Before connecting an AI tool to company email, cloud storage, or business files, ask specific questions about data access and handling. A statement such as “we don’t train on your data” does not explain which files the tool can read, what interaction logs it keeps, or how to remove the connection later.
Use this checklist before approving an integration. Ask the provider for clear answers in writing, then check that the permissions requested by the app match what you were told.
Start with the data the AI can access
First, ask the provider to describe exactly what the integration can reach. Avoid settling for broad terms like “your workspace” or “connected account.” Find out whether it can access email, attachments, calendars, shared drives, chat, or other connected services.
- Which mailboxes, folders, drives, or other data can the AI access?
- Can we limit access to selected users, folders, or files?
- Does it see only new information, or can it search older content too?
- Can it access shared or externally shared files?
Then review the app’s consent or permissions screen. Pay attention to whether permissions are read-only or allow the tool to edit, send, or delete information. Also ask whether access is granted on behalf of a signed-in user or through app-only permissions that can work without a user present. These permission types can allow different kinds of access; Microsoft’s overview of Microsoft Graph permissions explains the distinction in its platform.
If the provider cannot explain each requested permission in plain language, pause before granting access. A tool may request more access than your intended use requires, so ask whether any permissions can be removed or narrowed.
Clarify how prompts, files, and outputs are used
Ask whether the provider uses any business information to train or improve its models. Be precise about what “data” means in the answer. Does the promise cover the email or file content retrieved by the tool, the prompts your staff enter, and the AI-generated responses?
- Will you use our emails, files, prompts, or responses to train or improve any models?
- Does that answer change by product, account type, or settings?
- Do you retain prompts, retrieved content, or outputs in logs?
- Can your employees or subcontractors review any of this information, and under what circumstances?
A “not used for training” statement does not automatically answer questions about logging, human review, or retention. Ask the provider to explain each separately and point to the terms that apply to the specific tool and integration you plan to use.
Policies can differ across products, so check the documentation for the exact service you intend to use. For example, Google’s information about how Gemini in Google Workspace protects data describes its own services. Treat product-specific documentation as a starting point, not as an answer for every AI provider.
Find out where data goes and how long it stays
When an AI service processes business information, another provider or service may be involved. Ask who receives or processes your data, including AI model providers and subprocessors. Find out what each party does with it and which terms apply.
- Where is our data stored and processed?
- Which subprocessors or model providers receive or process it?
- How long are prompts, files, outputs, and logs retained?
- Can we request deletion, and what does that request cover?
- What happens to copies in backups or other retained records?
Ask for a useful distinction between the original information in your company systems and any copies or records the AI service may create. Also ask how the provider will confirm that a deletion request has been completed. A general privacy policy may not spell out what happens to information accessed through a particular integration.
Check who can review data and how issues are handled
Ask whether provider staff or subcontractors can review prompts, files, or email content. If review is possible, ask what triggers it, who is allowed to do it, and whether access is limited and recorded. A security or encryption statement is useful context, but it does not explain these operational details.
Also ask how the provider handles a suspected exposure involving your information. Who will contact your business, what information will they share, and where can you find the relevant process in the provider’s terms? Make sure your internal contact knows who should receive and assess a notice.
For a broader look at setting expectations for employee AI use, see Zevonix’s guide to a small business AI use policy. If your concerns extend to protecting sensitive business information beyond AI tools, its article on securing sensitive data offers related guidance.
Review the connection before and after a pilot
Do not begin with unrestricted access if a smaller test can answer the business question. A limited pilot lets your team check whether the tool works as expected while keeping the initial scope narrower.
- Choose a limited account or folder. Use information appropriate for testing, and avoid exposing sensitive material unnecessarily.
- Review existing sharing permissions. The AI tool may be able to retrieve information that a connected user or account can already access, including shared content. Check what is available before connecting.
- Inspect the consent screen. Compare each requested permission with the provider’s written explanation. Confirm whether access is read-only, restricted to selected users, or able to operate without a user signed in.
- Test the intended workflow. Check what information the AI returns and whether its access stays within the boundaries you set.
- Review and revoke when finished. Confirm how to inspect permissions, remove the connection, and request deletion of stored data and interaction history.
Write down who owns the connection and how your team will review it. If the provider cannot explain how to revoke access or what deletion covers, do not assume that disconnecting the app removes every copy or record.
Frequently asked questions
Is “we don’t train on your data” enough?
No. It answers only one part of the data-use question. Ask separately about logs, retention, human review, subprocessors, and deletion.
Should an AI integration have read-only access?
Use the narrowest permissions that support the task. If the tool only needs to find or summarize information, ask why it needs permission to edit, send, or delete data.
What should we do when we stop using the AI tool?
Revoke the app’s access, review any remaining permissions, and ask the provider how to request and confirm deletion of copied data, prompts, outputs, and interaction history.
Can we test an AI tool without connecting the whole company account?
Ask whether you can begin with a limited user or folder. Check the tool’s permission scope and existing sharing access before starting the pilot.
Make the decision with clear answers
Before connecting AI to company email or files, understand its access, data use, retention, human review, and offboarding process. Start with a limited pilot, and expand access only when the permissions and provider answers fit your needs. For help considering AI access as part of your broader IT planning and security, contact Zevonix at (386) 260-0777.