Amazon Bedrock cost management guide for FinOps and platform teams

AWS outlines a five-step operating model to control Amazon Bedrock costs, targeting FinOps and platform engineering teams. According to the announcement, generative AI cost management is a shared responsibility between application teams and those accountable for cloud spending. Agentic workloads, which trigger multiple calls per task, complicate spend tracking.

The recommended approach follows a continuous loop: cost attribution to an owner, observation of each request via metadata, budget enforcement with alerts and blocks, optimization of platform defaults, and regular data reviews. Fine-grained attribution is possible through the user ARN in AWS CUR 2.0, or via inference profiles per application or projects for OpenAI-compatible APIs. Centralized and automated request logging enables real-time or delayed tracking, depending on needs.

This framework targets FinOps teams, responsible for attribution and reviews, and platform engineers, in charge of logs and default settings. Its effectiveness depends on the order of steps: without attribution and observability, other measures remain approximate. Companies must adapt this model to their environment to avoid visibility gaps, especially with shared gateways or multi-model usage.