← Back to Insights

AI Governance

Staff AI Use Policy: Governing ChatGPT and Claude at a UK Firm

Michael K. Onyekwere··9 min read

Staff at most firms can open ChatGPT or Claude in a browser and paste in a client email, a contract or a set of HR notes. Whether that is acceptable depends on the account they are using and on what they enter, and the firm remains responsible for the personal data in either case. A staff AI use policy is the document in which the firm decides those questions in advance and gives its staff the answer.

Why staff instructions should be recorded in writing

A firm may hold personal data as a controller, or as a processor acting for a client, and it often does both. Article 29 of the UK GDPR provides that a processor, and anyone acting under the authority of a controller or processor, may process personal data only on the controller's instructions, unless domestic law requires otherwise. Article 32(4) requires controllers and processors to take steps to ensure that the people working for them keep to this. Article 24(2) adds that a controller must implement appropriate data protection policies where these are proportionate to its processing.

These provisions do not expressly require an AI policy, or require staff instructions to be in writing. A written policy is nonetheless the practical way for a firm to tell its staff what the rules are and to keep a record of what they were told, which matters because Article 24(1) requires a controller to be able to demonstrate that its processing complies with the Regulation. The policy works alongside the technical and organisational measures that put its rules into practice, such as the choice of plan and the settings on the firm's workspace.

Which accounts are approved

The rule on accounts matters most, because a vendor's consumer plans and its business plans operate on different terms.

On the business plans, which are ChatGPT Business and Enterprise and Claude Team and Enterprise, the vendor processes the firm's content under its data processing addendum and does not use that content to train its models by default. The terms of the consumer plans are different, and OpenAI states that on personal workspaces with ChatGPT Free, Plus or Pro, "data sharing is enabled by default". Anthropic's consumer terms allow it to use conversations for "training our models, unless you opt out of training through your account settings". Its consumer terms for the UK also contain the clause "Non-commercial use only. You agree not to use our Services for any commercial or business purposes".

The policy should therefore name the workspace the firm has approved and require that work involving personal data or client information is carried out only in that workspace. It should also prohibit the use of personal accounts for that work, including a personal account on a product the firm has approved under a business plan. The terms of each plan are set out in our guides to Claude at work and the ChatGPT plans.

What staff may enter

Article 5(1)(c) requires personal data to be "adequate, relevant and limited to what is necessary" for the purpose. In a policy, this becomes an instruction to remove what the task does not need before entering anything into the tool, in particular names, contact details and account or reference numbers.

The policy should then list the kinds of information that may not be entered without specific approval. These usually include special category data such as health information, information about criminal convictions or offences, information about children, passwords and other credentials, and material that is legally privileged or confidential to a third party. Which of these the firm permits, and for which tasks, should follow from its own assessment of the processing, so that the decision is not left to each member of staff.

Settings and feedback buttons

The rules on feedback differ by product and by plan. For its commercial products, Anthropic states that conversations reported through the thumbs up or thumbs down button may be used to train its models, and that a conversation submitted as feedback is stored "for up to 5 years". The owner of a Team or Enterprise workspace can disable feedback with the Rate chats setting. OpenAI states that on its consumer plans it may use the whole conversation associated with a thumbs up or thumbs down "even if you have opted out of model training". For ChatGPT Business, Enterprise and Edu, OpenAI states that response feedback "does not permit training on the associated conversation", and workspace administrators can enable or disable it.

Because the position depends on the product and the plan, the simplest rule for staff is not to use the feedback buttons on any conversation that contains personal data or client information. It should also reserve changes to workspace settings, and the connection of the tool to other systems such as shared drives or email, to a named person.

Free download

Get the staff instruction on AI tools

A one-page instruction you can adapt and issue to staff, covering approved accounts, what may be entered, feedback buttons, checking output and reporting mistakes. Each rule shows its legal or contractual reference.

For partners, directors and compliance leads who need to give staff written rules on ChatGPT, Claude or Copilot.

  • ·Eight rules on one page, with a signature line for each member of staff
  • ·The legal or contractual reference for each rule
  • ·A second page of notes on adapting it to your firm and keeping the record

Your email is used to deliver the PDF and (if you opt in) the newsletter. No spam. Privacy policy.

Checking output

These tools can produce text that is inaccurate or that presents invented material as fact. Where the firm processes personal data in an output, Article 5(1)(d) requires those data to be accurate and, where necessary, kept up to date. Every reasonable step must be taken to erase or rectify without delay personal data that are inaccurate, having regard to the purposes for which they are processed. The policy should state that the person who uses an output is responsible for it, and that outputs are checked before they are relied on or sent to anyone outside the firm.

Decisions about individuals need a separate rule. Since 5 February 2026, Articles 22A to 22D of the UK GDPR have governed significant decisions taken solely by automated processing, and whether a person was meaningfully involved in the decision is part of that test. The policy should state that AI output is not used to make or support a decision about an individual, such as a recruitment or disciplinary decision, unless that use has been separately assessed and approved. Our guide to automated decisions under Articles 22A to 22D explains the requirements.

