Using an AI tool does not make sensitive information safe by default. This practical AI data privacy checklist helps teams and creators classify data, remove or replace identifying details, assess an LLM provider, and respond sensibly if confidential information is disclosed. Use it before uploading customer, employee, client, source-code, or business data, then revisit it whenever a workflow or vendor changes.
Overview
The safest starting point is to assume that information pasted into an AI system may be stored, logged, reviewed, transferred, or processed according to the provider’s configuration and contract. The exact behaviour depends on the product, account type, settings, integrations, and applicable agreements. Never rely on a general statement such as “the tool is secure” when deciding whether a particular dataset can be used.
First classify the material. A simple classification scheme can make decisions consistent across a team:
| Classification | Examples | Default AI handling |
|---|---|---|
| Public | Published articles, public documentation, approved marketing copy | May be suitable for an approved tool, subject to normal quality and copyright checks. |
| Internal | Unpublished plans, internal procedures, non-public analysis | Use only in an approved workspace with suitable access and retention controls. |
| Confidential | Contracts, customer records, employee information, financial forecasts, private source code | Do not upload until the use case, provider, contract, controls, and lawful basis have been reviewed. |
| Restricted | Passwords, private keys, payment details, health information, identity documents, security incident data | Keep out of general-purpose AI tools. Use a specifically approved, controlled process if processing is genuinely necessary. |
Classification is a decision aid, not a substitute for your organisation’s information-security policy. If you are unsure, pause the upload and ask the relevant privacy, security, or legal owner.
For UK organisations, an AI GDPR compliance checklist should normally include purpose, data minimisation, transparency, security, retention, individual rights, processor arrangements, and international-transfer considerations. The appropriate obligations depend on the use case and organisation. Treat this article as an operational checklist, not legal advice, and consult your data-protection lead or qualified adviser where the risk is material.
Checklist by scenario
Before using customer or prospect data
- Write down the precise purpose: for example, categorising support themes or drafting a response.
- Remove names, email addresses, phone numbers, postal addresses, account numbers, order references, and free-text details that could identify a person.
- Check for indirect identifiers, such as a rare job title, exact date, small location, unique complaint, or combination of attributes.
- Use representative synthetic examples or aggregated data for prompt development and demonstrations.
- Confirm that the approved tool and workspace match the sensitivity of the task.
- Define who can access prompts, uploaded files, generated outputs, logs, and exported results.
- Require a human review before an output is sent to a customer or used to make a consequential decision.
Before using employee or applicant information
- Do not paste performance notes, disciplinary records, health information, payroll details, or application documents into an unapproved tool.
- Separate the task from the person’s identity. Ask the model to analyse a redacted policy scenario rather than an identifiable case whenever possible.
- Check whether the proposed use is necessary and proportionate for the stated purpose.
- Record the workflow owner, approved users, retention period, and deletion process.
- Prevent the model from making an unreviewed hiring, promotion, disciplinary, or employment decision.
Before using business-confidential material
- Remove trade secrets, unreleased product details, customer lists, negotiated terms, credentials, security architecture, and private financial information.
- For code assistance, replace secrets with placeholders and scan files for tokens, keys, connection strings, and environment variables.
- Limit the context to the minimum files and passages needed for the task.
- Check whether plugins, browser tools, retrieval systems, or connected storage can send the material to additional services.
- Keep an audit record of the approved use case without copying the sensitive content into the record.
Before building an LLM application
- Map every data flow: collection, preprocessing, model request, provider logs, application logs, retrieval index, output storage, and deletion.
- Redact or pseudonymise before data reaches the model rather than relying only on a prompt instruction such as “do not reveal personal information.”
- Use access controls, encryption, environment-specific credentials, and separate development data from production data.
- Test whether the application can reveal source documents through direct questions, prompt injection, or excessive output.
- Define retention and deletion behaviour for prompts, completions, embeddings, caches, backups, and monitoring systems.
- Monitor quality and unexpected data exposure. The LLM observability runbook can help structure checks for errors, drift, and operational signals.
Redaction and pseudonymisation workflow
Use this repeatable workflow when a useful task appears to require real data:
- Define the task. State what the model must produce and what information it does not need.
- Extract only the necessary fields. A support-summary task may need issue text and product category, but not a customer name or full order history.
- Detect sensitive content. Search structured fields and free text for names, contact details, identifiers, credentials, financial data, health information, confidential business terms, and embedded documents.
