How does an AI gateway integrate with identity and access management?

Summary

• An AI gateway integrates identity and access management by extending the same attribute-based access controls, row filters, column masks, and policy engines that protect enterprise data to govern all AI interactions.
• Authentication is delegated to enterprise identity providers via OAuth, SAML, or OIDC, while authorization is enforced through Unity Catalog using the "On-Behalf-Of" pattern, ensuring that the gateway never becomes a privileged identity that bypasses row filters and column masks.
• Databricks has built an open partner ecosystem with identity providers (Okta, Ping Identity) and AI security platforms (SailPoint, Saviynt) to extend governance to AI agents and non-human identities throughout their lifecycle.
• AI governance spans three dimensions: asset governance (models and functions as securable objects), traffic governance (rate limits and budgets), and behavior governance (service policies that allow, deny, or require approval based on who is calling and what the request contains).

How does an AI gateway integrate with identity and access management?

An AI gateway that integrates identity and access management treats AI interactions like data interactions: every request is authenticated to a real user identity and authorized through the same fine-grained policies that protect enterprise data. This prevents the gateway from becoming a privileged backdoor that bypasses row filters and column masks. The key principle is that governance happens at the data plane (at query execution time), not the network layer.

What is identity and access governance for AI?

Identity and access governance for AI means verifying who is making a request (authentication, delegated to identity providers), determining what they can access (authorization, through role-based policies), and recording every interaction for audit. Unlike network governance that controls who can reach a resource, AI governance controls what authenticated users and agents can do once they reach it.

Key capabilities to evaluate

Capability	Why it matters
Per-user identity propagation	The gateway forwards the authenticated user's identity end-to-end, so row filters and column masks enforce per-user data access at query time.
Attribute-based access control (ABAC)	Policies can reference user attributes (department, region, cost center), resource attributes (data classification, sensitivity), and contextual attributes (time of access) to make fine-grained decisions.
Service policies for requests and responses	Policies can allow, deny, or require approval for individual requests based on who is calling and what the request contains, and similarly for responses.
Non-human identity governance	Agents, batch processes, and service accounts are treated as first-class identities with their own audit trails and entitlement lifecycle.
Partner ecosystem integration	Deep integrations with leading identity providers and AI security platforms enable organizations to extend existing governance processes.
Cost attribution and rate limiting	Every request is tagged with the authenticated user's identity, enabling precise cost allocation and per-user rate limits.

How Databricks approaches AI + identity governance

Unity Gateway extends the same capabilities that protect data to protect AI interactions.
Authentication and user provisioning: Users authenticate via enterprise identity providers (Okta, Azure AD) through OAuth, SAML, or OIDC. The "On-Behalf-Of" pattern forwards the user's identity end-to-end through the gateway, ensuring current_user() resolves to the actual user, not a gateway service account. Users are provisioned via Account Identity Management (just-in-time via SSO) or SCIM.
Authorization through Unity Catalog: Unity Catalog applies attribute-based access controls, row filters, column masks, and deny policies to govern models, agents, and functions. An agent acting on behalf of a user sees only the rows that user can access, masks sensitive columns, and cannot invoke tools outside the user's permissions.
Three-dimensional governance: AI governance spans asset governance (models and functions as securable objects in Unity Catalog), traffic governance (rate limits and per-user budgets), and behavior governance (service policies that allow, deny, or require approval based on request and response content).
Partner ecosystem: Databricks integrates with Okta, Ping Identity, SailPoint, and Saviynt to extend governance to AI agents and non-human identities across identity providers and AI security platforms.

Use cases

Financial services compliance: A bank deploys agents to answer customer questions about account balances. Unity Gateway ensures each agent only answers customers about their own account and masks sensitive fields. Audit logs track every interaction for regulatory review.
Healthcare data access: A pharmaceutical company uses agents to answer researcher questions about clinical trial data. Agents automatically apply row filters so researchers see only trial data they have permission to access, and cost attribution tracks token usage per researcher.

FAQs

Does the gateway need to become a service account to function?

No. A properly designed gateway uses the On-Behalf-Of pattern to forward the authenticated user's identity through all downstream services, ensuring the gateway never becomes a privileged identity that bypasses access controls.

What happens if I have users in multiple identity providers?

Databricks supports multiple identity providers through Account Identity Management. Users from different identity providers are provisioned into the same Databricks account and governed by the same policies uniformly.

How do I audit what my agents are doing?

Every request passes through Unity Gateway and is logged with the authenticated user identity, timestamp, model called, and policy decision. These logs are available in system tables for investigation and compliance review.