Opt out of Snowflake AI features¶

Most Snowflake AI features are initially available to all users in your Snowflake account. Access to most features is controlled by the SNOWFLAKE.CORTEX_USER database role, which is initially granted to the PUBLIC role. All users are granted the PUBLIC role, giving them access to Cortex features by default. (Access to Snowflake Copilot is controlled by the SNOWFLAKE.COPILOT_USER database role, also granted to PUBLIC by default.) Two features, Cortex Analyst and Document AI, are opt-in features that are not accessible to users by default.

Opt out of default features¶

To revoke access to all Snowflake AI features that are available to users by default, revoke the CORTEX_USER and COPILOT_USER database roles from the PUBLIC role. You can grant these roles to specific roles that you want to have access to the features, then grant those roles to specific users as needed. (You cannot grant database roles directly to users, but must grant them to roles that can be assumed by users.)

Use SQL like the following to revoke access to the CORTEX_USER and COPILOT_USER roles from the PUBLIC role, then grant them to specific roles and users.

-- Revoke access to most Snowflake AI features from all users in the account
REVOKE ROLE CORTEX_USER FROM ROLE PUBLIC;
REVOKE ROLE COPILOT_USER FROM ROLE PUBLIC;

-- Optionally, grant access to specific roles
GRANT ROLE CORTEX_USER TO ROLE my_cortex_role;
GRANT ROLE COPILOT_USER TO ROLE my_copilot_role;

-- Then grant those roles to specific users
GRANT ROLE my_cortex_role TO USER alice;
GRANT ROLE my_copilot_role TO USER bob;

Copy

Note

If you granted CORTEX_USER and COPILOT_USER to other roles, revoke them from those roles to completely block users from using Snowflake AI features.

Revoke access to opt-in features¶

Some Snowflake AI features are opt-in. Access to these features is disabled by default, so unless you grant access to them, your users cannot use them. If you have granted access to any of these features, you can revoke access to individual features:

Cortex Analyst: Set the ENABLE_CORTEX_ANALYST account parameter to FALSE:
```
ALTER ACCOUNT SET ENABLE_CORTEX_ANALYST = FALSE;
```
Copy
Cortex Fine-tuning: Revoke the CREATE MODEL privilege on schemas from any roles you have granted it to.
```
REVOKE CREATE MODEL ON SCHEMA my_schema FROM ROLE my_role;
```
Copy
Document AI: Revoke the SNOWFLAKE.DOCUMENT_INTELLIGENCE_CREATOR database role from any roles you have granted it to:
```
REVOKE DATABASE ROLE SNOWFLAKE.DOCUMENT_INTELLIGENCE_CREATOR FROM ROLE my_role;
```
Copy
Note

Existing pipelines continue to operate after access to Document AI is revoked. To stop the pipelines, suspend or drop the tasks associated with the pipelines.
Provisioned Throughput: Revoke the CREATE PROVISIONED THROUGHPUT privilege on schemas from any roles you have granted it to.
```
REVOKE CREATE PROVISIONED THROUGHPUT ON SCHEMA my_schema FROM ROLE my_role;
```
Copy

Access control by feature¶

The following table has more detailed information on access control for individual Snowflake AI features:

Feature	Opt in	Main access control method	Additional access control methods
Cortex Agents		CORTEX_USER database role	USAGE on the search service that the agent queries, plus USAGE on the database, schema, and table used by the search service
Cortex AISQL		CORTEX_USER database role
Cortex Analyst	✔	ENABLE_CORTEX_ANALYST account parameter
Cortex Fine-tuning	✔	CREATE MODEL on the schema where you create fine-tuned models
Cortex Knowledge Extensions		CORTEX_USER database role	Relies on access control for the underlying Cortex Search Service
Cortex Provisioned Throughput	✔	CREATE PROVISIONED THROUGHPUT privilege on the schema where you create provisioned throughput objects
Cortex Search		CORTEX_USER database role	USAGE on the search service, database, schema, and table used by the search service
Document AI	✔	SNOWFLAKE.DOCUMENT_INTELLIGENCE_CREATOR database role	Numerous object-level privileges for creating models and pipelines
Snowflake Copilot		COPILOT_USER database role
Snowflake Intelligence		CORTEX_USER database role	Relies on access control for the underlying Cortex Agent or Search Service

Opt out of specific models and AISQL functions¶

Because the cost of using different large language models varies, you can limit access to specific LLMs via an account-level allowlist, by role-based access control, or by a combination of both. For more information, see Control model access.

