Measure AI cost per successful business outcome
On this page
Use cost per quality-qualified business outcome as the primary unit metric. Include incurred costs from original attempts, failures, retries, evaluations, and allocated supporting services within a defined scope. Count an outcome only once after it passes documented business and quality thresholds.
Why this is worth a look
Cost per token is a resource-efficiency metric, while FinOps unit economics ties technology costs more directly to business outcomes. Provider telemetry supplies operational inputs such as requests, tokens, errors, throttles, latency and cache activity. Your application must add outcome identity, quality decisions, retry lineage and evaluation cost so technical activity can be related to business value and each outcome counted once.
Run this check
CHECKLISTUse this read-only contract as a worksheet for one FinOps scope and attribution window. It defines application fields and calculation logic, but performs no cloud changes and requires no cloud permissions. Record monetary amounts in the billing currency and usage in the provider-reported units.
AI OUTCOME COST CONTRACT
Scope and time window
[ ] FinOps scope: ______________________________
[ ] Reporting period: _________________________
[ ] Attribution window: _______________________
[ ] Billing currency: _________________________
[ ] Supporting-cost allocation rule: __________
Outcome definition
[ ] outcome_type: one business result, such as an accepted case resolution
[ ] outcome_id: stable internal identifier, stored in the approved analytics boundary
[ ] success_rule: business completion condition
[ ] quality_rule: evaluator, threshold, version, and human-review policy
[ ] Count a unique outcome once within the attribution window
Attempt ledger
[ ] attempt_id and parent_attempt_id for retries or fallbacks
[ ] outcome_id, provider, model, region, deployment, and processing mode
[ ] start time, end time, status, and error or throttle class
[ ] input, output, cached-read, and cached-write usage when supplied
[ ] incurred request cost, provisioned-capacity allocation, and supporting-service allocation
[ ] evaluation run ID, evaluator version, evaluation cost, and quality result
Monthly calculation
1. Group attempts and evaluations by unique outcome_id within the scope and window.
2. Sum only costs actually incurred or allocated for successful, failed, throttled, retried, fallback, evaluation, and agreed supporting activity.
3. Count each outcome_id once only when success_rule and quality_rule both pass inside the attribution window.
4. cost_per_successful_outcome = total_attributed_cost / qualified_unique_outcomes
5. Report zero-denominator periods as undefined, not zero.
Controls
[ ] Reconcile model cost to provider billing and provisioned-capacity totals.
[ ] Publish attempt success rate, retry rate, quality pass rate, and unattributed cost beside the unit cost.
[ ] Version all definitions so trend breaks are visible.How to confirm it
- 01
Define one valuable outcome
Choose a business event that can be counted once, such as an accepted case resolution, not a model response. Document completion, quality, and attribution-window rules with the product owner. Keep sensitive outcome data in the organization's approved analytics system.
- 02
Create attempt lineage
Assign each initial call, retry, fallback, and evaluation a distinct ID linked to one outcome ID. Capture provider, model, deployment or mode, status, error or throttle class, available token categories, and timestamps. Use the provider's documented telemetry as an input, not as the business-outcome count.
- 03
Build the complete numerator
For the defined reporting period and FinOps scope, reconcile metered model usage and fixed provisioned capacity to provider billing. Add failed and throttled activity only when its cost is actually incurred or allocated, along with evaluation calls and agreed retrieval, vector, safety, orchestration, and other supporting costs. Record shared-capacity allocation rules.
- 04
Apply the quality gate
Count a unique outcome only after the business completion rule and versioned quality threshold pass. Record human overrides and evaluator changes. A response that is technically successful but rejected, unsafe, or unusable remains attributed cost without a successful outcome.
- 05
Publish the ratio and drivers
Report cost per qualified outcome with the reporting period, scope, currency, volume, attempt success rate, retry rate, quality pass rate, allocated fixed cost, and unattributed cost. Trend one stable definition over time within that scope, then investigate changes in both numerator and denominator.
Before making changes
Provider monitoring supplies technical signals, not business acceptance, and coverage varies by endpoint, deployment type, model and logging configuration. The cited Google Cloud page covers fully managed MaaS endpoints and excludes self-hosted models from its dashboard. Do not infer cost from error counts alone. Reconcile activity to billed usage or allocated capacity, document shared-cost allocation, version evaluator changes and keep automated evaluation separate from business acceptance.