Skip to content

PATTERN Cited by 1 source

Tagged inference-profile cost attribution

Intent

Attribute use of a shared foundation model to organizational owners without creating a distinct AWS caller identity for every owner. The application resolves an authenticated request to a trusted cost boundary, then invokes a tagged application inference profile that points to the common model.

Structure

authenticated request
        │
        ▼
trusted owner / department resolution
        │ validate against known mapping
        ▼
owner → tagged inference-profile ARN
        │
        ▼
shared IAM role invokes Bedrock + underlying model
        │
        ▼
Cost Explorer groups billed usage by activated profile tag

Three elements are required:

  1. Trusted ownership resolver — claims, directory lookup, or a server-side session field identifies the department. Never accept an arbitrary client-supplied tag or ARN.
  2. Allowlisted routing map — maps the known boundary to a known profile ARN and rejects unknown values.
  3. Billing-plane tag activation — profile tags must be activated at the payer level before Cost Explorer can group costs by them.

Why this is distinct from caller-identity attribution

Per-caller billing works when every team invokes Bedrock under its own IAM principal. It fails when a shared application role is intentionally the only AWS principal. This pattern preserves that operational model while making the selected Bedrock resource, rather than the caller, the attribution carrier. It is a Bedrock-specific realization of patterns/chargeback-cost-attribution and a sibling of patterns/eks-cost-allocation-tags, which derives billing dimensions from Kubernetes runtime metadata instead.

Operational rules

  • Give the application permission on both the profile and underlying-model ARNs.
  • Treat the profile ARN as versioned application configuration; profile deletion or recreation changes the identifier.
  • Use CloudFormation when the profile population is large or managed centrally.
  • Keep near-real-time token telemetry separately: Cost Explorer may take 24–48 hours to show newly activated tags.
  • Add budget, anomaly, quota, or model-allowlist controls separately; resource attribution alone is observability, not enforcement.

Trade-offs

Advantage Trade-off
Shared application credentials remain simple Application becomes responsible for correct boundary resolution
One foundation model serves many owners Billing tag does not isolate prompts, data, or model access
Cost Explorer provides an AWS-wide financial view Reporting is delayed and therefore not a per-request guardrail
Wildcard profile policy makes onboarding easy It broadens the role's ability to invoke profiles

Failure modes

  • Client-controlled department selection — users can shift their spend to another team; resolve ownership server-side.
  • Unknown-boundary fallback to raw model ID — traffic becomes untagged; reject rather than silently bypass the profile.
  • Inactive or inconsistently cased tag — no expected Cost Explorer breakdown appears.
  • Profile lifecycle drift — a recreated profile has a new ARN and causes invocation failures until the map is redeployed.
  • Treating delayed bills as a budget limit — cost spikes remain unbounded for the reporting delay; pair with runtime telemetry and quota enforcement.

Seen in

Last updated · 622 distilled / 1,953 read