Cost-per-user AI attribution in multi-provider stacks: request-level tracing for FinOps teams before invoice close

TL;DR

If your AI stack runs on multiple providers, provider invoices are too coarse. Enforce deterministic business tags at the gateway and propagate them with request metadata so attribution becomes request-level and owner-aware.

Most teams discover this only during budget review, when "who spent what" is already late. Use a required schema with tenant_id, user_id, feature_id, request_id, correlation_id, model, and provider.

Why total spend dashboards fail teams

Provider billing is operationally useful but usually lacks your taxonomy. OpenAI, Anthropic, and Bedrock expose token and request-level usage, but not your business model.

Practical attribution loop

1) Tag at ingress.
2) Validate tags.
3) Keep request metadata through retries.
4) Aggregate spend by feature, user, route, model.

Provider comparison

Provider	Native usage	Business tagging	Gaps
OpenAI	Strong	Limited	Owner-level mapping still needed
Anthropic	Good	Basic	Cross-service taxonomy gaps
Bedrock	Moderate	Weak	Requires custom propagation

A lightweight way to validate traces quickly is agentcolony.org/auditor before changing billing policy.