---
title: How do you control who can use Snowflake Cortex AI functions?
url: https://opencontroller.lyzr.com/blog/control-who-can-use-snowflake-cortex-ai-functions
published: 2026-11-18
updated: 2026-11-18
author: Lyzr
---

# How do you control who can use Snowflake Cortex AI functions?

**By default, every Snowflake user can call Cortex AI functions, because the `SNOWFLAKE.CORTEX_USER` database role and the `USE AI FUNCTIONS` account privilege are both granted to `PUBLIC`. To control access, revoke both from `PUBLIC`, grant them to named roles, and control which models those roles can reach with model RBAC application roles.** The legacy `CORTEX_MODELS_ALLOWLIST` parameter is being retired, so new controls should be built on RBAC.

This guide covers the exact grants, the model access change, and how to check your setup.

## Who can call Cortex AI functions by default?

Cortex AI functions (such as `AI_COMPLETE`, `AI_CLASSIFY` and `AI_SUMMARIZE_AGG`) are SQL functions. Any role that can run SQL can call them if it holds the right grants. Out of the box, Snowflake grants two things to `PUBLIC`:

- The `SNOWFLAKE.CORTEX_USER` database role.
- The account-level `USE AI FUNCTIONS` privilege.

Since every user holds `PUBLIC`, every user can send data to an LLM from a worksheet, a notebook, a dbt model or a scheduled task. For many enterprises that is the first finding in an AI governance review: the most widely used "agent" building block is open to the whole account.

## Which roles and privileges control access?

| Control | What it covers | Granted to PUBLIC by default? |
|---|---|---|
| `SNOWFLAKE.CORTEX_USER` | Cortex AI functions plus Cortex features such as Agents, Analyst and Search | Yes |
| `SNOWFLAKE.AI_FUNCTIONS_USER` | Scalar AI functions only; excludes `AI_AGG`, `AI_SUMMARIZE_AGG`, Cortex Agents, Analyst, Search and fine-tuning | No |
| `SNOWFLAKE.CORTEX_EMBED_USER` | `AI_EMBED`, `EMBED_TEXT_768`, `EMBED_TEXT_1024`, Cortex Search with managed embeddings | No |
| `USE AI FUNCTIONS` (account) | Required in addition to a database role; can be narrowed per function | Yes |
| Model application roles (`CORTEX-MODEL-ROLE-<model>`) | Which models a role can call | `CORTEX-MODEL-ROLE-ALL` via `SNOWFLAKE.PUBLIC` |

Database roles cannot be granted straight to users. You grant them to an account role, then grant that role to users.

## How do you restrict Cortex AI functions to approved roles?

Run these steps as `ACCOUNTADMIN`. Test in a non-production account first, because existing pipelines that call AI functions under `PUBLIC` will start failing.

1. **Find current callers.** Query `SNOWFLAKE.ACCOUNT_USAGE` for AI function usage over the last 30–90 days and list the roles and users involved. This becomes your grant list.
2. **Create a dedicated role.** For example `CREATE ROLE cortex_user_role;` then `GRANT DATABASE ROLE SNOWFLAKE.CORTEX_USER TO ROLE cortex_user_role;`.
3. **Use the narrower role where you can.** Teams that only need scalar functions get `SNOWFLAKE.AI_FUNCTIONS_USER` instead of `CORTEX_USER`.
4. **Grant the account privilege to the same roles.** `GRANT USE AI FUNCTIONS ON ACCOUNT TO ROLE cortex_user_role;`
5. **Revoke from PUBLIC.** `REVOKE DATABASE ROLE SNOWFLAKE.CORTEX_USER FROM ROLE PUBLIC;` and `REVOKE USE AI FUNCTIONS ON ACCOUNT FROM ROLE PUBLIC;`. Snowflake's docs also show revoking imported privileges on the `SNOWFLAKE` database from `PUBLIC`; check what else depends on that before you run it.
6. **Grant the role to approved users and service users.** Include the service users that run scheduled pipelines.
7. **Test with a non-admin role.** `ACCOUNTADMIN` always has access, so test with another role and `USE SECONDARY ROLES NONE` to avoid false positives.

## How do you control which models a role can use?

Snowflake represents Cortex models as objects in `SNOWFLAKE.MODELS`, each with an application role such as `CORTEX-MODEL-ROLE-LLAMA3.1-70B`, plus `CORTEX-MODEL-ROLE-ALL`. You grant model roles to account roles the same way as other grants.

