Can your agency put client data into an AI tool? A UK GDPR checklist

Identify personal data hidden in briefs and attachments

Start with the material, not the prompt. For a proposed brief upload, inspect the main document alongside comments, tracked changes, email signatures and attachments. For a contact list, examine each column. For support transcripts, read the free-text content rather than assuming that removing a customer-name field finishes the job.

Use a cautious screening question: could this material tell the tool something about an identifiable person? Look for names, contact details, account references and descriptions that point to someone. Where classification is uncertain, hold the upload and check with a professional adviser. Separately, screen for client-confidential material and credentials; keep those checks in your agency’s AI data rules too.

• Briefs: inspect stakeholder details, interview extracts, comments and linked files.
• Contact lists: check names, email addresses, phone numbers and notes about individuals.
• Support transcripts: look for identifying details within the conversation, not just in headings.
• Attachments: review each file on its own before including it in an upload.

Check your role, lawful basis and client instructions

Write down the proposed task in plain English: for example, ‘summarise client-supplied support conversations to suggest help-centre topics’. Then ask who decides the purpose and how the information is used. Are you acting on the client’s instructions, deciding the use yourself, or facing a more complicated arrangement? Have an adviser confirm the relevant controller or processor roles and lawful-basis questions rather than choosing a label for convenience.

Read the client agreement and existing instructions before seeking approval for the task. Ask whether they address AI use, external suppliers or restrictions on sharing information. Treat client permission as one check, not a replacement for the wider review. If the instructions or legal position are unclear, pause the upload and ask a professional adviser what documentation, permissions or changes are needed. The supplied ICO questions page flags that its guidance is under review, so check current guidance rather than relying on an old checklist.

Review retention, training use and supplier contract terms

Review the exact service, subscription and account your team will use. Ask the supplier to explain what happens to prompts, uploaded files and outputs, including how long they are kept and whether they are used for model training or other purposes. Request answers in the relevant terms or documentation rather than relying only on a product headline.

Do not end the review when you find a training opt-out. Continue through deletion, support access, other suppliers involved and contractual responsibilities. Ask an adviser to confirm whether the supplier terms and any data-processing agreement are suitable for the proposed arrangement. Keep a dated record of the documents and settings you reviewed; if an important answer is missing, leave that use unapproved.

Check international transfers and access controls

Ask where the supplier stores and processes the information, and from where staff or other suppliers can access it. Include support and administration access in your questions. If the answers involve overseas processing or access, ask a professional adviser whether international-transfer requirements apply and what arrangements are needed. Do not treat a hosting-location answer as the whole assessment.

The ICO says appropriate technical and organisational measures must safeguard people’s rights. For your agency’s implementation, consider work-managed accounts, restricted workspace membership and a clear process for removing access when someone leaves a project. Before enabling a connection to a shared drive or inbox, review exactly which folders or messages you intend the tool to access. Approve the scope deliberately rather than connecting everything for convenience.

Minimise personal data and assess whether a DPIA is needed

Design a version of the task that uses less information. For a brief rewrite, try supplying the audience, tone, deliverables and constraints without stakeholder details. For help-centre planning, consider manually prepared issue summaries instead of full transcripts. For a contact-list task, first test the method with invented sample records. Treat removing names as an editing step, not automatic proof that the remaining material is anonymous.

Before approving the workflow, ask whether a data protection impact assessment, or DPIA, is needed. Record the purpose, information involved, people affected, proposed safeguards and unresolved risks for an adviser to review. The supplied research does not establish the detailed DPIA triggers or transfer rules, so check GOV.UK or a professional adviser for your circumstances. This early assessment follows the ICO’s emphasis on data protection by design and throughout the processing lifecycle.

Create an approved-tool policy and an incident response route

Turn the review into instructions a busy account manager or freelancer can follow. List approved tools and work accounts, permitted tasks, prohibited inputs and the person who approves exceptions. Make approvals specific: ‘draft generic campaign ideas without client records’ is clearer than ‘AI is allowed’. Include a human check before outputs are sent or relied on, because AI can be wrong.

Give accidental uploads a clear reporting route too. Ask staff to stop further sharing and promptly tell the named owner what went in, which tool and account were used, and when it happened. Have that owner coordinate containment and professional advice about any client, regulator or individual notifications. Do not invent your own legal reporting timetable: check current GOV.UK guidance or a professional adviser.

1. 1Record the approved tool, account, task, client restrictions and permitted data.
2. 2Name the person responsible for new-tool requests, exceptions and questions.
3. 3Require a human review before an AI output reaches a client or informs an action.
4. 4Document how staff report an accidental upload and who coordinates the response.
5. 5Revisit approval when the task, supplier terms, settings or connected data changes.