Retention and deletion

Article 5(1)(e) limits the keeping of identifiable personal data to the period necessary for the purposes of the processing. The policy should explain when staff must delete conversations and when a conversation must be preserved, for example for an incident investigation or because the firm's retention rules require it. Those instructions should reflect the retention settings on the firm's workspace and the vendor's deletion terms.

Reporting mistakes

Under Article 33, a firm acting as controller must notify the regulator of a personal data breach without undue delay and, where feasible, within 72 hours of becoming aware of it, unless the breach is unlikely to result in a risk to the rights and freedoms of individuals. A firm acting as processor must notify the controller without undue delay after becoming aware of a breach. In either role, the firm can assess an incident only once it has been told about it.

The policy should therefore require staff to report at once if they have entered information the policy does not allow, used a personal account for work, or noticed personal data in an output where it should not appear. It should name the person to report to, and it should make clear that the member of staff reports the incident and the firm decides whether it amounts to a breach.

Monitoring staff use

A firm may wish to review how its staff use the tools, for example through the records available to the administrator of its workspace. That review is itself processing of staff personal data, so the firm should tell staff in the policy that it may take place and describe the processing in its staff privacy notice. The wider rules are covered in our guide to monitoring staff with AI.

Acknowledgement, ownership and review

A record of who received the policy, and when, strengthens the firm's evidence that it communicated its instructions. Each member of staff should acknowledge the policy, new joiners should do so before they are given access to the workspace, and the firm should keep the record. The policy should also name its owner, carry a version number and a date, and be reviewed when the firm adds a tool, changes plan, or when a vendor changes its terms.

How the policy relates to the firm's other documents

The policy records the firm's rules for staff. It does not establish that the underlying processing is lawful, which depends on the firm's lawful basis, its contract with the vendor, its transfer position and, where one is required, its data protection impact assessment. Where an assessment is required, the controls it identifies are the ones the policy should record, and our guide explains when a DPIA is required for an AI system.

Our DPIA tool screens whether a data protection impact assessment is required and helps you draft one in a conversation. It is free to complete, and downloading the finished DPIA as a Word document costs £39.

For firms that use these tools on client files, the related guide on using Claude or ChatGPT on client files covers confidentiality, lawful basis and the full set of documents. A one-page staff instruction covering the points above is available to download from this page.

Checked on 16 November 2026 against the UK GDPR as published on legislation.gov.uk (Articles 5, 24, 29, 32, 33 and 22A to 22C), Anthropic's Consumer Terms of Service for the UK (effective 8 October 2025), Commercial Terms of Service and Privacy Center, and OpenAI's Data Processing Addendum, data controls FAQ and help article on how data is used to improve model performance. This guide is general information about UK data protection law, not legal advice.

Free download

Get the staff instruction on AI tools

A one-page instruction you can adapt and issue to staff, covering approved accounts, what may be entered, feedback buttons, checking output and reporting mistakes. Each rule shows its legal or contractual reference.

For partners, directors and compliance leads who need to give staff written rules on ChatGPT, Claude or Copilot.

  • ·Eight rules on one page, with a signature line for each member of staff
  • ·The legal or contractual reference for each rule
  • ·A second page of notes on adapting it to your firm and keeping the record

Your email is used to deliver the PDF and (if you opt in) the newsletter. No spam. Privacy policy.

Frequently Asked Questions

Is a staff AI use policy a legal requirement in the UK?

The UK GDPR does not expressly require a document called a staff AI use policy. Article 32(4) requires controllers and processors to take steps to ensure that people acting under their authority process personal data only on the controller's instructions, unless domestic law requires otherwise. Article 24(2) requires a controller to implement appropriate data protection policies where these are proportionate to its processing. Written instructions on AI tools, with a record that staff have acknowledged them, help a firm to put these measures in place and to evidence them.

Can staff use a personal ChatGPT or Claude account for work?

A personal account should not be approved for work involving personal data or client information. The consumer plans do not give the firm a processing agreement, and conversations on them may be used to train the vendor's models. The training settings do not remove every exception, because feedback can still permit training and Anthropic also uses content flagged for safety review. Anthropic's consumer terms for the UK prohibit commercial or business use.

Do staff need to sign the policy?

The UK GDPR does not require a signature. A signature or a recorded acknowledgement allows the firm to show that each person received the instruction, and Article 24(1) requires a controller to be able to demonstrate that its processing complies with the Regulation. A policy carries more weight as evidence when the firm can show who received it and when.

Does a staff AI use policy replace a DPIA?

No. The policy gives staff their instructions, while a data protection impact assessment examines the processing itself and identifies the controls that the policy then records. Where the firm's use of AI requires an assessment, the assessment comes first and the policy follows from it.

Start with a £500 scoping review

If you need GDPR documentation, AI Act work, or a compliant AI build, the first step is a written scoping review. You get a written report you can act on.

staff AI use policyChatGPT policy for employeesClaude at work policyAI acceptable use policy UKgovern ChatGPT and Claude usageUK GDPR Article 32