When we built per-user quotas, the idea was simple. You already put cost controls on your infrastructure, so you should be able to put them on your people too. Customers who are rolling AI out to large parts of their organization keep asking for the same two things:
- Visibility into AI spend for each individual
- A way to cap that spend so one person can't run up a surprise bill
Per-user quotas give you both. Since the first release we've shipped a lot based on what customers told us they needed: roles and named users for scoping, weekly limits, per-user overrides for exceptions, coverage for Cortex AI Gateway, a self-service usage view for end users and a new account usage view of AI spend by user. This post covers what quotas do, how they work and what's new.
The gap budgets leave
Snowflake Budgets let you set monthly credit limits on groups of resources: a set of warehouses, an AI service, a tagged cost center. They work well for tracking aggregate spend across a team or a workload.
Aggregates have a blind spot, though. Say an analyst runs a Cortex AI function against a million-row table, or a developer leaves a Cortex Agent running in a loop while debugging. That cost lands on the whole group. The budget shows the team's total going up, but it doesn't show which person caused it. The admin gets an alert, and the user who spent the money usually finds out last.
A quota watches spend per person instead of per resource. Every user in scope gets their own credit limit, and each user is evaluated separately. When someone gets close to their limit, they get a notification. If block enforcement is on and they hit the limit, they lose access to the governed service until the next cycle. Nobody else is affected.
This is what lets you open AI up to everyone. You can give every analyst AI Functions, every developer Cortex Agents and Snowflake CoCo, and every business user Snowflake CoWork, and still control what each person spends.
What quotas govern
Quotas track user-attributable spend across six domains:
| Domain | What it covers |
|---|---|
| Warehouse | Query execution credits attributed to the user, based on Query Attribution History |
| AI Functions | Cortex AI SQL functions such as AI_COMPLETE, AI_CLASSIFY, AI_EXTRACT, AI_FILTER, AI_SENTIMENT, AI_SUMMARIZE and AI_TRANSLATE |
| Snowflake CoWork | Credits consumed in Snowflake CoWork |
| Cortex Agents | Credits from Cortex Agents |
| Snowflake CoCo | Cortex Code usage across Snowsight, the CLI and Desktop |
| Cortex AI Gateway | Inference requests routed through Cortex AI Gateway, including OpenAI- and Anthropic-compatible calls from SDKs and coding agents |
Warehouse spend and AI spend need separate quotas because they're billed in different credit types. For AI, one quota can cover a single domain or several. For AI Functions and Cortex Agents you can also target a specific function or a specific agent. That comes in handy: You might set a tight limit on AI Complete, where one bad query can get expensive, and a looser limit on everything else.
Cortex AI Gateway support
Cortex AI Gateway gives your apps and third-party agents a single, governed endpoint for reaching models. You can now put that traffic under a quota:
CALL my_quota!ADD_SHARED_RESOURCE('AI GATEWAY');
Once the gateway is added, inference that a user sends through it counts toward their limit, and it's blocked at the limit like any other AI domain. If you want team-level tracking too, you can add the gateway to a custom budget as a shared resource.
Choosing who's in scope: Tags, roles and names
A quota can apply to every user in the account, or to a group of users you define. The way you define that group has changed the most since launch.
(NEW) Roles. You can now select users by the roles that have been granted to them directly, alongside or instead of tags. For most Snowflake customers, roles already describe how people are organized, and many identity providers already sync them through SCIM. If roles drive scope, granting or revoking a role moves a user into or out of a quota with no tagging step in between.
CALL my_quota!SET_QUOTA_USER_SCOPE(PARSE_JSON('{
"applies_to": {"type": "GROUP_OF_USERS", "roles": ["ANALYST", "DATA_ENG"]},
"excluded_by": {"roles": ["ACCOUNTADMIN"]}
}'));
Keep in mind that this selects users. It isn't about which role is active when they run a query. A user is in scope if they've been granted the role directly. Inherited roles don't count.
Tags. Tags still work the way they always have. Each user can carry key-value tags like COST_CENTER = 'engineering', and the quota picks up whoever currently matches. Tag changes take effect without editing the quota. With multiple tags, UNION (the default) covers anyone who matches at least one tag, and INTERSECTION covers only users who match all of them. You can write the snowflakeTags tag and value through SCIM from your identity provider into Snowflake as people join, move teams or leave. One caveat: Microsoft Entra ID can add and update tags through SCIM but can't remove them, so check how departures show up before you rely on tags alone.
