AWS Publishes Jamf’s Bedrock Budget Gate, Restricting Premium Models at 80% Spend

A developer who reaches a daily AI budget does not have to lose access to Amazon Bedrock entirely. Jamf’s production setup instead removes access to progressively pricier Claude models while preserving Claude Haiku, a lower-cost option. AWS published the implementation and its accompanying code on September 1, offering a concrete pattern for per-user AI spending controls.

The system is designed for a cost problem that arrives at the level of user behavior, not provisioned infrastructure. Jamf gave its engineering organization broad Bedrock access, then built controls that calculate each engineer’s daily use from invocation logs containing token counts, model identity and user identity.

A budget becomes an access decision

In AWS’s example configuration, access to Anthropic Claude Opus is denied once a person reaches 80% of a daily budget. At 100%, Claude Sonnet is also denied, but Claude Haiku remains available. The design is a graduated constraint: it reduces the cost of the next request instead of treating a budget threshold as a complete work stoppage.

The control loop changes permissions, not sessions

The enforcement handler cross-references spending with a DynamoDB table of individual exceptions, then publishes updated IAM Customer Managed Policy versions for users above each threshold. The policies identify users through a saml:sub condition and sit on the IAM permission set. Because IAM evaluates the new policy on the next Bedrock call, restrictions take effect within minutes without a new sign-in.

The reset is built into the same recomputation process. The Athena view uses a daily window beginning at midnight in a chosen reference time zone; after that window rolls over, users no longer over the threshold drop from the newly calculated restriction list. A fresh policy version then lifts the restriction, eliminating a separate unblock workflow.

Exceptions and pricing are part of the control plane

Jamf’s exception path is deliberately separate from editing access policies by hand. Administrators can use a Slack command to issue time-boxed higher limits; the system records who granted the exception, when it was granted and, optionally, its authorizing ticket. The next Lambda run incorporates an active exception into that engineer’s threshold.

The spend calculation depends on an explicit map of model token prices. Each model family needs its own pricing branch, and AWS’s sample prices an unrecognized model at the highest tier rather than at zero, so a new model cannot silently bypass enforcement. That fail-safe also creates an operational requirement: teams must update the price map promptly when they enable a new model.

The cheap components are not the whole bill

AWS says Lambda, DynamoDB and S3 cost well under $10 a month for hundreds of engineers in this use case. Athena is the component requiring closer sizing: its cost rises with data scanned and query frequency, and JSON logs force Athena to scan every byte even when a query selects only a few fields. AWS recommends consolidating queries or converting logs to a columnar format such as Parquet.

There are implementation constraints beyond cost. IAM managed policies retain at most five versions, so the enforcement function must remove the oldest non-default version before publishing another. Athena queries are asynchronous, requiring the function to submit, poll and then read results; its timeout must accommodate that sequence.

AWS presents the result as a governance pattern that enabled more AI access at Jamf because leadership could see and cap individual spending. The published sample makes that approach deployable for Bedrock users, but its reliability still rests on accurate model pricing, carefully sized log queries and correctly maintained IAM policy versions.