Skip to content

FinOps

Target

This page describes intended behaviour or structure. No implementation is claimed unless a section is explicitly marked Current implementation or Evidence.

Cost is attributable or it is not governable. Usage facts are owned by Usage Accounting in the Control plane (BC-05); this plane owns budget policy, the FinOps views and the reconciliation process. Commercial entitlements and invoicing remain SaaS foundations (BC-33) and are not decided here (open question Q-15).

Attribution dimensions

Every usage record carries: tenant, workspace/project, environment, user or workload identity, agent and release, skill, run, step, attempt, model profile and the actually used provider/model/deployment, compute pool, tool invocation where applicable, pricing revision and currency.

Amount states

State Meaning Source
Estimated Predicted before dispatch from profile, context size and history Context planner, model gateway
Reserved Budget held before a model or tool call Usage Accounting
Observed Tokens, calls, duration reported by the provider or tool Model gateway receipts
Reconciled Matched against provider invoices or usage exports Reconciliation job
Uncertain Call timed out or reply was lost; the provider may still have billed Model gateway; stays visible until reconciled

A cancelled or timed-out call is never recorded as zero cost. Duplicate receipts are counted once (idempotency key).

Budget policy

  • Per-tenant and per-project soft and hard budgets with predictive alerts.
  • "If predicted run cost exceeds a threshold, require approval" is a policy decision evaluated at acceptance and bound to the exact run (approvals).
  • Fallback chains respect budget and privacy policy together: a cheaper external provider is never selected for a workload whose policy forbids it.
  • Quota denial is explicit and observable; a denied run is Rejected, not Failed.

Optimization levers (later)

Local-first routing, provider comparison, model downgrade within profile constraints, caching, prompt compression, batching, GPU autoscaling and serverless execution. Each is evaluated with the regression suite before adoption; a cost improvement that degrades task success is not an improvement.

Current implementation

  • ConnectSoft.Saas.MeteringTemplate has transactional usage counters, quota transitions and integration events for generic tenant usage.
  • ConnectSoft.Extensions.Saas.Metering and .Billing are marker contracts.
  • No AI token or pricing capture, reservation, receipt or reconciliation exists; the prototype's cost views are fixtures.
  • Packet proposal PK-17 (AI usage mapping onto existing metering) is the reuse path.