- Redact or generalise. Replace content with labels such as
[CUSTOMER_NAME], “a UK customer,” or an approximate time period. Redaction removes detail; generalisation reduces precision. - Pseudonymise consistently where needed. Replace a person with a stable token such as
CUSTOMER_014if the model needs to track references within one task. Store the mapping separately, protect it, and do not send it to the model. - Inspect the result manually. Automated detection can miss unusual names, contextual clues, images, screenshots, and combinations of facts that identify someone.
- Limit output and retention. Ask for the smallest useful response, remove sensitive material from the output, and delete temporary files according to the approved process.
A prompt can support this workflow but cannot replace it. For example: Analyse the redacted case below. Do not infer or reconstruct identity. Return three issue categories and evidence using only the supplied text. This is a useful instruction, but the input should already have been minimised.
For prompt-based research and reporting, use a structured process that keeps source checks separate from unnecessary personal information. The prompt template library provides a useful companion for repeatable research tasks.
What to double-check before approving a tool
Record answers rather than relying on a sales page or a colleague’s assumption. The following questions form a practical vendor privacy comparison:
- What data is collected in prompts, uploads, outputs, telemetry, and support requests?
- Are customer inputs used to train, improve, evaluate, or otherwise develop the provider’s models? Is the setting controllable and documented for the specific product?
- How long are inputs, outputs, backups, and logs retained, and can an administrator delete them?
- Where is data processed and stored? Are regional processing options available for your requirements?
- Which subprocessors or connected services may receive the data?
- What encryption, identity, role-based access, audit logging, tenant separation, and administrator controls are available?
- Can the provider support relevant contractual, data-protection, deletion, access, and incident-notification requirements?
- What happens when the account is closed, a user leaves, or the provider changes its terms?
- Can the application prevent prompts and outputs from appearing in developer logs, analytics tools, error reports, or browser history?
Save the provider’s answers, configuration screenshots, approval date, owner, and review date. Recheck them when the product, plan, region, contract, or workflow changes. A broader AI tool evaluation framework can help balance privacy with accuracy, cost, and workflow fit.
Common mistakes
- “It is only a draft.” Drafts can still contain personal data, trade secrets, or regulated information. Classify the input, not just the intended output.
- Using real records for prompt testing. Build tests from synthetic, anonymised, or carefully minimised examples whenever possible.
- Leaving metadata behind. File names, document properties, tracked changes, screenshots, hidden spreadsheet tabs, and URLs can reveal more than visible text.
- Confusing pseudonymisation with anonymisation. A dataset remains potentially identifying if someone can reconnect tokens to people. Protect the mapping and apply the correct risk assessment.
- Trusting a system prompt to enforce privacy. Instructions can be ignored, bypassed, or exposed. Apply technical access controls and minimise data before inference.
- Forgetting connected tools. Retrieval plugins, automation platforms, ticketing systems, and monitoring services can create additional data flows. Review the security of the whole workflow.
- Ignoring outputs. A generated summary may reproduce sensitive details or combine harmless facts into a revealing conclusion. Review, filter, and restrict distribution.
- Failing to prepare for mistakes. Teams need a clear route for reporting, containing, assessing, and documenting accidental disclosure.
If an accidental disclosure occurs, stop further processing, preserve the relevant facts without unnecessarily duplicating the sensitive content, revoke exposed credentials where applicable, restrict access, notify the internal privacy or security owner, and follow the organisation’s incident process. A useful initial report records who discovered it, what type of data was involved, which system received it, approximate time, affected users or records, actions already taken, and outstanding questions. Do not decide alone whether notification is required; escalate for a documented assessment.
When to revisit this checklist
Make this review part of normal AI workflow management rather than a one-off approval. Revisit it before a seasonal planning cycle, when a team launches a new campaign or product, when a new dataset is introduced, and whenever an AI provider, model, integration, account plan, contract, region, or retention setting changes.
Also review after an incident, a security assessment, a material change in UK data-protection guidance, or a change in the purpose of processing. Schedule a quarterly check if the workflow handles ongoing customer, employee, or confidential business data; higher-risk processes may need a more frequent review set by your privacy and security owners.
Use this short sign-off sequence:
- Name the workflow and its owner.
- Classify the input and output data.
- Confirm minimisation, redaction, and pseudonymisation steps.
- Record the approved provider, settings, region, retention, access, and connected services.
- Test for leakage, inappropriate access, and unsafe outputs.
- Document human review and incident escalation.
- Set the next review date and a trigger for earlier review.
Once these checks are complete, teams can use AI more deliberately: share only what the task requires, keep control of the surrounding data flow, and treat privacy settings as configuration that must be verified rather than assumed.