Accepted-work economics
AI Agent Cost per Accepted Task
Measure the complete cost of accepted agent work with stable cohorts, transparent review effort and a denominator that cannot hide rejected or pending tasks.
Cost per accepted task measures the cost of producing business work that meets an agreed acceptance contract. It is more informative than cost per model call when an agent can retry, generate rejected drafts or transfer substantial work to reviewers. The metric is only credible when its cost scope and accepted-task denominator describe the same work.
This guide helps finance, operations and AgentOps owners build that denominator, reconcile costs and compare changes fairly. The worked example uses supplier-catalogue enrichment with hypothetical SAR amounts. These are internal modelling assumptions, not market prices, service quotations or observed customer results.
Choose the unit the business actually accepts
In the example, one task enriches one admitted supplier-catalogue record with an approved product classification, required attributes and source references. Acceptance requires a reviewer or agreed validation process to confirm the correct product identity and supported values. A generated record, a completed API call and a record accepted into the catalogue are different events.
Define whether corrections remain part of the original task. Usually, reworking the same admitted record should add cost to that task rather than create extra accepted units. If the business genuinely requests a new revision, identify it explicitly and explain how the metric handles versioned work.
The outcome-evaluation contract supplies the acceptance criteria. Finance should not independently redefine success as “the agent returned a response”, and engineering should not redefine the cost boundary to exclude inconvenient review. Agreement between those owners is the first control against a misleading productivity claim.
Use a cohort with a stated observation window
A cohort groups tasks admitted under a consistent rule, such as a defined intake period, task class and release. Give outcomes enough time to mature under a stated observation window. Do not divide this month's spend by whatever happened to be accepted this month if those acceptances largely belong to a different intake cohort.
For the hypothetical calculation, each compared cohort contains 1,000 admitted catalogue records and uses the same 7-day observation window after admission. The example assumes comparable record difficulty and source availability. Those assumptions would need verification in a real release comparison.
Keep pending and rejected work visible. A pending record has not yet contributed an accepted unit, but its incurred cost remains in the cohort. A rejected record also consumes resources. Removing their costs makes the metric a price for successful attempts rather than the economics of delivering accepted work.
Build the numerator from nonoverlapping cost categories
| Category | Assumed amount | Boundary |
|---|---|---|
| Model inference | 480 | All attributed calls, including failed attempts. |
| Chargeable tools and data access | 320 | Attributed operations under the chosen scope. |
| Allocated platform operation | 700 | Documented share of runtime and operating infrastructure. |
| Initial review | 1,900 | Human review effort before corrective rework. |
| Corrective rework | 850 | Additional effort and resources not already counted above. |
| Total | 4,250 | Nonoverlapping costs for the admitted cohort. |
Make the allocation method explicit. Platform operation might be attributed using measured consumption, a documented capacity share or another justified driver. The FinOps Foundation unit-economics framework connects resource costs with meaningful business units and emphasises clear metric definitions. The catalogue contract and allocation choices here are a proposed operational design.
Decide whether the view includes development investment, supplier management, incident recovery and taxes. Different views can be useful: a variable operating metric for optimisation and a broader fully attributed metric for service decisions. Label them clearly and do not compare one scope against another.
Avoid double counting. If corrective model calls already sit in inference, the rework category should not include them again. If reviewer cost is derived from recorded effort, use an approved internal costing method and make clear whether it represents incremental cash spending or allocated staff capacity.
Calculate accepted-task cost and show the unaccepted work
Suppose cohort A has 850 accepted records, 100 pending records and 50 rejected records at the stated cutoff. Its cost per accepted task is SAR 4,250 divided by 850, or SAR 5.00. The denominator is 850, not 1,000 and not the number of model attempts.
Report the accepted share as a separate operational measure: 850 out of 1,000, or 85%. That helps explain why the cost has its current value. The cost metric alone cannot distinguish expensive high-quality delivery from a cheap process that accepts very little work.
When pending cases later resolve, update the cohort under a published restatement rule or show a later maturity view. Preserve the earlier snapshot. Do not quietly add accepted outcomes while omitting the extra review cost required to obtain them.
A cheaper model call can produce a more expensive task
Now assume a comparable cohort B uses a cheaper routing configuration. Model spending is SAR 400, tools SAR 300, platform allocation SAR 700, initial review SAR 2,400 and corrective rework SAR 1,000. Total cost is SAR 4,800. At the same observation point, 800 records are accepted, 150 pending and 50 rejected.
| Measure | Cohort A | Cohort B |
|---|---|---|
| Admitted records | 1,000 | 1,000 |
| Model spending | SAR 480 | SAR 400 |
| Total attributed cost | SAR 4,250 | SAR 4,800 |
| Accepted records | 850 | 800 |
| Cost per accepted task | SAR 5.00 | SAR 6.00 |
The assumed model saving did not create a complete-task saving. More review and rework, combined with fewer accepted records, increased the unit cost. This arithmetic illustrates a decision method; it does not prove that cheaper models generally perform worse or that a particular model route caused the difference.
