UK GDPR and AI tools: what your business needs to get right
A practical guide to UK GDPR and AI tools: lawful basis, DPIAs, processor contracts, human oversight and what the ICO expects before you deploy.
If you put personal data into an AI tool, you are still the data controller, and UK GDPR applies exactly as it does to any other processing. That means you need a lawful basis, a written contract with the provider, a data protection impact assessment where the risk is high, and real human oversight of anything that affects a person. The ICO has been clear that AI does not create a legal grey area: existing data protection law already covers it.
We are not lawyers and this is not legal advice. What follows is the practical version we would give an operations lead who has just been handed a budget for AI and wants to know what to sort out first.
What does UK GDPR require when you use AI tools?
The core duties are unchanged. You must process personal data lawfully, fairly and transparently, collect only what you need, keep it accurate, retain it no longer than necessary, and keep it secure. Article 35 of UK GDPR requires a data protection impact assessment before processing that is likely to result in a high risk to individuals. The ICO says that in the vast majority of cases the use of AI involves processing likely to result in a high risk to people's rights and freedoms, which triggers that legal requirement (ICO guidance on AI and data protection). The ICO is reviewing that guidance after the Data (Use and Access) Act 2025, so check the current version before you rely on the detail.
A DPIA for an AI tool is not a formality. Done properly it covers the necessity and proportionality of the processing, your lawful basis, consultation with stakeholders, and AI-specific risks such as automated decision-making, profiling, bias and discrimination. It should also set out how you will apply meaningful human oversight, how you will be transparent with people about the processing, what the procurement contract says, and how you will review the arrangement as the tool changes.
The point that catches most businesses out is that AI tools change under you. A model update can alter how data is used or how outputs are produced. Build a review date into the DPIA rather than treating it as a one-off document.
What does the ICO expect from businesses using AI?
The ICO's position, in short, is that you cannot outsource your accountability by buying a tool. If you decide why and how personal data is processed, you are the controller, even when a supplier does the technical work. The provider is usually your processor, and that relationship needs a contract.
In practice the ICO's expectations come down to a few questions it would ask if someone complained or you had a breach:
- Do you know what personal data goes into the tool, and why?
- Can you show your lawful basis for each purpose?
- Have you assessed the risk before going live, and recorded it?
- Is there a written contract that limits what the provider can do with your data?
- Can a human review and challenge outputs that affect a person?
- Could you answer a subject access request or erasure request that touches the tool?
If any of those gets a shrug, that is your work list.
Is there a lawful basis for putting customer data into an AI tool?
You need one of the lawful bases in Article 6 of UK GDPR, and for any special category data you also need a condition under Article 9. Legitimate interests is often the practical choice for internal efficiency work, but it requires a legitimate interests assessment that you can show. Consent is a poor fit for most back-office AI use because it is hard to withdraw cleanly and awkward to evidence later.
Be honest about purpose. Training a model on customer emails to improve your own response quality is a different purpose from using AI to summarise a support ticket. If the purpose changes, your basis and your transparency information may need to change with it.
One line we would draw firmly: do not paste special category data, or anything you would not put on a postcard to the customer, into a consumer-grade tool. Employee health details, safeguarding notes, criminal records checks and detailed financial data deserve a deliberate decision, not a shortcut.
What should be in the contract with an AI provider?
A processor contract under Article 28 of UK GDPR should set out what the provider may and may not do with your data, where it is stored, how long it is kept, what happens on termination, and how breaches are reported to you. If your website collects the data in the first place, our guide to GDPR for small business websites covers the basics there. For AI tools, add three things that generic contracts often miss.
- Training use. Get in writing whether your data is used to train or improve models, and turn it off if you can.
- Sub-processors. AI stacks lean on third parties. You need to know who they are and how changes are notified.
- Transfers. If data leaves the UK, you need a valid transfer mechanism and a record of it.
Also check retention in the tool itself: how long it keeps conversation history, and whether you can shorten it. Anything kept longer than you need is a retention problem sitting inside your estate, and it is your problem, not the vendor's.

