Whenever we talk to a UK business owner about AI agents, the same question comes up within the first ten minutes. "What about our data? What about GDPR?" It is the right question to ask, and the answers are clearer than most people think.
This is a practical guide. Not legal advice (always talk to your lawyer or DPO for the formal version), but a working framework for thinking about AI data safety in a UK business.
The five things UK GDPR cares about with AI
You don't need to memorise the regulation. You do need to be able to answer five questions about anything that processes personal data in your business, AI included:
- What personal data is being processed?
- Where is it being processed (which countries, which providers)?
- What is the legal basis (consent, legitimate interest, contract performance, etc.)?
- Who has access?
- How long is it kept?
If you can answer those five questions for any AI agent in your business, you're 90% of the way to being defensible.
Where your data actually goes
This is the bit most owners are unsure about. When you use a model from OpenAI, Anthropic, Google, Mistral, or any of the others, what happens to the data you send?
The honest answer depends on which tier you're on. The safe ones for business use share a few common features:
- Your data is not used to train future models. All major providers offer this on their business and enterprise tiers. Make sure you're on the right tier.
- Data is processed in regions you've agreed to. EU and UK data residency is available with most providers, sometimes for an extra fee. If you handle EU citizen data, you may need this.
- You have a Data Processing Agreement in place. This is standard with business tiers and required by GDPR.
If you're using consumer ChatGPT to handle client data, none of those guarantees apply. Stop doing that.
How the major providers compare
A simplified view of the providers we work with most often:
- Anthropic (Claude). Business and enterprise tiers do not train on your data. Data residency options available. Strong default privacy posture.
- OpenAI (GPT models). API and business tiers do not train on your data by default. Data residency available. Consumer ChatGPT does train on your data unless you opt out.
- Google (Gemini). Workspace and Vertex AI tiers do not train on your data. EU residency available.
- Microsoft (Azure OpenAI). Excellent enterprise privacy posture, EU regions, very controllable.
- Mistral. European, GDPR-friendly defaults, EU hosting available.
None of these is automatically "the right one". The right one depends on what you're building, where your customers are, and your existing infrastructure.
Practical safeguards for your AI agent
Beyond the choice of provider, the agent itself needs to be designed safely. The basics:
Minimise data exposure
Send the agent only the data it needs to do its job. If it's drafting a customer reply, it doesn't need the full client database. If it's processing an invoice, it doesn't need access to your salary data. Scope tight.
Redact what you don't need to send
For some workflows, sensitive details (NHS numbers, payment data, full addresses) can be redacted before they reach the model and reinserted afterwards. Worth designing in from day one.
Log everything
Every prompt and response should be logged, with timestamps and user attribution. Not because you'll review them all, but because you need them when you have to.
Match access to your existing permissions
If only HR can see salary data in your file system, the agent shouldn't be able to surface it to anyone else. Build the existing access controls into the agent design.
Get a human in the loop for sensitive actions
Anything client-facing or compliance-sensitive should be reviewed by a human before it goes out. Not forever, but at least until trust is established.
GDPR doesn't ban AI. It just asks you to be deliberate about how you use it. Most of the work is upfront design, not ongoing burden.
Questions to ask any AI vendor
Whether you're talking to us or anyone else, these are the questions worth asking before you sign anything:
- Which model providers will you use, and on which tiers?
- Where will our data be processed and stored?
- Will our data ever be used to train any model?
- What's your Data Processing Agreement, and is it GDPR-compliant?
- How is access controlled to the agent and its data?
- What happens to our data if we cancel?
- Do you have a working incident response plan if something goes wrong?
If a vendor stumbles on those, that tells you something important.
The DPIA question
For most agent projects, you'll want to do a short Data Protection Impact Assessment. This is not as scary as it sounds. It's a structured one-pager describing what data you're processing, why, where, and what you're doing to keep it safe. Your lawyer or DPO will have a template.
If your business is too small to have a DPO, the ICO's website has good free guidance and templates aimed at SMEs.
How we approach this in our builds
We build every agent with UK GDPR baked in from day one. That includes choosing the right provider tier, configuring data residency, designing access controls, redacting sensitive fields where appropriate, and writing clear documentation that you can hand to your DPO or your auditors.
This applies most visibly to private internal knowledge agents, where access control and data residency really matter, but the same principles run through every agent we build.
It's not glamorous work, but it's the difference between an AI project you can defend and one that becomes a liability. If you'd like to talk through the data safety side of a project you're considering, that's a conversation we have all the time.