Named users. Rules always have exceptions, like a contractor who isn't tagged yet or a service account that shouldn't be capped. You can now include or exclude specific users by name. Naming a user overrides the tag and role rules: An included user is monitored even without a matching tag or role, and an excluded user stays out even if they match. Their tags and grants aren't touched.
CALL my_quota!INCLUDE_USERS(['ALICE', 'BOB']);
CALL my_quota!EXCLUDE_USERS('USER', ['CAROL']);
Put together, a quota's scope works in three steps, applied in this order: include a group (all users, or users selected by tags, roles or both), exclude groups by tag or role, then always include or always exclude specific users by name. You can now also exclude by tag on quotas that are scoped by tag, not only on all-users quotas. The SET_QUOTA_USER_SCOPE, ADD_QUOTA_USER_SCOPE and REMOVE_QUOTA_USER_SCOPE methods handle all three steps, and GET_QUOTA_SCOPE_V2 returns the full picture. The quota wizard in Snowsight follows the same three steps.
Setting limits: Monthly, weekly and daily
Each quota sets a default credit limit for every user in scope. Users are evaluated separately, and there's no shared pool. Three cycles are available, and you can set any combination of them:
- Monthly resets at 00:00 UTC on the first of the month. Use it to manage total spend over time.
- Weekly follows the ISO week and resets every Monday at 00:00 UTC.
- Daily resets at 00:00 UTC every day. Use it to catch runaway jobs before they burn through a month's allowance in an afternoon.
CALL my_quota!SET_PER_USER_LIMIT(500, 'WEEKLY');We added weekly limits because customers told us monthly was too slow to react and daily was too strict for people whose work comes in bursts. A weekly limit lets someone have a heavy Tuesday without getting blocked, as long as the week stays reasonable. A common setup is a monthly limit for the total plus a weekly or daily limit to catch spikes. Each cycle is checked on its own, so a user is blocked as soon as they hit any of their limits.
New: Overrides for individual users
A single limit for everyone is easy to manage, but it doesn't always fit. One engineer is prototyping an agent this month. A new hire should start lower. Someone has a quarter-end project. Before, the only fix was to create a separate quota or raise the limit for everyone.
You can now override the limit for specific users on a specific cycle, with an optional expiration date:
CALL my_quota!SET_PER_USER_LIMIT_OVERRIDE(
['ALICE'],
2500,
'MONTHLY',
'2026-12-01'
);
Alice gets 2,500 credits a month until December 1, and then she goes back to the quota default on her own. An override can be higher or lower than the default. It keeps the quota's domains, notifications and enforcement settings, and notification thresholds scale to the new limit because they're percentages. If Alice is blocked when you raise her limit, the block clears within a few minutes with no separate unblock step. Each quota supports up to 200 overrides. Admins can also set overrides in Snowsight from the quota or user detail pages, and users with an override are marked so they're easy to find.
Block enforcement
For AI domains, a quota can block users who reach their limit, not just alert on them. With block enforcement on:
- A user who hits their daily limit is blocked until the next 00:00 UTC
- A user who hits their weekly limit is blocked until the following Monday
- A user who hits their monthly limit is blocked until the first of the next month
- The block covers only the governed domain. If the quota covers AI Functions, a blocked user can't run AI Function queries, but everything else they do in Snowflake keeps working.
Blocks lift on their own when the cycle resets. They also lift when you raise the limit, set an override above the user's current spend or remove the quota.
This takes away the usual tradeoff. You don't have to choose between giving everyone access and accepting the cost risk, or holding access back and slowing adoption. You can roll AI out broadly and let the quota stop individual outliers.
Visibility for users and admins
New: Users can check their own usage
Getting blocked without warning is frustrating. Users already receive notifications as they approach a limit, and now they can check their usage whenever they want:
SELECT * FROM TABLE(SNOWFLAKE.LOCAL.MY_BUDGET_SUMMARY());
This returns one row for each quota and cycle that applies to the user, with their usage, their limit (including any override), the percentage used and when the cycle resets. Any user can call it, and it only returns the caller's own data, so admins don't need to grant anything. Users who can see where they stand ask the help desk "why am I blocked?" a lot less often.
