Permission Inheritance in Enterprise AI
A practical test protocol for AI permission inheritance: controlled identities, a per-connector status table, and the evidence needed to assess enterprise AI access control.
An enterprise AI agent may connect to applications using credentials that can access more data than an individual employee is allowed to see. If the agent uses those credentials without enforcing the employee’s own access rights, it can expose confidential information through an otherwise valid query.
AI permission inheritance is the mechanism by which source-level access restrictions are preserved when an AI system retrieves information. It is a core requirement for enterprise AI access control, especially when an agent searches across multiple applications.
Why permission inheritance is the hard problem
Enterprise systems contain data with different access requirements: customer records, internal discussions, financial information, product plans, and employee documents. Those restrictions apply regardless of whether the information is accessed through the original application or an AI interface.
If an employee cannot access a document in the source system, an AI agent should not disclose its contents simply because its integration can read it. The same applies to information extracted from that document and returned in an answer.
This is a security and compliance concern because the system may expose information outside its intended access boundary. Whether a specific incident also violates a regulatory or contractual requirement depends on the data, applicable obligations, and circumstances.
The permission test matrix
The test should use controlled identities and known source permissions. Create User A, who can access a private document, and User B, who cannot. Add a shared document that both users can read. Use synthetic content so unexpected disclosures can be identified without putting real confidential information at risk.
Run the same queries as both users. Record the expected result before testing, rather than deciding after the fact whether an answer looks acceptable.
| Scenario | Setup | Pass criteria | Fail criteria |
|---|---|---|---|
| Private document | Only User A can access the document. | A can retrieve permitted content; B cannot retrieve or receive its restricted contents. | B receives information from the private document. |
| Shared document | Both users can access the document. | Both users can retrieve its permitted contents. | A user is incorrectly denied access or receives unrelated restricted information. |
| Revoked access | Remove User A’s access and repeat the query. | Future requests follow the documented permission-update behavior and its verified timing. | A can still retrieve restricted content beyond the accepted update window. |
| Cross-source query | User B can access a Slack message but not a related Linear issue. | The answer uses only information B is authorized to access. | The answer reveals details from the restricted issue. |
| Permission change | Change a user’s access rights and rerun the same query. | Results reflect the updated permissions within the defined window. | The old access rights continue to govern results beyond that window. |
This is a test protocol. A scenario should be marked as passed only after its result has been observed and recorded.
Record evidence for every result
A pass/fail label without supporting evidence is difficult to reproduce or investigate. Each test record should include:
- Identity: the test user and the permissions assigned to them.
- Source state: the document or record ID, its access settings, and any relevant group membership.
- Request: the exact query, timestamp, and connector used.
- Outcome: the returned answer, source references, and any restricted information disclosed.
- Permission changes: when access changed and when the connected system reflected that change.
- Retrieval evidence: available logs or traces showing which source records were selected and, where observable, included in the model’s context.
The distinction between retrieval and disclosure matters. If a system retrieves an unauthorized document but generates an answer that does not mention it, the visible response may look harmless. That does not prove the retrieval control worked.
Where internal retrieval traces are unavailable, report that limitation. Do not infer that the system passed a test simply because the answer contains no obvious leak.
How each connector handles indexed permissions
Permission inheritance is not a single property that can be assumed across every integration. Source applications represent access differently, and connectors may supply different permission information. The table describes how the connectors work today. It does not report results from running the test scenarios.
| Connector | What to test | Current implementation |
|---|---|---|
| Google Drive | Document-level access and behavior after permissions change or are revoked. | We map Drive file access rules into indexed permissions for retrieval. |
| Slack | User access to indexed messages, membership changes, and cross-source queries. | We associate indexed messages with channel membership. The bot must be invited before it can index a private channel. |
| Linear | Individual access restrictions and behavior after permission changes. | Team filters limit which issues we sync. We fetch projects and documents workspace-wide within the API key’s access. Projects, project updates, and documents without retained team keys are visible to all users in the Omni workspace. |
| HubSpot | Individual HubSpot user permissions carrying over to indexed data. | Individual permission inheritance is not supported; indexed data is visible to all users in the Omni workspace. |
A complete result for any connector should identify the tested configuration, scenarios executed, observed outcomes, and any known limitations. An architecture description or a successful test against another connector is not a substitute for that evidence. See our integrations directory for connector details.
What to ask any enterprise AI vendor about permissions
1. Are permissions enforced before content reaches the model?
Ask how the system excludes records the requesting user cannot access. Request evidence from retrieval tests, not just screenshots showing that restricted documents are absent from the final answer.
2. Whose identity determines access?
Find out whether the system evaluates the employee’s permissions, the integration account’s permissions, or both. Ask how individual access rights and group membership are represented for each connector.
3. What happens after access is revoked?
Ask the vendor to revoke access to a test record and measure how long it remains retrievable. Clarify how the system handles permission-update delays and requests already in progress.
4. Can information leak through queries spanning multiple sources?
Test a question that combines an accessible record with a related restricted record. Ask for evidence that access to one source does not authorize disclosure from another.
5. Can the test results be reproduced?
Request the test identities, source permissions, queries, expected outcomes, actual results, timestamps, and known limitations. If the vendor cannot show how a result was verified, treat the behavior as unverified.
How Omni handles source permissions
We associate permission records with indexed documents and apply an access control filter during search. The filter checks whether a document is public or whether the requesting user is authorized through the listed user or group information. The AI assistant’s search tool uses the same permission filter.
That filter alone doesn’t tell us whether every connector passes every scenario in the test matrix. Indexed search and separately authorized MCP actions also have different permission scopes.
Omni does not inherit individual HubSpot permissions, and indexed data is visible to all users in the Omni workspace. Teams handling confidential HubSpot records should account for this limitation when deciding what data to index and who should be able to access the resulting information.
This article does not include measured pass rates or revocation times. When running the matrix, record the observed outcomes, including failures and any delays after permissions change.
For more on the retrieval architecture, read our enterprise RAG benchmark.
FAQs
Does enterprise AI respect document permissions? +
It depends on the product and connector. Source-system permissions do not automatically carry over to an AI agent. Test with separate user identities, private and shared documents, revoked access, and queries spanning multiple sources.
What happens when access is revoked? +
The system’s behavior depends on how the connector processes permission changes and on its documented security contract. Revoke access in the source application, repeat the query, and record when the content stops appearing.
Which Omni connectors inherit per-user permissions? +
It depends on the connector. Google Drive maps file access rules, Slack uses channel membership for indexed messages, and Linear uses team permissions where available. HubSpot does not inherit individual record permissions; indexed records are visible to all users in the Omni workspace.