Track 09 · Systems & unit costLesson 09 of 11
FinOps and chargeback
Attribute cost per team, so the enterprise will scale the spend.
- Lesson library · reference
- 40 min read
- 75 min hands-on
- Pro
What you'll be able to do
01- Explain the FinOps Framework phases (Inform, Optimize, Operate), the FOCUS billing specification, and the difference between showback and chargeback, to a finance audience.
- Design an attribution schema (tenant, business unit, cost centre, use case, agent version, traffic type) and enforce it at the model gateway so every call is tagged.
- Implement a month-end allocation that combines direct token spend from gateway logs with shared platform costs split by usage drivers, and reconciles to the invoices to the cent.
- Choose allocation drivers for shared pools (platform, vector database, observability, committed capacity) and defend them to the business units that pay.
- Diagnose untagged spend and cost anomalies, and set budgets and alerts that catch a runaway agent on day one rather than at month end.
Before you start
- Cost per resolved case: cost per call, per task and per resolved case. This file assumes you can compute them.
- Routing models to fund the rollout: the gateway is where tags are enforced and where hard budget limits live.
- Tracing agents with OpenTelemetry: how attribution metadata rides on spans across an agent's sub-calls.
- Authorization and multi-tenancy: the tenant identifier you attribute against.
- Registries for models, prompts and agents: the source of valid agent IDs, versions and owners.
The short version
Finance will not let a business unit scale AI spend it cannot see. Attribution means every model call carries tags (who, which use case, which agent version, which environment) from the gateway onward, so token spend can be summed per team from the gateway's usage log. Shared costs that no single call owns (the platform, the vector database, observability, committed capacity) are split by an agreed usage driver. Start with showback (reporting what each team consumed) and move to chargeback (posting the cost to their budget) once tag coverage is high and the drivers are agreed. Budgets and anomaly detection on the same data catch runaway agents within a day.
- The gateway's per-call log is the source of truth for token spend; the cloud bill tells you totals, not teams.
- Enforce tags at the gateway from day one. Retrofitting attribution is a reconstruction.
- Shared costs need a driver that reflects what causes the cost, agreed with the people who pay, before the first bill.
- Allocations must reconcile exactly to invoices, or finance will not trust any of it.
- Untagged spend gets its own line and an owner chasing it, never quietly spread across everyone.
Practise this in the lab
Pro lesson
The rest of this lesson is in Pro.
The mental model, worked example, field note, failure table, exercise and self-check, with the voice tutor.
- Every stage and specialisation, in full
- Every lab and every role project
- Halden weeks 2–10, on your record
- Every drill, live in the playground
- All solutions, plus the server judge
- Lab Assistant and every Teach with AI