AI tools are only useful when they work with real information, and real information usually includes personal data: customer emails, CVs, call recordings, invoices, support tickets. That's where GDPR comes in. In the UK, it's the UK GDPR and Data Protection Act 2018, updated by the Data (Use and Access) Act 2025. In the EU, it's the GDPR, now alongside the EU AI Act.

This guide is practical, not legal advice. For anything high-risk, check the ICO's (UK) or your national regulator's guidance, or speak to a specialist.

The core rules, applied to AI

GDPR principleWhat it means for AI use
Lawful basisYou need a reason to process the data (contract, legitimate interests, consent, legal obligation)
TransparencyYour privacy notice should say you use AI tools and for what
Purpose limitationUse data for the purpose you collected it for
Data minimisationOnly give the AI the personal data it actually needs
AccuracyCheck AI outputs about people; wrong outputs are inaccurate data
Storage limitationDon't keep prompts, transcripts and outputs forever
SecurityChoose secure tools, control access, use business accounts
AccountabilityDocument your decisions

Step 1: Map where AI touches personal data

List every AI use in your business:

  • Assistants staff use (ChatGPT, Gemini, Claude, Copilot).
  • AI features inside existing software (CRM, helpdesk, accounting, email).
  • Chatbots and voice agents talking to customers.
  • Meeting recorders and transcription.
  • Recruiting tools screening CVs.
  • Automations sending data to AI models.

For each, note what personal data goes in, who the provider is and where data is processed.

Step 2: Use business plans with proper terms

Personal or free accounts are the biggest risk. Business plans from major providers typically offer:

  • A data processing agreement (DPA), making the provider your processor.
  • No training on your data by default.
  • Admin controls, retention settings and security certifications.

Check international transfers too. If data leaves the UK or EEA, the provider should rely on an adequacy decision or approved transfer mechanisms such as standard contractual clauses.

Step 3: Choose your lawful basis

For most business uses (summarising support tickets, drafting replies, meeting notes), legitimate interests or contract usually apply. Do a short legitimate interests assessment: what's the benefit, is AI necessary for it and does it override the person's interests?

Be more careful with:

  • Special category data (health, ethnicity, religion, biometrics, trade union membership). Avoid sending it to AI tools unless you have a clear legal condition and strong safeguards.
  • Children's data.
  • Recordings of calls and meetings, where people should be told in advance.

Step 4: Update your privacy notice

Add plain-language lines such as:

We use AI tools from trusted providers to help answer customer messages, summarise meetings and draft documents. These providers process data on our behalf under contracts that protect your information and don't allow them to use it to train their models. A member of our team reviews AI-assisted responses before decisions are made about you.

Adjust to what you really do.

Step 5: Minimise what goes in

  • Remove names, emails and account numbers when the task doesn't need them ("summarise the themes in these complaints" doesn't need customer names).
  • Don't paste whole databases into chats.
  • Configure integrations to send only the fields required.
  • Set short retention for prompts and transcripts where the tool allows it.

Step 6: Keep humans in decisions about people

GDPR restricts solely automated decisions that have legal or similarly significant effects on people, such as rejecting a job application or a loan, without meaningful human involvement. The UK's 2025 reforms relaxed some of these rules but still require safeguards such as the right to human review and to contest the decision.

The safe approach: AI can summarise, sort and suggest. A trained person decides. Read AI recruiting done fairly.

Step 7: Do a DPIA when risk is high

A data protection impact assessment is required for high-risk processing: new technology on a large scale, systematic monitoring, profiling or decisions about people. Even when not required, a one-page assessment of each AI tool (what data, why, risks, safeguards) is cheap insurance.

Step 8: Handle people's rights

People can ask for access to, correction of or deletion of their data. Make sure you can find data held in AI tools: chat histories, transcripts, generated notes. Choose tools that let you search and delete.

Step 9: Train staff and write a policy

Most AI data incidents start with a well-meaning employee pasting something into the wrong tool. A short policy covering approved tools, what not to paste and when to get approval fixes most of that. Our AI policy guide has a template.

A quick checklist

  1. List AI tools and the personal data each uses.
  2. Move all work use onto business plans with DPAs.
  3. Confirm no-training settings and data location.
  4. Record your lawful basis for each use.
  5. Update your privacy notice.
  6. Minimise personal data in prompts and integrations.
  7. Keep humans making decisions about people.
  8. Run a DPIA for high-risk uses.
  9. Make sure you can answer access and deletion requests.
  10. Train staff and publish an internal AI policy.

The short version

You can use AI tools with personal data lawfully if you use business plans with proper contracts, have a lawful basis, tell people in your privacy notice, share only what's needed, keep humans making decisions about people and document your choices. Avoid sending special category data, and run a DPIA for high-risk uses.

Our AI automation team builds AI workflows with data protection designed in, and tools like Zelvorm Assist and Zelvorm Notes are designed with these principles in mind.