LLM Data Privacy Checklist: A Practical Workflow for Prompts, Logs, and Third-Party APIs
Before sending business information to an LLM, use this repeatable privacy workflow to identify personal data, minimise what leaves your systems, assess the provider, control logs and access, and document decisions for future reviews.
Overview
LLM applications can process much more than the text a user deliberately enters. Prompts may contain names, customer messages, support tickets, source code, contracts, internal policies, identifiers, or confidential business plans. The same information can also appear in application logs, traces, analytics platforms, evaluation datasets, backups, and support tickets.
A useful privacy assessment therefore follows the complete data flow rather than focusing only on the model. Map what is collected, where it is transformed, which systems receive it, who can access it, how long it is retained, and how it can be deleted or corrected. This is a practical risk-control process, not a substitute for advice from your data protection or legal team.
For each proposed use case, record the purpose, data categories, users, provider, processing location, retention settings, access controls, and review owner. Treat the record as a living document: an API change, new logging destination, model switch, or product feature can change the privacy position without changing the user interface.
What to track
1. Data entering the workflow
Start with a prompt and input inventory. List every field sent to the LLM, including system instructions, retrieved documents, conversation history, tool results, file uploads, and error messages. Classify each field as public, internal, confidential, personal, sensitive, or unknown. Do not assume that text is harmless because it is unstructured; a paragraph can contain several identifiers or sensitive details.
Ask whether the use case genuinely needs the data. Replace full names with stable pseudonyms where possible, remove direct identifiers, shorten long documents to the relevant passages, and exclude unrelated conversation history. Data minimisation should be designed into the application rather than left to users to remember while writing prompts.
2. Redaction and transformation controls
Document the controls applied before transmission. Examples include pattern-based removal of email addresses, phone numbers, account numbers, access tokens, and private keys; entity detection for names and addresses; field allowlists; truncation; and pseudonymisation. Record known limitations: redaction can miss unusual formats, inferential identifiers, images, tables, or information spread across several fields.
Test the controls with realistic but synthetic examples. Confirm that redaction does not create a new problem, such as removing the context needed for a safe answer or leaving a secret in an exception message. Keep production secrets and credentials out of prompts altogether.
3. Provider terms and processing settings
Complete an AI vendor privacy assessment before production use. Obtain current answers to the following questions from provider documentation or your commercial contact:
| Question | Why it matters | Evidence to retain |
|---|---|---|
| Are API inputs and outputs used to train or improve models? | Determines whether business data may be used beyond the requested service. | Current API terms, product setting, or written provider answer. |
| How long are prompts, outputs, abuse records, and support data retained? | Retention may differ by data type and product route. | Retention configuration and relevant documentation. |
| Where is data processed and stored? | Regional requirements and transfer assessments may depend on the processing path. | Region selection, subprocessors, and transfer information. |
| Who are the subprocessors and how are changes communicated? | Other organisations may handle data or infrastructure. | Subprocessor list and change-notification process. |
| Can data be deleted, corrected, exported, or isolated by tenant? | Supports operational requests and separation between customers. | API procedure, service commitment, and test result. |
| What security controls and incident-notification processes apply? | Helps align the service with your risk and response requirements. | Security documentation and internal approval record. |
Do not rely on a general consumer product statement when assessing an API, hosted enterprise service, or embedded third-party feature. The product route, account configuration, and contract can affect the applicable controls. Recheck the provider’s documentation whenever these details change.
4. Logs, access, and secondary copies
Review application logs as carefully as prompts. Search for request bodies, full model responses, headers, user identifiers, uploaded filenames, tool arguments, and exception traces. Set an explicit policy for what may be logged. Prefer metadata such as latency, model version, token count, request status, and a non-reversible request identifier over raw content.
Where content logs are necessary for debugging or evaluation, restrict access by role, encrypt them appropriately, define a short and justified retention period, and record deletion ownership. Separate production data from test and evaluation datasets. Mask sensitive values in dashboards and prevent them from appearing in alert notifications or third-party observability tools.
5. A compact data-flow worksheet
Complete one row for each stage of the workflow:
- Source: Where does the data originate, and who provides it?
- Purpose: What specific task requires an LLM?
- Fields: Which data elements are collected, generated, retrieved, or inferred?
- Transformation: What is removed, masked, summarised, or pseudonymised?
- Recipient: Which model provider, tool, storage system, or subprocessors receive it?
- Location: Which regions and environments process or store it?
- Retention: How long is it kept at each stage, and who deletes it?
- Access: Which roles can view, export, change, or delete it?
- Control owner: Who reviews the workflow and responds to incidents or requests?
Cadence and checkpoints
Use a simple review rhythm so privacy checks do not depend on memory. A monthly operational check can examine log samples, redaction failures, access activity, deletion jobs, new data fields, and unexpected provider or model routes. A quarterly review can revisit provider terms, regional processing, subprocessors, retention settings, access permissions, and the list of approved use cases.
Run an additional checkpoint before launch and after material changes, including a new provider, model, region, retrieval source, tool integration, user group, logging platform, or prompt that introduces new business data. Include privacy tests in release acceptance criteria: sensitive test values should be removed or blocked, raw prompts should not appear in ordinary logs, and deletion should work across the systems you control.
Keep an evidence folder or register containing the data-flow worksheet, provider review, configuration exports, test results, approvals, incidents, and review dates. Link each production workflow to an owner and a next-review date. This makes an annual data-flow assessment faster and gives reviewers a clear trail of decisions.
How to interpret changes
Not every change has the same significance. A model version change may require output and safety testing, but it also warrants a check that the endpoint, region, retention behaviour, and provider terms remain appropriate. A small interface change can be more important if it adds file uploads or automatically includes conversation history.
Escalate a change when it introduces a new category of personal or confidential data, expands the user population, sends information to a new recipient, changes the purpose, increases retention, weakens access control, or makes deletion difficult. Pause release where you cannot establish what data is sent, who can access it, or how long it remains available.
Track both control indicators and real incidents. Useful indicators include the percentage of requests passing redaction checks, the number of raw-content log events, failed deletion jobs, privileged-access exceptions, unresolved provider questions, and prompts blocked for containing secrets. A rising number can indicate better detection rather than worse privacy, so interpret trends alongside sampling and investigation notes.
When to revisit
Revisit this LLM data privacy checklist at least quarterly for active systems, and sooner when recurring data points change. Review it before onboarding a new AI vendor, approving a new use case, enabling content logging, changing retention settings, switching models, adding retrieval or tool calling, moving regions, or responding to a privacy or security review.
Use the following pre-release checklist:
- Purpose and data categories are documented.
- Unnecessary fields, identifiers, secrets, and unrelated history are removed.
- Redaction and input validation have been tested with representative synthetic data.
- Provider training, retention, location, subprocessors, deletion, and incident terms have been checked using current documentation.
- Raw prompts and outputs are excluded from routine logs or protected by a documented exception.
- Access is role-based, monitored, and reviewed.
- Deletion, correction, export, and incident-response procedures have an owner.
- The data-flow record, approval, test evidence, and next-review date are stored.
Finally, assign the next review before closing the current one. Privacy work is most effective when it is treated as part of prompt engineering, LLM app development, deployment, and evaluation—not as a one-time approval performed after the system is already in production.