AI Privacy Checklist: 15 Questions Before Uploading Client Data

Tool Tests

Uploading client information to an AI tool is not one decision. It is a chain of decisions about what enters the system, who can access it, how long it remains, what other services receive it, and what happens when something goes wrong. A useful privacy review turns those questions into a documented operating boundary. Start with the broader AI tool evaluation process if the workflow itself has not yet been defined.

Start with the proposed task

Write one sentence describing the task, not the product: “Turn sanitized meeting notes into a draft action list” is testable; “use AI in the business” is not. List the input, expected output, person reviewing it, and system receiving the result.

Then classify the information. Public material and invented examples are suitable for early testing. Internal process notes deserve more care. Client records, personal information, contracts, credentials, health information, or unpublished financial details require explicit authority and an approved environment.

Questions about collection and purpose

1. What exact information will enter the tool? Include prompts, files, metadata, connected accounts, and information added automatically by integrations.
2. Can the same result be achieved with less data? Replace names, identifiers, exact figures, and confidential context with neutral labels whenever they are unnecessary.
3. Is the client expecting this use? Check contracts, confidentiality commitments, professional rules, and the client’s own policies before treating consent as implied.
4. Is the information used to improve models? Check the terms for the specific free, individual, team, or enterprise account—not a different plan.

Questions about storage and access

5. How long are prompts, files, outputs, logs, and backups retained? “Deleted” may describe the user interface rather than every stored copy.
6. Where is the information processed? Location can matter for contracts, regulation, and client expectations.
7. Who inside the provider can access it? Look for support, abuse review, security, and subcontractor access—not only other customers.
8. Who inside your business can access the account? Require individual accounts, appropriate roles, and multifactor authentication. Shared credentials destroy accountability.

Questions about connected services

9. Which integrations can read or write data? A calendar, drive, email, browser, or CRM connection can expand the boundary far beyond one prompt.
10. Does the tool request more permission than the task needs? Prefer read-only and narrowly scoped access. Remove unused integrations.
11. Can external content instruct the system? Documents and websites may contain misleading instructions. Treat their content as untrusted input.

Questions about control and recovery

12. Can you export the useful work? Test an export before relying on the tool.
13. Can you delete prompts, files, history, and the account? Perform a deletion test with synthetic material and record the result.
14. Can you identify and contain a mistake? Keep source information, version history, activity logs, and an owner who can revoke access.
15. What is the fallback? Define how work continues if the service is unavailable, changes terms, raises prices, or no longer meets the privacy boundary.

Run a synthetic-data test

Create an example with the same format, length, edge cases, and ambiguity as the real work but no real client facts. Test normal input, missing information, conflicting instructions, and a request the system should refuse. Record errors and the time required to verify the output.

A successful test does not authorize real data by itself. It shows whether the workflow is useful enough to justify completing the contractual, security, and client-approval checks. The local-versus-cloud privacy comparison explains how the data boundary changes between common setups.

Write the operating rule

Finish the review with a rule another person can follow: approved tasks, prohibited information, required account settings, permitted integrations, human review, deletion schedule, and review date. “Be careful” is not an operating rule.

The strongest privacy control is often a narrower workflow. Give the tool the least information and authority needed to produce a useful draft, keep consequential decisions human, and preserve a practical exit.

Published September 5, 2026. Provider terms and controls change; verify the documentation for your exact account before using real client information.