Jamf builds real-time AI cost caps for Amazon Bedrock

The challenge Jamf addressed is that AI costs scale with user behavior, not with provisioned infrastructure, making them invisible until the bill arrives. This makes proving ROI difficult, which can slow down broader AI adoption within a company. By building a system to measure, restrict, and notify on per-user spend, Jamf created the accountability needed to confidently expand access.

The solution uses a serverless architecture on AWS: an Athena view turns token counts into dollar amounts, a Lambda function runs every 15 minutes to compare spending against budgets, and Customer Managed Policies enforce the restrictions. A notable design choice is the system's idempotency, meaning each run recomputes the full restriction list from the day's cumulative spend, which simplifies the logic and requires no rollback procedures.

A key insight from running the system is that governance drove adoption, not restriction. The hard per-user cap gave leadership the confidence to increase the number of AI-enabled engineers. The article recommends keeping a low-cost model always available and treating the pricing map as a critical artifact that must be updated when new models are enabled.