove it: it still depends on who has access and what gets uploaded. That’s why the usage policy and human review matter as much as the tool.

What do I do if someone has already uploaded client data to a personal account?

Delete those conversations, turn off training on that account and move that work to a business plan. If what was uploaded includes sensitive data or could affect clients, ask your data protection adviser what else needs to be done.