Before attributing causality, check record mix, source quality, reviewer availability and acceptance policy. A controlled comparison or carefully matched cohorts can strengthen the inference. Do not credit or blame the model for a staffing change that happened at the same time.
Protect the denominator from gaming
Keep acceptance authority independent of the agent's own completion claim. Track whether validation rules or review strictness changed between periods. Lowering the acceptance standard can improve the metric while making the catalogue less useful. Record policy revisions alongside the cost series.
Do not split one record into several trivial “tasks” solely to inflate accepted volume. Conversely, a genuine change in task granularity may be justified, but it creates a new metric definition and requires a bridge comparison. Stable units are necessary for a trustworthy trend.
Watch admission changes as well. Excluding hard records can lower accepted-task cost while reducing service coverage. Report the intake, admitted and excluded populations with their reasons. A narrow service can be valuable, but its economics should not be presented as covering the full catalogue workload.
Segment before deciding what to optimise
Break the metric down by meaningful cost drivers such as source format, product family, language or missing-attribute complexity. Avoid creating tiny segments whose averages look precise but rest on very few records. Show counts and uncertainty alongside material differences.
For bilingual catalogues, compare equivalent acceptance requirements. Arabic transliteration review may differ from English attribute extraction; both can be legitimate work. Do not call one language inefficient without examining the task mix, reference quality and required review.
Use runtime budget evidence to identify loops and repeated context, then inspect whether reducing them preserves acceptance. Use calibrated semantic evaluation only where it measures a relevant property. A judge score is not itself a cost denominator unless the business contract explicitly makes that judgement sufficient.
Connect the metric to a practical decision
A rising cost can justify a targeted investigation, not an automatic shutdown. The agent may now handle harder records, include previously hidden review or operate under a stronger quality standard. Explain those changes before comparing against an old baseline.
Choose actions that address the observed driver: improve source structure, remove repeated retrieval, adjust review sampling within approved risk limits or narrow unsupported tasks. Estimate the expected effect and verify it on a comparable cohort. Do not present an assumed saving as realised before the acceptance and cost evidence arrives.
Use portfolio FinOps to reconcile the unit view with the total bill, and the AgentOps scorecard to place it alongside quality, turnaround and intervention. The unit measure should help the service owner choose, while the portfolio view ensures the costs add up.
Handle zero accepted output honestly
If a cohort has incurred cost but no accepted tasks, the ratio has no finite value under this definition. Do not display zero cost per accepted task or substitute the number of attempts to avoid an uncomfortable result. Show the total cost, admitted count, pending and rejected states, and the reason no outcome has yet qualified.
For small cohorts, a single acceptance can move the ratio substantially. Preserve the counts and maturity window, and avoid presenting a precise-looking decimal as a stable forecast. If management needs a planning estimate, label the assumptions separately from observed performance. The distinction is particularly useful during launch, when review delays and incomplete attribution can make early unit economics provisional.
Specify the reporting contract
Publish the task definition, admission rule, acceptance version, observation window, cost inclusions, allocation driver, currency basis and treatment of late charges. Identify the owners who can change each definition. Keep estimates visibly separate from final reconciled figures.
Discuss agent cost-per-accepted-task measurement with a masked cohort extract, the review process and a reconciled cost sample. The practical objective is a unit metric that reveals the full effort required to deliver usable work and remains comparable when the agent changes.
FAQ
Your questions, answered.
What is the formula for AI agent cost per accepted task?
Divide the attributed cost of a defined task cohort by the number of tasks in that cohort that meet the published acceptance contract at the stated observation point. Include the costs of pending and rejected work within the chosen scope. State cost inclusions, allocation rules and maturity window so another reviewer can reproduce the result.
Should rejected and pending tasks contribute to the cost numerator?
Yes, when their costs belong to the measured admitted cohort. They have consumed resources even though they have not produced accepted units. Removing those costs understates delivery economics. Show their counts separately and use a consistent rule for later acceptance, extra rework and restated cohort results.
How do we include human review without double counting it?
Separate initial review from corrective effort and use an approved internal costing method. Ensure model or tool costs already recorded elsewhere are not repeated in a broad rework estimate. Label whether staff cost represents incremental expenditure or allocated capacity, and preserve the same treatment across compared cohorts.
Can lower model spending increase cost per accepted task?
Yes. A cheaper configuration may require more review, retries or correction, or produce fewer accepted results. The worked example shows that possibility using assumed costs. Test actual comparable cohorts before attributing the difference to the model, because workload mix, source quality and reviewer capacity can also change the result.
How should late acceptance change the metric?
Use a published restatement rule or a later maturity view for the original cohort. Add the corresponding extra costs as well as newly accepted units, and preserve the earlier snapshot. Moving late acceptances into an unrelated current-period denominator can disconnect the outcome from the spending that produced it.
What prevents a team from gaming the accepted-task denominator?
Use a stable business unit, independent acceptance authority and versioned criteria. Report admission exclusions and avoid splitting one task into many trivial units. If task scope or quality standards legitimately change, declare a new definition and bridge the comparison rather than presenting the new ratio as an unchanged trend.