New: One view of AI spend by user
Admins also need a single place to see AI spend by user, whether or not a quota covers it yet. The new USER_AI_USAGE_HISTORY view in ACCOUNT_USAGE (with a matching view in ORGANIZATION_USAGE) shows AI credits by user, by service and by object (a specific function, agent, CoWork object or gateway), aggregated by hour. It covers AI Functions, Cortex Agents, Snowflake CoWork, Snowflake CoCo and Cortex AI Gateway, and includes service users.
SELECT user_name, SUM(credits) AS total_credits
FROM SNOWFLAKE.ACCOUNT_USAGE.USER_AI_USAGE_HISTORY
WHERE usage_hour >= DATEADD('day', -30, CURRENT_TIMESTAMP())
AND user_id > 0
GROUP BY user_id, user_name
ORDER BY total_credits DESC;
It also helps with setting your first limits. Look at what your heaviest users actually spend over a month or a week, then pick limits based on that data.
Notifications and custom actions
Quotas send notifications through two separate channels.
User notifications go to the user as they approach or reach a limit. You choose the thresholds, for example 50%, 80% and 100%. Each threshold can fire on actual spend or on projected spend, meaning the user's current pace would reach that level by the end of the cycle. Projected alerts are the early warning. If someone uses 3% of their monthly limit on day one, they might get an 80% projected alert and slow down before anything is blocked. Thresholds work for monthly, weekly and daily cycles.
Admin summaries collect every threshold crossed across users in scope and send a digest to the admins you choose after each evaluation. Each entry shows the user, the threshold, whether it was actual or projected, and their usage at the time. Summaries can go to email or to a Snowflake notification integration such as a webhook, Slack, PagerDuty, Amazon SNS or Azure Event Grid.
If notifications and blocks aren't enough, custom actions run a stored procedure when a threshold is crossed. The procedure receives the list of users who crossed it, and the logic is up to you: revoke a role, write an audit row, send an email from your own template or start a downstream workflow. Each quota can also have one cycle-start action that runs at 00:00 UTC at the beginning of each cycle, which you can use to reset whatever your custom actions changed.
How fast it reacts
Quotas evaluate spend within minutes of the usage. AI costs can climb quickly, and with block enforcement on, users are blocked within minutes of going over their limit, not at the end of the day. Configuration changes take about five to 10 minutes to propagate, and scope changes from tag or role updates can take about 15 minutes.
Creating and managing quotas
You can create quotas in Snowsight under Admin » Cost Management or with SQL. The Snowsight wizard walks you through scope (users and the domains to track), limits, notifications and custom actions. You can also create and adjust quotas in plain language with the Cost Intelligence skill in Snowflake CoCo.
How quotas fit with budgets
Quotas and budgets work together at different levels:
- Budgets track aggregate spend for a group of objects or users, such as a set of warehouses, an AI service or a cost center. Use them to see whether a team or project is on track for the month.
- Per-user quotas track and enforce spend for each individual. Use them to make sure no single person drives an outsized bill.
- USER_AI_USAGE_HISTORY and the other AI usage views in ACCOUNT_USAGE and ORGANIZATION_USAGE give you detailed per-user, per-service reporting.
- MY_BUDGET_SUMMARY shows each user their own limits and usage.
- The Cost Intelligence skill in Snowflake CoCo puts all of this behind a natural language interface for anyone with access to the underlying data.
Why this matters now
Customers keep telling us the same thing: They want to roll AI out widely, and they need cost governance in place before they can do it with confidence.
We think that's the right way to look at it. Cost controls make broad adoption possible. Once finance and platform teams trust that one person can't create a surprise bill, the conversation about expanding AI access gets much easier. Governance stops being the reason to wait and becomes the reason you can move forward.
Every feature in this post came from that conversation. Roles and named users make scope easier to manage. Weekly limits and overrides fit the way people actually work. AI Gateway coverage reaches the agents and tools your developers are adopting. And the new usage views mean neither users nor admins are left guessing.
Get started
- Per-user quotas documentation: Full reference for scope, limits, overrides, enforcement, notifications and custom actions
- USER_AI_USAGE_HISTORY view: Per-user AI spend for reporting and sizing limits
- Check your quota usage with MY_BUDGET_SUMMARY
- Cortex AI Gateway
- Access control for cost management: How to grant quota admin and viewer access to your team

