Multi-cloudAI costs4 min read

Measure AI cost per successful business outcome

Sources checked September 10, 2026Varies by scope
On this page

This check is due for a source refresh. Confirm the current documentation before you rely on provider-specific details.

THE SHORT ANSWER

Divide one workload's spending for one period by unique accepted outcomes, including in-scope failed attempts and retries. Otherwise request metrics can understate the cost of delivering usable business value.

What you need first

Tool
Use a text editor or spreadsheet as a read-only worksheet for existing reports. Start with cost and usage reports plus application outcome records for one workload and reporting period.
Access
Ask authorized cloud, finance and application colleagues for read-only cost, usage and outcome results for the agreed scope. The worksheet itself needs no direct cloud access.
If you do not use that tool
Give a cloud engineer the accounts or projects, models or deployments, supporting services and dates. Request usage metrics with units and coverage limits, matching cost totals, and accepted outcomes linked to attempts.

Why this is worth a look

Skipping this measure can make AI delivery look efficient when requests rose but accepted results did not. Amazon Bedrock Invocations counts successful API requests, while Azure OpenAI Requests counts API calls. Neither proves business acceptance. Pair scoped spending with application acceptance records and compare the same workload over time.

Run this check

CHECKLIST

Complete this worksheet in a text editor or spreadsheet from existing reports. These are measurement choices, not provider query columns. Keep billing currency and usage units visible.

AI cost per accepted outcome worksheet
AI COST PER ACCEPTED OUTCOME
Read-only worksheet. Review reports without changing cloud resources.

1. Define scope
[ ] Workload and accounts/projects: __________________________
[ ] Models/deployments and supporting services: ______________
[ ] Reporting start/end and time zone: ______________________
[ ] Attribution window for linking costs to outcomes: _________
[ ] Billing currency and shared-cost allocation rule: _________

2. Define acceptance
[ ] Business outcome or value proxy: _________________________
[ ] Completion rule: ________________________________________
[ ] Quality threshold, evaluator and rule version: ___________
[ ] Stable application outcome identifier: __________________
[ ] Count each outcome once after both rules pass.

3. Reconcile records
[ ] Link attempts, retries, fallbacks and evaluations to outcomes.
[ ] Record usage metrics with units and coverage limits.
[ ] Count accepted outcomes from application records, not requests.
[ ] Include in-scope failed-attempt, retry and evaluation costs.
[ ] Allocate supporting-service and capacity costs once.
[ ] Reconcile attributed costs to the scoped billing total.
[ ] Record unattributed costs; do not silently drop them.

4. Calculate
[ ] Total attributed cost: _________________________________
[ ] Unique accepted outcomes: ______________________________
[ ] Cost per accepted outcome = attributed cost / outcomes.
[ ] If outcomes = 0, report the ratio as undefined.
[ ] Report scope, period, window, currency, outcome count and
    unattributed cost beside the ratio.
[ ] Mark the ratio incomplete if links or reconciliation are incomplete.
[ ] Record scope, acceptance and allocation changes before comparison.

How to confirm it

  1. 01

    Define one accepted result

    Choose a countable outcome with the product owner, such as an accepted case resolution. Record completion and quality rules. Set the attribution window for linking costs and outcomes.

  2. 02

    Request matching reports

    Ask cloud, finance and application owners for the same workload and period. Request links from attempts to outcomes so retries do not become extra accepted results. Keep usage units visible, such as token counts or requests per second.

  3. 03

    Reconcile spending before dividing

    Compare attributed costs with the scoped billing total and record any gap. Apply the allocation rule once. For Bedrock, keep cache-read input tokens separate because they are charged at a reduced rate. Do not infer a charge from an error or throttle count alone.

  4. 04

    Calculate and qualify the result

    Divide attributed spending by unique accepted outcomes. Report billing currency per accepted outcome beside the outcome count and unattributed cost. Mark zero-outcome ratios undefined and incomplete records incomplete. Check scope and rule changes before comparing periods.

Before making changes

Treat the ratio as complete only with stable outcome identifiers, matching cost records and an agreed allocation rule. Keep the defined scope, period and attribution window fixed. Bedrock metrics here cover bedrock-runtime, not bedrock-mantle. The Gemini Enterprise Agent Platform dashboard covers fully managed model endpoint API calls, not self-hosted models or console activity. Monitoring metrics are not a complete billing record.

Skip this ratio for experiments with no countable business outcome or agreed value proxy. Track spending and quality separately until the product owner defines a consistent result and a rule for linking it to costs.

Primary sources