ACCOUNTADMIN and AI features¶

The ACCOUNTADMIN role has complete access to all features in a Snowflake account, including Snowflake AI features. Revoking the CORTEX_USER and COPILOT_USER roles from PUBLIC does not prevent ACCOUNTADMIN from using these features. Even if an ACCOUNTADMIN’s access to AI features is revoked, a user with access to ACCOUNTADMIN can always grant access to that role (or any other role) again.

For this and other reasons, it is a best practice to grant the ACCOUNTADMIN role to trusted users only, or even more strictly, to a single user in the account which is not used for any purpose other than Snowflake account administration and whose login credentials are tightly controlled. Use ACCOUNTADMIN only for account setup and maintenance, and use other adminitstarive roles with more limited scope (that is, SECURITYADMIN, SYSADMIN, or USERADMIN) for day-to-day administration.

It is possible to prevent ACCOUNTADMIN from using Snowflake AI features that are gated by means other than role-based access control. For example, even a user with ACCOUNTADMIN can’t use Cortex Analyst if the ENABLE_CORTEX_ANALYST account parameter is set to FALSE. Of course, this user can always set this parameter to TRUE.

Monitor AI feature usage¶

To make sure that Snowflake AI features are not being used, monitor usage of Snowflake AI features using the Cortex-related views in the SNOWFLAKE.ACCOUNT_USAGE schema. These views are:

Note

CORTEX_FUNCTIONS_QUERY_USAGE_HISTORY and CORTEX_FUNCTIONS_USAGE_HISTORY log essentially the same events. It is not necessary to monitor both.

Create alerts on new data in these views to notify you when new AI features are used in your account. For example, the following SQL statement creates an alert that sends a Slack message when any AI function is used:

CREATE ALERT my_alert
  IF (EXISTS (
    SELECT * FROM SNOWFLAKE.ACCOUNT_USAGE.CORTEX_FUNCTIONS_QUERY_USAGE_HISTORY))
  THEN
    BEGIN
      CALL SYSTEM$SEND_SNOWFLAKE_NOTIFICATION(
        SNOWFLAKE.NOTIFICATION.TEXT_PLAIN('AI function used in account'),
        '{"my_slack_integration": {}}'
      );
    END;

Copy

Such alerts incur a nominal compute cost when new data is added to a Cortex usage history view, but if no AI features are used, there is no cost because no data is ever added and the alert is never triggered.

Control access to ML features¶

Snowflake ML features are not AI features, and access to them is not controlled by the CORTEX_USER role.

ML Functions¶

ML Functions employ classical machine learning techniques for forecasting, anomaly detection, classification, and other data analysis tasks. Creation of models by ML Functions is opt-in and controlled by a function-specific privilege, such as CREATE SNOWFLAKE.ML.FORECAST, on schemas. Access to trained models is controlled by the USAGE privilege on the model object. If you have granted these privileges already, revoke them to prevent users from creating or using ML Functions models. You may want to DROP any models that have already been created.

Owners of schemas can create ML Functions models in them, regardless of whether they have CREATE privileges for a specific type of model, so limit ownership and creation of schemas to trusted users. Grant specific privileges to create models within each schema only to users who need them.

Snowflake ML¶

Snowflake ML lets you build, deploy, and manage custom machine learning models developed in Python, at Snowflake scale. Creation and use of Snowflake ML objects, including the model registry, the feature store, and models and their versions, is not controlled by the CORTEX_USER role.

Snowflake ML objects are schema-level objects, which means that users can create Snowflake ML objects in any schema on which they have OWNERSHIP or an appropriate CREATE privilege (for example, CREATE MODEL REGISTRY). Therefore, access to Snowflake ML is best controlled by limiting ownership and creation of schemas to trusted users. Grant specific privileges to create Snowflake ML objects within each schema only to users who need them.

Note

Users with the CREATE MODEL privilege in a schema can also create models using Cortex Fine-tuning. However, actually using Cortex fine-tuned models requires the CORTEX_USER database role.