Before you connect an AI tool to company email or files, check exactly what it can access, what data leaves your systems, and what happens to that data afterward. The key question is not only whether the tool is useful, but whether its access and handling practices fit your organization’s needs.
Use this checklist to evaluate an integration before enabling it. For sensitive or regulated information, involve the appropriate security, privacy, compliance, or legal staff before sharing data.
Start with the access the tool is requesting
Ask the vendor to show the permissions the integration requires and explain why each one is needed. Look for access to email, attachments, calendars, shared drives, chat, or other connected services. Check whether the tool can only read information or can also send messages, edit files, change settings, or take other actions.
Pay particular attention to whether access is granted on behalf of an individual user or to an application. With delegated access, an app acts on behalf of a signed-in user, within the permissions available to that user. Application permissions allow an app to act without a signed-in user and may reach organization-wide resources, depending on the permissions granted. Microsoft’s overview of Microsoft Graph permissions explains this distinction for Microsoft Graph integrations.
Apply least privilege: grant only the access needed for the intended task. If the tool is meant to summarize messages in a specific mailbox or search documents in a particular workspace, ask whether you can restrict it to those locations rather than granting broad access. Confirm who can approve permissions and whether an administrator can review or revoke them.
Trace where company data goes
Map the data path from the moment a user invokes the tool. Find out which messages, files, attachments, prompts, and generated outputs are sent to the provider. Ask whether the information is processed by the vendor or by additional service providers, and whether copies appear in logs, backups, or support systems.
Get a clear answer on model training. Does the provider use company data, prompts, or outputs to train or improve models? Does that depend on the product plan, an account setting, or a separate agreement? Ask for the answer to be documented for the specific service your organization will use, rather than relying on a general statement about the vendor’s products.
Clarify retention and deletion as well. Ask how long the provider keeps prompts, email, files, outputs, and logs; what happens to backups; and what deletion means after an account is disconnected or closed. Find out whether the organization can request deletion and whether the vendor explains any exceptions.
Review vendor access and safeguards
Ask who at the provider and its subcontractors can access company data, for what reasons, and under what controls. Understand how access is authorized and recorded, and whether support personnel may see customer content while troubleshooting. Check how the provider handles security incidents and communicates relevant information to customers.
Review the vendor’s security documentation and contract together. Look for specific commitments about data use, retention, deletion, subprocessors, security controls, incident handling, and your organization’s ability to end access. Confirm that the documents apply to the exact AI feature and service being connected. If terms are unclear or conflict with internal requirements, pause and ask the appropriate reviewer to resolve them.
For organizations handling sensitive personal information, consider whether state requirements apply to the data and workflow. Florida’s statute on security of confidential personal information is available from the Florida Legislature. Have your privacy or legal team assess how it applies to your situation instead of treating a vendor’s general security statement as a compliance determination.
Healthcare organizations should also ask their compliance or legal team whether the AI service’s role and the information involved call for a business associate agreement (BAA). Do not assume a standard service agreement or privacy statement answers that question.
Test with limited access before expanding
A staged review helps uncover problems without immediately exposing every mailbox or file repository:
- Identify sensitive locations. Note which mailboxes, shared folders, and records contain confidential, personal, or regulated information. Decide which sources the tool should not access.
- Start with a narrow scope. Use a limited group of users and the smallest practical set of mailboxes or folders. Avoid broad organization-wide permissions until you have a reason to expand.
- Use suitable test material. Where possible, test with non-sensitive or approved sample content. Confirm that the integration behaves as expected before connecting it to live business records.
- Inspect activity. Review the permissions granted, the content accessed, the outputs produced, and any actions taken. Check whether logs provide enough detail for your team to understand what happened.
- Revoke access when needed. Confirm how to disable the integration, remove permissions, and request deletion of retained data. Include those steps in your internal process.
Expand access only when the test supports it and the vendor’s handling practices meet your organization’s requirements. Keep a record of who approved the connection, what scope was approved, and when the decision should be reviewed again.
Decide where human approval is required
Set clear limits on what the AI may do on its own. A summary or draft may still contain errors, omit context, or disclose information to the wrong audience. Decide which outputs require a person to verify them before use, particularly when they affect customer communications, employee records, business decisions, or regulated information.
Require review before the tool sends messages, shares files, changes records, or takes other actions that are difficult to undo. If the integration only generates suggestions, make sure employees know they remain responsible for checking the result before relying on it. The NIST Generative AI Profile can serve as a reference for organizing risk review and oversight; it does not replace your organization’s own requirements.
Frequently asked questions
How can we limit an AI integration to specific mailboxes or folders?
Ask the vendor and your administrator which permission scopes and configuration options are available. Test the narrowest access that supports the task, then verify through administrative settings and activity records that the integration cannot reach unrelated locations.
Does the AI vendor use company data to train its models?
Do not assume the answer. Check the terms and settings for the specific product, and request written confirmation about whether prompts, files, and outputs are used for training or improvement.
What should we do before connecting sensitive personal information?
Identify the data involved, review the tool’s permissions and handling practices, and involve your privacy, security, compliance, or legal staff as appropriate. They can assess applicable requirements and whether the proposed use is acceptable.
When should a healthcare organization consider a BAA?
Ask your compliance or legal team to assess the service, the vendor’s role, and the information the AI tool will process. Confirm whether a BAA is needed and have the relevant terms in place before sharing the information.
Make the connection a reviewed business decision
Before enabling an AI integration, document its access, data handling, retention, vendor access, test results, and human approval rules. Revisit that review if the tool’s permissions or use change. For help assessing AI integration security, data access, or compliance needs, Florida businesses can contact Zevonix at (904) 658-0777.