AWS has published a detailed five-stage playbook for managing Amazon Bedrock costs, aimed at FinOps and platform teams. Its central advice is less “pick a cheaper model” and more “build the plumbing to see, attribute and control what each workload spends”.
AWS AI Watch analysis
What happened
The AWS playbook sets out a loop covering attribution, observability, budget enforcement, optimisation and regular review. AWS says Bedrock inference costs can be attributed to the IAM principal making each call through Cost and Usage Report 2.0 data. Teams can tag those identities with business details such as team, application and cost centre, then use the tags to group costs in Cost Explorer.
A shared gateway role can hide which end user generated the bill. AWS’s proposed fix is for the gateway to assume an invocation role separately for each user, pass that user’s identity and tags into the session, and cache the credentials for up to an hour. For visibility into individual requests, the playbook recommends invocation logs plus request metadata identifying the application, workflow, tenant or agent.
Why it matters
An agent can make many model calls while a person sees only one task. That means user interactions alone are a poor guide to cost: autonomy and context size also shape consumption, AWS says. Teams that cannot connect spend to an owner and workload may end up arguing over averages while the bill carries on being extremely specific.
The playbook also distinguishes between seeing a problem and stopping it. Alerts and dashboards help teams spot rising usage; AWS describes harder enforcement options for cases where a warning is not enough. The right balance depends on how quickly a workload can spend and how much overspend the organisation is prepared to tolerate.
Our read
This is useful operational guidance, not a magic switch for making AI cheaper. Start by finding out who is spending what, and make request logging and metadata consistent before trying to optimise. Otherwise, cost-cutting risks becoming guesswork with a spreadsheet attached.
What to watch
- Whether teams adopt per-user attribution behind shared gateways.
- How they balance budget alerts against controls that can block use.
- Whether request metadata stays consistent across applications and teams.
Discussion spark: When AI workloads can spend faster than a person can notice, should companies favour hard budget blocks or let teams keep experimenting with alerts in place?
Sources and evidence
- Managing Amazon Bedrock costs end to end: A playbook for FinOps and platform teams (30 October 2026, 22:42 UTC)
not affiliated with or endorsed by Amazon Web Services (AWS)