FinOps for AI
Financial accountability belongs
at the same chokepoint as safety.
Most governance tooling stops at safety and compliance, leaving cost to a separate product that reconciles invoices after the money is spent. Consumption control is the same call, the same ledger, and the same enforcement point.
The Argument
Cost is a governance problem, not a reporting problem
Agentic systems make consumption unpredictable in a way that traditional software does not. A single request can fan out into dozens of model calls, and the cost of a workflow is not knowable until it runs.
Treating that as a finance-team reporting exercise means the first signal arrives on an invoice, weeks after the behaviour that caused it. By then the only available action is explanation.
Enforcement has to sit where the spend happens. Because every model call already passes through one governed entry point, that point is the natural place to check a budget — before the provider is contacted, not after the bill arrives.
Budgets
Scope, threshold, and cap
A budget is policy, not consumption. It defines a limit and what should happen as usage approaches and crosses it.
Scope
Set at organization level for a whole company, or at use-case level for a single project. Both can be active at once.
Period
Monthly by default. Usage is aggregated across the period from ledger rows written in that window.
Alert threshold
A percentage — 80% by default — at which the budget moves into warning status and surfaces on the meter in amber.
Hard-cap flag
Determines whether crossing the limit blocks or merely alerts. This is the difference between a guardrail and a dashboard.
Enforcement
Only a hard cap stops work
The distinction matters operationally. Most budgets should inform; a small number should enforce.
A hard-capped budget at or over its limit blocks the call and returns an explicit rate-limit response with the reason. The work stops, visibly, with a cause attached.
A budget without a hard cap surfaces warning and over-budget status on the meter but never interrupts a request. It alerts; it does not stop the work. This is the right default for most teams — the cost of an unexpected block usually exceeds the cost of an overage.
Reserve hard caps for scopes where runaway consumption is worse than interrupted work: experimental workspaces, untrusted tenants, and automated pipelines with no human watching.
Reconciliation
Estimate versus billed reality
Per-call token estimates are useful for control but are not the invoice. The gap between the two is where confidence in cost governance usually breaks down.
Per-call estimate
Tokens captured at call time and converted to approximate cost using blended per-model rates. Available instantly, on the meter.
Provider-reported usage
Metered usage and cost pulled from the provider's own admin API for the current period, giving the billed figure alongside the estimate.
The delta
Closing the gap between estimate and billed reality is what makes the number defensible in a budget conversation.
Derived, not stored
Consumption is always aggregated from the audit ledger rather than kept as a separate counter, so the meter cannot drift from what executed.
Know the spend before the invoice explains it
See how budgets, thresholds, and hard caps work at the same enforcement point as safety controls.