UK GDPR and AI tools at work: a plain-English guide to ChatGPT GDPR risk for businesses
The ChatGPT GDPR risk for businesses comes from six UK GDPR duties, not from AI itself: lawful basis, data minimisation, controller and processor roles, international transfers, breaches and accountability. Each one applies the moment staff paste personal data into ChatGPT or Copilot. Each one also has a plain control that meets it. This guide takes them in turn, then maps them in one table.
AI is welcome at work. Client data needs a reason to go into it. That is the whole idea.
Who this guide is for
Owners, practice managers and data protection leads at small firms. Accountants, solicitors, advisers, recruiters, agencies. Anyone who has to explain, in plain words, what the law asks when staff use AI tools.
It sticks to well-established principles. It is not legal advice.
1. Lawful basis
Every use of personal data needs a lawful basis. Contract, legal obligation and legitimate interests are the usual ones for a small firm. Sending client data to an AI tool is a use of that data. So it needs a basis too.
Guidance summarised by AI Governance puts it plainly: any organisation processing personal data of UK individuals using AI systems must be able to identify and document the lawful basis for each processing activity, including where AI is used in training pipelines or inference.
In practice, that means writing down what staff may use AI for. Drafting a letter to a client you act for may sit within your existing basis. Pasting a whole client list to "see what it says" probably does not.
Control: a short written AI policy listing approved uses, and a dated record that each person has agreed it.
2. Data minimisation
Use only the personal data you need. Nothing more. This is the principle staff break most easily, because pasting the whole email is quicker than trimming it.
Most AI questions do not need the name, the bank details or the National Insurance number. "Rewrite this reminder about invoice 4471" works just as well as a stand-in such as "Client A".
Rule of thumb: if the answer would be the same without the personal detail, the detail should not go in.
The duty sits on both sides. The ICO says a developer or deployer remains responsible for assessing and mitigating the risk that personal data used to train your models may be exposed as a result of the way your clients have deployed the model. Your firm still owns its own choices about what goes in.
Control: teach staff to strip details first, and back that up with a tool that takes them out before sending. Our guide on how to redact sensitive data before pasting into ChatGPT covers the manual way.
3. Controller and processor roles
Your firm is the controller of its client data. You decide why and how it is used. An AI provider used on a proper business account usually acts as your processor, working on your instructions under a contract.
That contract matters. A business plan normally comes with data processing terms. A personal free account signed up with a work email usually does not, so the provider's own consumer terms apply instead.
Training is part of this. For business ChatGPT plans, OpenAI Help Centre states that OpenAI won't train on your workspace's data, and that business data is excluded from training by default and is encrypted in transit and at rest. For more on plans, see ChatGPT Team vs Enterprise vs free.
Control: approve named business accounts, keep their data processing terms on file, and steer staff away from personal accounts.
4. International transfers
Most AI providers are based outside the UK. Sending personal data to them can be a restricted transfer. UK GDPR allows it with the right safeguards, such as an adequacy decision or approved contract terms.
Business plans from the large providers usually set out how they handle this. Free tools, and newer tools from unfamiliar providers, may say little. If you cannot find the answer, treat the tool as unapproved.
Control: check where each approved tool processes data and what safeguards it relies on. Keep a list of tools that are not approved, and know which ones staff actually use.
5. Breaches
Personal data going somewhere it should not can be a breach. Not every paste is one. It depends on what went in, where, and the likely risk to the people involved.
Where a breach is likely to risk people's rights and freedoms, the clock starts. The ICO example is clear: you notify the ICO within 72 hours of becoming aware of the breach, explaining that you don't yet have all the relevant details, but that you expect to have the results of your investigation within a few days. Article 33(4) allows you to provide the required information in phases, as long as this is done without undue further delay.
You cannot judge a breach you never knew happened.
That is the practical problem with AI tools. A paste leaves no trace in your own systems unless something records it. We cover the judgement in more depth in is pasting a client email into ChatGPT a data breach.
Control: a record of what kind of data went to which AI site, when and by whom, plus a named person who decides whether to report.
6. Accountability
You must be able to show you comply, not just say so. The ICO's strategic approach notes it has developed an overarching Accountability Framework and AI-specific guidance on accountability that it continues to update.
For a small firm, evidence is simple. A policy. Signed agreements. A list of approved tools. A training record. A log of what was blocked or taken out. Clients, insurers and auditors increasingly ask for exactly this.
Control: keep these records in one place and review them monthly.
Summary table: principles and controls
| UK GDPR principle | What it means for ChatGPT and Copilot | Practical control |
|---|---|---|
| Lawful basis | Each AI use of personal data needs a documented reason | Written AI policy with approved uses, dated staff agreement |
| Data minimisation | Only the personal data the task needs goes in | Strip details first, stand-ins such as "Client A", automatic removal |
| Controller and processor | Your firm is controller, provider should be processor under contract | Business accounts only, processing terms on file |
| International transfers | Data sent abroad needs safeguards | Check each tool, list unapproved tools, find what staff use |
| Breaches | Risky disclosures may need reporting within 72 hours | Record of AI sends, named person to decide on reporting |
| Accountability | You must show compliance | Policy, agreements, training and logs kept together |
Worked example: Priya at Ashcombe wants a polite chaser for Fenwick Joinery. She asks with "Client A" in place of the name. The question still goes. No client data left the firm.
Where a tool helps
Several of these controls can run by hand. Minimisation and breach records are the hard ones to keep by hand. Helio Guard checks each paste, message and file upload on listed AI sites, checked on the PC with fixed rules. It takes out or stops items such as bank details and National Insurance numbers, and records agreement to your AI policy. We never see what your team types. It is £1 a person a month, from 5 people, with a 14-day free trial and no card needed. The pricing page has the detail.
Common questions about ChatGPT and UK GDPR
Is it illegal for staff to use ChatGPT with client data?
Using ChatGPT with client data is not illegal in itself under UK GDPR. It becomes a problem when there is no lawful basis, when more personal data goes in than the task needs, when no processing contract is in place, or when the firm cannot show what happened. A business account and a clear policy address most of that.
Does Microsoft Copilot raise the same GDPR issues as ChatGPT?
Microsoft Copilot raises the same six UK GDPR questions as ChatGPT: lawful basis, minimisation, roles, transfers, breaches and accountability. The answers can differ because Copilot within a business Microsoft 365 account comes under your existing Microsoft terms. The free consumer version of Copilot does not. Check which version staff are actually using.
Do small firms need a DPIA before using AI tools?
A data protection impact assessment is required under UK GDPR when processing is likely to result in high risk to individuals. Routine drafting with personal details removed may not reach that level. Feeding sensitive client records into an AI tool more likely would. Many small firms write a short assessment anyway, because it doubles as accountability evidence.
Related: AI policy for employees: a UK template you can adapt today, Is it safe to put client data into ChatGPT? What every firm should know