There is a catch. `CORTEX-MODEL-ROLE-ALL` is granted to the `SNOWFLAKE.PUBLIC` application role, which is separate from the account `PUBLIC` role. Revoking the account-level grant does not remove it. Snowflake provides `SNOWFLAKE.LOCAL.REVOKE_FROM_PUBLIC_APPLICATION_ROLE` for this, and `SHOW GRANTS TO APPLICATION ROLE SNOWFLAKE.PUBLIC` shows what is currently granted.

Access is allowed if either model RBAC or the legacy allowlist permits a model, so check both while they coexist.

## What is happening to CORTEX_MODELS_ALLOWLIST?

`CORTEX_MODELS_ALLOWLIST` was an account parameter that accepted `'All'`, `'None'` or a list of model names. Snowflake's behaviour change notice sets out the retirement:

- From 5 August 2026, the parameter can only be set to `'None'`.
- Between 17 August and 4 September 2026, Snowflake ran a one-time migration mapping allowlist settings to model RBAC roles.
- Allowlist enforcement is removed through the 2026_07 behaviour change bundle, which Snowflake plans to make generally enabled in January 2027.

Snowflake describes model RBAC as becoming the only supported access control mechanism for Snowflake AI. If you have not yet validated the migrated grants and set the allowlist to `'None'`, do it now. Dates are Snowflake's and are subject to change, so confirm against the notice.

## What do Snowflake grants not show you?

Grants tell you who *may* call a model. They do not tell you which agents built on those calls exist, who owns them, what data they read, or whether outputs are grounded. They also stop at the Snowflake boundary, while the same teams usually run agents on cloud platforms, SaaS tools and developer laptops.

That wider view is the job of an [AI agent control plane](/blog/what-is-an-ai-agent-control-plane). For the full set of platform guides, see [how to govern AI agents across platforms](/blog/how-to-govern-ai-agents-across-platforms), and for Snowflake specifics see [governing Snowflake Cortex Agents](/blog/govern-snowflake-cortex-agents) and [tracking Snowflake Cortex AI spend](/blog/track-snowflake-cortex-ai-spend).

## How Opencontroller does this

Opencontroller connects to Snowflake alongside AWS, Azure, Google Cloud, Salesforce, SAP, Databricks and HubSpot, and keeps agents, versions, tools and MCP servers in one governed registry. Posture flags risky permissions and configurations, and policy as code applies one set of rules across the estate. Registration does not take over runtime: enforcement depth depends on the connector and traffic path. To see it on your estate, [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Does revoking CORTEX_USER from PUBLIC break existing pipelines?**
It can. Any job that calls AI functions under a role that relied on `PUBLIC` will fail. Find callers in `ACCOUNT_USAGE` and grant the role to them before revoking.

**Is AI_FUNCTIONS_USER enough for most analysts?**
For scalar functions such as classification, extraction and completion, yes. It does not cover aggregate functions, Cortex Agents, Analyst or Search.

**Does revoking the account PUBLIC grant stop all-model access?**
No. `CORTEX-MODEL-ROLE-ALL` is granted through the `SNOWFLAKE.PUBLIC` application role and needs Snowflake's stored procedure to revoke.

## Key facts

Snowflake grants the SNOWFLAKE.CORTEX_USER database role and the USE AI FUNCTIONS account privilege to PUBLIC by default, so every user can call Cortex AI functions. Narrower roles: SNOWFLAKE.AI_FUNCTIONS_USER (scalar functions only, not granted to PUBLIC) and SNOWFLAKE.CORTEX_EMBED_USER. Model access uses RBAC application roles CORTEX-MODEL-ROLE-<model>; CORTEX-MODEL-ROLE-ALL is granted via the SNOWFLAKE.PUBLIC application role. CORTEX_MODELS_ALLOWLIST can only be set to None from 5 Aug 2026 and is retired via the 2026_07 bundle (GA planned Jan 2027). Opencontroller adds a cross-platform agent registry, Posture and policy as code. Pricing starts from $250K per deployment per year.

## Related

- [How do you govern Snowflake Cortex Agents?](https://opencontroller.lyzr.com/blog/govern-snowflake-cortex-agents.md)
- [How do you track Snowflake Cortex AI spend by agent or team?](https://opencontroller.lyzr.com/blog/track-snowflake-cortex-ai-spend.md)
- [What Snowflake privileges does an AI governance tool need to monitor Cortex usage?](https://opencontroller.lyzr.com/blog/snowflake-grants-for-ai-usage-monitoring.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