Where does human oversight actually matter?
Human oversight is the difference between an AI tool that helps and an AI tool that makes decisions about people without a person involved. Since the Data (Use and Access) Act 2025, Articles 22A to 22D of UK GDPR govern decisions made solely by automated means that have legal or similarly significant effects on a person. Such decisions are allowed in more situations than before, but only with safeguards: people must be told about the decision, be able to make representations, get a human to review it and contest it. Where special category data is involved, the rules are much tighter. A tool that scores job applicants, sets credit terms or decides a refund dispute is in that territory; one that drafts an email is not.
The GOV.UK summary of the Act's changes sets out the new framework, and the ICO's draft guidance covers the safeguards and when special category data can be used in automated decision-making.
Sensible oversight is not a tick box. Someone named should be able to see why the tool produced an output, challenge it, and override it. Log the overrides. If nobody ever overrides anything, that usually means nobody is looking.
Where does Varsuite fit?
We build bespoke AI agents for UK businesses, and we build them so that the data protection questions above have answers. That means working out which data the agent touches before we write any prompts, keeping personal data out of places it does not need to be, logging what the agent did and on whose instruction, and leaving a clear point where a person reviews the output.
We are Cyber Essentials certified, which covers a defined set of technical controls. It is a useful baseline and nothing more: no certification guarantees security or compliance.
Just as importantly, we are honest about where we are not the right fit. If you need a decision made about a novel processing activity, a complex international transfer, or an AI use that makes significant decisions about people with no human involved, you need a solicitor or a data protection specialist, not a development team. We will build to the design your advisers sign off, and we will say so when the question is above our pay grade.
What should you do first?
Before you buy anything, do four things in this order.
- List every AI tool already in use, including the ones people signed up for themselves.
- For each, write down what personal data goes in and what comes out.
- Decide whether a DPIA is needed, and do it if the answer is yes.
- Fix the contracts, retention settings and training-use flags that the DPIA exposes.
That sequence costs very little and prevents most of the problems. Buying a tool first and asking these questions afterwards is the expensive route, because by then data has already moved.
The sensible next step is a short conversation about what you actually want the AI to do, so you can scope it before any data moves. If you want to talk it through, our team can look at your specific use case and tell you plainly whether it is a build job, a compliance job, or both.
Frequently asked questions
Do I need a DPIA for every AI tool I use?
No, but you need to make that judgement deliberately rather than by default. A DPIA is required where processing is likely to result in a high risk to individuals, and a new AI tool that handles customer or employee personal data sourced from an external provider will often meet that bar. Record your reasoning either way. The ICO expects one in the vast majority of AI uses.
Can I use AI tools on customer data at all?
Yes, provided you have a lawful basis, a proper processor contract, appropriate security and a retention plan. The law restricts how you process personal data, not the technology itself. The mistakes that get businesses into trouble are using consumer tools with no contract and keeping everything forever.
Who is responsible if an AI tool gets something wrong?
As controller, accountability for the processing sits with you, even where a supplier provides the technology. That is why contract terms, oversight and logging matter: they are how you show you took the risk seriously.
Does Cyber Essentials make an AI tool compliant?
No. Cyber Essentials covers a defined set of technical controls and is a useful baseline. It does not assess your lawful basis, your DPIA, your contracts or your oversight arrangements, and no certification guarantees compliance or security.
Images from Pixabay
Jamie is Technical Director at Varsuite and leads the technical development team, setting how we design and build everything we ship. He builds the AI models that power our agents and manages the AI s...
More from Jamie WoodruffGet The Signal
Practical notes on AI, websites and automation for UK businesses. One useful email at a time, unsubscribe whenever.
Varsuite is an AI-accelerated, human-perfected digital production company based in Rishton, Lancashire. Agents build, people perfect: websites, stores, software and AI automation.
Ready to put AI to work?
Let our agents design, build and manage it for you.