I've been in IT since 1998, and most of the data protection problems I've seen weren't caused by clever attackers. They were caused by the wrong account, the wrong setting or a contract nobody read. AI hasn't changed that. It has made it faster, because AI tools now plug straight into your email, files and accounts.
The legal examples here use UK and EU GDPR. If you operate elsewhere, check the rules there, but the questions are the same.
Is AI GDPR compliant?
No AI tool is compliant on its own. It depends on how you use it: which account, which data, on what terms. If you decide what customer data goes into an AI tool, you're responsible for it, even when the tool belongs to Google, Microsoft or OpenAI.
If the AI provider handles your data only on your instructions, it's your processor. Check its terms say so. The law then requires a written contract between you, with specific terms covering security, confidentiality, sub-processors and what happens at the end. The ICO lists them in its guidance on what a processor contract must include. For most AI tools that contract is the provider's data processing agreement, often called a DPA. Check you've actually accepted or signed it.
On 20 October the ICO reprimanded an organisation partly because the contract with its IT provider lacked those terms. We covered the case in this week's AI news.
Which AI accounts are safe for business data?
Business plans with a data processing agreement. Not personal or free accounts. The terms really are different.
OpenAI's enterprise privacy page says it doesn't train its models on business data by default for ChatGPT Business, Enterprise and its API, and offers a DPA for those plans. Personal plans aren't on that list. Google's Gemini privacy help says that on personal accounts some chats are read by human reviewers while Keep Activity is on, and asks people not to enter anything confidential. It notes that work accounts come with different terms.
That matters now Google is rolling out small business tools in the Gemini app that connect to Xero and Shopify. Plenty of sole traders will connect them to a personal Gmail.
What data should never go into an AI tool?
Some data shouldn't go into any AI tool without a specific reason and a proper review. Our default list:
- Passwords, card numbers and bank details. There's no good reason to put these into an AI chat.
- Health and other special category data, and criminal offence data. The law treats these as higher risk, and large-scale use triggers extra duties.
- Anything about children. In the ICO case above, most of the 3,300 people affected were children.
- Other people's confidential material. Client contracts, supplier pricing and anything under an NDA.
For anything else, ask why the tool needs it and who will be able to see it.
Do you need a DPIA for AI?
Sometimes. A DPIA, or data protection impact assessment, is a written check of the risks before you start. The law requires one where processing is likely to result in a high risk to people.
The ICO's guidance on when you need a DPIA lists warning signs. They include using new technology, evaluating or scoring people, automated decisions with serious effects and sensitive data. It says two of those together usually means you need one, and if you're in doubt, do one. An AI tool that drafts marketing copy probably doesn't need one. Screening job applicants with AI is a strong reason to check whether you do.
What should you check before connecting AI to your email or accounts?
Treat each one like a new member of staff. It needs an account, a clear job and only the access that job requires.
Assistants are starting to work across all of these at once. Claude Duet, in a US waitlist preview on Claude's individual Pro and Max plans, checks your inbox, invoices and calendar every few hours and drafts work for you. Approval before it sends or pays is built in. The data protection question comes earlier: should it be reading that data at all? Connect one app at a time and remove what nobody uses. Our checklist for giving an AI agent access covers the detail.
How we handle data on client builds
Every engagement starts with four written answers, agreed in the statement of work before anything is built. Where does the data live? Who can access it? How long is it kept? What leaves your environment?
Take the AI service desk we're building for a US managed service provider, which is in delivery now. Support tickets are full of personal data: names, phone numbers and whatever a caller decides to type. So the access questions came first, and the AI triages tickets against rules the client agreed, with tight limits on what it can change. We build on your existing systems and permissions rather than making a parallel copy of your data. We use the providers' commercial routes, never consumer tools, and we don't use client data to train models. Our trust and security page sets out the rest.
The decision we'd make for most owners is simple. If a job involves personal data, it goes through a business account with a signed DPA and a named owner, or it doesn't go through AI.
AI data protection checklist
- Every AI tool used for work is on a business plan, not a personal account.
- Each provider's data processing agreement has been accepted or signed, and you know where it's filed.
- You've checked whether each provider trains on your data, and switched it off where you can.
- Staff know what must never go into AI: passwords, payment details, special category data and confidential client material.
- Anything that scores or decides about people has been screened for a DPIA.
- Connected assistants can only see what their job needs, and someone reviews the list monthly.
- Your privacy notice says how you use AI with customers' personal data.
- Your IT provider's contract says who looks after the security of the AI tools they set up.
Where we fit
Start by listing the tools your staff use, then check the accounts, terms and connections with your IT provider. The harder part is building AI that does a real job with the right access from day one. That's what we do, and we run it afterwards so the access doesn't drift.
The AI Readiness Diagnostic is free and no-obligation. We look at how you work, pick the jobs worth doing with AI and hand you a statement of work, a wireframe and an ROI case you keep whether or not we build anything. If you're an IT provider, our AI for MSPs page shows how we work alongside you.
Common questions
Does turning off AI training make customer data safe?
No. It stops your data being used to improve the provider's models, but the data is still processed and stored. Google, for example, keeps Gemini chats for 72 hours even with Keep Activity off. You still need a business account, a data processing agreement and limits on who can see what.
Who is responsible if an AI tool leaks customer data?
It depends on each party's role and what went wrong. The business that decided to use the tool has duties, and so does the provider if it's your processor. The ICO's October 2026 reprimand of a client and its IT provider shows responsibility can sit with both. A clear contract saying who does what protects you.
Do I need to tell customers I'm using AI?
If AI processes their personal data, your privacy information should explain what you use it for and who it's shared with. The ICO says people must be told about new uses of their data before that processing starts. In practice that usually starts with updating your privacy notice.
See it in practice: AI service desk for a US managed service provider →