Under UK GDPR you can use ChatGPT and similar tools with customer information, but only if you have a lawful reason, tell people, put in no more than you need and have the right contract with the provider.
The simplest safe rule is to keep personal data out unless you have decided, in writing, that it should go in. Most everyday uses, such as drafting a letter or summarising a public document, need none.
This guide is for owners and managers who know their staff are already using AI tools. It covers when the law applies, the questions the ICO says to answer, the contract you need, and a three-colour policy you can give your team today.
Does UK GDPR apply when you use ChatGPT?
It applies whenever someone on your team types personal data into the tool, because that is processing, and your business is the controller of it.
Personal data means anything that identifies a living person, alone or with other information. A name and email address obviously qualify.
So does a case description that someone could trace back to a person, an employee's performance notes or a photo of a customer. If none of that goes in, UK GDPR is not triggered by the act of using the tool.
The first control is therefore not a technical one: you need to know what your staff paste in. A short conversation or an anonymous survey will usually show you more than a policy document does.
LegalVision's UK guide to ChatGPT privacy makes the same starting point: the rules bite when personal data is entered, so you begin by establishing what is being entered.
Its summary of the privacy considerations of using ChatGPT also notes that free and entry-level consumer plans may use inputs for training, and that business plans with data controls can be a safer choice.
What does the ICO say you must be able to answer?
The ICO has set out eight questions for anyone using generative AI with personal data, and each one needs an answer before you start.
They come from an April 2023 post by Stephen Almond, the ICO's Executive Director for Regulatory Risk, titled generative AI: eight questions that developers and users need to ask. In plain terms:
- What is your lawful basis for processing the personal data?
- Are you a controller, joint controller or processor?
- Have you carried out a data protection impact assessment (DPIA)?
- How will you be transparent with the people concerned?
- How will you manage security risks?
- How will you limit unnecessary processing?
- How will you handle requests from individuals about their data?
- Will the tool make solely automated decisions about people?
For a small firm, the answers do not need to run to pages. A paragraph for each, kept somewhere findable, is a reasonable record. The DPIA is the one that catches people out: the ICO says risks should be assessed before processing begins, and the assessment updated as the use changes.
Question eight matters more than it looks. Using AI to draft an email is one thing. Using it to score applicants, approve credit or decide who gets a service is a decision about a person, and it raises separate rules.
Do you need a contract with the AI provider?
Yes, if the provider handles personal data on your behalf, UK GDPR requires a written contract with it.
When you put customer details into a hosted AI tool, the provider is normally acting as your processor.
The ICO says a controller-processor contract must cover the subject matter and duration of the processing, the types of personal data involved, and the processor's duties on confidentiality, security, sub-processors, help with individuals' rights and breach notification, and deleting or returning the data at the end.
The full list is in the ICO guidance on what the contract must include.
In practice, this is a reason to prefer business plans over personal accounts. The major providers generally offer data processing terms on their business plans, so ask for them.
A personal account used for work usually does not, and the contract you accepted when you signed up may allow the provider to use your inputs to improve its models.
Read the current terms for the exact plan you use. Providers change them often, and a setting that was off last year may not be now.
Also ask where the data is processed. If it leaves the UK, the transfer needs a recognised safeguard, and a reputable provider will state which one it relies on.
What can staff put into an AI tool?
A traffic-light rule works best: green for anything public or non-personal, amber for business-confidential material in an approved tool, and red for personal data that needs a decision first.
Green means go ahead. Public web pages, your own published marketing copy, generic questions, anonymised examples with all identifiers removed, and drafts that mention no real person.
Amber means approved tools only. Internal documents, pricing, contract templates and meeting notes with no personal detail. Use a business plan with training switched off and a processing agreement in place.
Red means stop and ask. Customer or patient records, case files, ID documents, health information, children's data, HR files, anything with a name next to a sensitive fact. Do not enter these in a consumer account under any circumstances. Decide first whether the tool should see them at all.
Anonymising properly takes more than deleting a name. A job title, a small town and a date of an unusual event can identify a person. If someone who knows the individual could work out who it is, treat it as personal data.
What if someone has already pasted customer data into a personal account?
Treat it as a possible personal data breach: find out what went in, ask the provider to delete it where you can, and record what you decided.
Under UK GDPR, if a breach is likely to result in a risk to the people affected, you must tell the ICO within 72 hours of becoming aware of it. A name and an order number pasted into a drafting tool is unlikely to meet that bar. A set of patient notes or a client's financial details might.
Whatever you decide, write it down. The ICO expects you to keep a record of breaches and the reasoning, including the ones you judged not worth reporting. Then close the gap that allowed it, usually by giving the person an approved tool and a clear rule so the same shortcut does not look attractive next week.
Do this without blame. Staff who fear being disciplined stop telling you what they use, and that leaves you with the worst of both worlds.
Which sectors need extra care?
Any sector that handles health, legal or financial matters needs more, because those categories carry higher risk and stricter duties.
Health information is special category data under UK GDPR. A clinic that pastes a patient's notes into a chatbot to write a referral letter has made a serious decision, not a convenient one.
For healthcare practices, the safe route is a clinical system with AI built in under a proper agreement, or no personal data in the prompt at all.
Law firms and solicitors face a confidentiality duty on top of data protection. Client matters should not go into a general-purpose tool unless the firm has assessed the contract, the retention and the risk, and has the client's informed agreement where appropriate.
For anything involving special category data or a regulated profession, take advice from a data protection specialist.
What do you tell customers?
Tell them in your privacy notice if AI tools process their personal data, and say what for.
Transparency is one of the ICO's eight questions. A line saying that you use AI-assisted tools to handle enquiries, and who the providers are, fits easily into an existing notice.
If you run a chatbot on your website, put a clear message in the chat that the visitor is talking to an AI, and link to the notice. We cover the design side in when AI chatbots convert and when they annoy.
Do the same where an automated workflow reads incoming enquiries. The workflow tools themselves are processors too, so they need the same contract.
What should you do this week?
Five steps, in this order:
- Ask your team which AI tools they use, and what for, without blame.
- Pick one or two approved tools on business plans, with training off and an agreement signed.
- Circulate the green, amber and red rules.
- Write short answers to the ICO's eight questions for your main use.
- Update your privacy notice, and add a DPIA for anything involving sensitive data.
Our one-page AI audit template has a data column built in, so you can mark the risky tasks as you list them. If you plan to build agents that read customer emails or documents, the same rules apply at greater volume, with a human check before anything is sent.
For wider context on choosing a supplier, see what people mean by "AI agency". We build automations with the data questions answered up front, and you can see how on our AI automation services page or contact us to talk through your own set-up.
