# Security and compliance review of an AI training tool

Review the knowledge workflow before granting access. An AI training tool security review should establish which sources the tool reads, what it can do, who approves its output and how access is revoked before protected data is shared. This page is for IT and compliance reviewers scoping a pilot: Elvina reads only the sources you approve and takes no actions in your systems on its own.

## Trace one workflow end to end

Identify the source, the acting identity, the permitted operation, the output and the person who approves release. Separate reading knowledge from writing to a business system. A successful connection test does not prove that permissions are minimal or that the intended action is safe. Begin with the least access necessary for the accepted scenario.

## Inspect the same boundary in Web and Chat

A restricted source must remain restricted when a user asks a question conversationally. Test an unauthorized request, access removal, stale content and a failed connector. The interface should explain a blocked action without exposing confidential details. Do not assume that an assistant’s ability to describe an operation means it is allowed to perform it.

## Agree on the data lifecycle

Document what enters the workflow, where derived material appears, who may review it and what happens when the source is withdrawn. Define retention, deletion, exports and consent requirements for the pilot. Record which controls are implemented, configured, contractually agreed or still unverified. Uploaded materials are used to build your training. Model-training terms depend on your plan and applicable agreement. See the ContentBuilder Privacy Policy and your workspace agreement for data use, service providers and available opt-out choices. Data is encrypted in transit. SSO (SAML), SCIM provisioning, granular roles and an audit log are on the Enterprise plan; hosting, residency, certification and service levels require evidence rather than marketing inference.

## Use an evidence pack that supports a decision

The acceptance record should name the tested version, scenario, role, source boundary and observed result, with sensitive data removed. Analysts need metric definitions and missing-data handling. Legal and compliance reviewers need the actual terms and intended use. A clean technical test is useful, but it does not independently approve the deployment for your organization.

## An acceptance check to run first

Use synthetic accounts with different source permissions. Verify the permitted read and the refusal of an unauthorized question in both interfaces. Remove access, withdraw a source and repeat the relevant checks. Save only sanitized evidence. The reviewer should be able to identify what was tested without receiving credentials or confidential source content.

## Questions

**Q:** Is every listed connector and enterprise control live?

Not all of them. Google Drive and Notion connect to Company Knowledge directly, SSO (SAML), SCIM provisioning, granular roles and an audit log are on the Enterprise plan. Dropbox and SharePoint are planned, not available today. Confirm the exact action, identity, environment and contractual requirement for your pilot.

**Q:** Does a successful demo satisfy our security review?

No. A demo proves only what was demonstrated. Review configuration, data handling, refusal behavior, incident ownership and any required contractual evidence independently.

No launch date is committed for planned features.