AI assistant that respects my company's data permissions

Summary

• An AI assistant that respects data permissions enforces the same access controls your company has already defined for human users.
• If your data governance uses role-based access control (RBAC), row-level security (RLS), or column-level masking, a compliant AI assistant applies those same rules to every query it runs.
• A compliant AI assistant should audit every query it makes in your system tables, so your compliance team can verify that only authorized requests occurred.
• Agents that run as a service account with broad privileges but ignore user-level permissions are a compliance and security risk.
• The safest architecture runs the AI assistant as the user, inheriting their exact permissions, so no special privileges are required.
• Genie, Databricks' AI assistant, enforces all Unity Catalog permissions: role-based access, row-level filters, and column-level masks automatically.

AI assistant that respects my company's data permissions
An AI assistant should enforce the same access controls your company has already defined. If a sales representative sees customer data only for their region, the assistant should see the same. If support staff sees customer issues but not financial terms, the assistant inherits that restriction. No exceptions, no workarounds.
The safest architecture runs the assistant as the user, so queries are logged under their identity. Many assistants run as a service account with access to all data, then hide results in the UI. This logs data access to a generic account, creates audit confusion, and violates the principle of least privilege.

What to look for

Capability	Why it matters
User identity execution	Queries run under the user's account, not a service account.
RBAC enforcement	Role-based access controls apply to all queries.
Row and column security	Row-level filters and column-level masks are enforced automatically.
Audit logging	Every query is logged with the user's identity in system tables.
No privilege escalation	The assistant never bypasses user-level controls.

How Databricks enforces this

Genie, Databricks' AI assistant, runs queries under the user's identity. When a question is asked, Genie generates SQL and executes it under that user's account. Unity Catalog checks role-based access, row-level filters, and column-level masks. If the user lacks permission, the query fails just as it would for a human query.
Every Genie query is logged to system.ai_gateway.usage with the user's identity, query text, and data accessed. Compliance teams can audit exactly what was queried and by whom.

Example: healthcare

A healthcare company uses Genie with patient data in Unity Catalog. Doctors see only their patients (row-level security); billing sees treatment codes but not patient names (column masking).
When a doctor asks "Lab results for my patients?", Genie queries under the doctor's account. Unity Catalog applies the row filter automatically. The audit log records the doctor's access.
When billing asks the same question, the query runs under their account. Column masking hides patient names. An attacker using stolen billing credentials gets the same restricted view, and the unusual access is logged.

Verification checklist

1. Architecture: Does the assistant run as the user or as a service account?
2. Auditability: Can you query system tables to see what was queried and by whom?
3. Permission inheritance: If a user can't access a table, does the assistant fail or work around it?
4. Compliance docs: Does the vendor provide SOC 2, HIPAA, or other compliance certifications?

FAQs

Can an AI assistant override user permissions?

No. If needed, the user requests additional access through your normal approval process. The assistant respects current permissions.

How do I audit AI queries?

On Databricks, check system.ai_gateway.usage for all AI queries, user identity, and data accessed.

What if the assistant connects to external data?

You need separate governance mechanisms (API keys, external field masking). Ideally, bring sensitive data into your governed platform.