1. Which parts actually run locally?

Ask whether the model, document search, memory, workflow engine, and activity log run on your hardware. A product may be called local because it installs an app while still sending prompts and files to an outside service. Request a plain-language data-flow diagram, not a one-word answer.

2. What information leaves the computer?

List each connected provider and the exact data it receives. Email services, calendars, online storage, analytics, AI model APIs, and remote support tools can all create external data flows. Useful answers distinguish required information from optional telemetry and explain retention settings.

3. Can access be limited by task?

An assistant that only prepares summaries should not need permission to send email or delete files. Look for least-privilege access, separate accounts where practical, read-only modes, narrow folder scopes, and explicit approval before consequential actions. Broad administrator access should have a specific reason and additional controls.

4. How are credentials handled?

Passwords and API keys should not be pasted into prompts, ordinary documents, scripts, or chat transcripts. Ask whether credentials are held by a protected system store, how they are rotated, who can retrieve them, and what happens when an employee or service provider no longer needs access.

5. What is recorded, and who reviews it?

A useful activity record should show what the assistant attempted, which tool it used, whether a person approved the action, and whether the result succeeded. Logs can also contain sensitive information, so ask where they are stored, how long they remain, and who may read them. Good logging supports accountability without collecting unnecessary content.

6. Are backups protected and tested?

Local data is not safe merely because it sits in the office. Hardware can fail or be stolen. Confirm that important files and configuration are encrypted, backed up to a separate location, and periodically restored in a test. The recovery instructions should identify the owner and the expected recovery time.

7. Can data be exported and deleted?

The business should be able to obtain its documents, workflow rules, memory, logs, and configuration in useful formats. Ask how a customer record is removed from indexes and backups, how cloud copies are handled, and what the provider retains after the relationship ends. Ownership should be written into the agreement.

8. What happens during remote support?

Clarify whether a developer can connect without an invitation, what the owner sees during access, and whether sessions are logged. Time-limited, revocable support access is safer than a permanent hidden route. There should be an emergency method to disable remote access without breaking the business's files.

9. Which decisions always remain human?

Privacy controls and action controls belong together. A system should identify messages, publications, purchases, account changes, and destructive steps that require approval. For regulated or high-impact work, the business may also need legal, compliance, or security advice beyond the AI implementation.

Turn the answers into acceptance tests

Before final payment or handoff, verify the promises. Disconnect the internet and observe which features remain. Revoke a test account. Restore a backup. Request an export. Trigger an approval gate and confirm the action cannot proceed alone. These checks reveal more than a privacy label.

For the underlying architecture choice, see what should stay local or use the cloud. The developer hiring checklist covers ownership, support, and handoff in more detail. If you want these controls included in a scoped build, describe your privacy requirements.