# AI Inference Economics: What Does an Accepted Result Cost?

> Calculate cost per accepted AI output using all cohort attempts, tools, review and rejected work. Includes fictional arithmetic and tenant attribution.

Source: https://grandrounds.tech/insights/ai-inference-economics · Published 2026-11-13 · CTO Grand Rounds

**Cost per accepted workflow result = all in-scope cost for a closed cohort ÷ distinct outputs accepted under the agreed standard.** Inference cost—the charge for running the model, including retries—is one numerator line beside tools, review and support. Rejected work stays in that numerator; only accepted outputs enter the denominator. For current token-price comparisons rather than workflow unit cost, see LLM Provider Cost Comparison.

*Figure: Receipts for attempts and review sit above a fraction line, with a distinct accepted-output document below, illustrating a full-cohort unit-cost calculation.*

**What does one accepted AI result cost after retries, tools, and human review?** A token rate cannot answer that question. Neither can the cost of the final successful model call if earlier attempts and rejected work disappear from the calculation.

For an AI SaaS company with real usage and paying tenants, the useful starting unit is one defined workflow output accepted against a stated quality standard. Calculate the cost of producing those outputs—including the work that did not become one. This is a cost-per-outcome method, not a diagnosis of your margin, a billing design, or a recommendation to change prices.

## Define “accepted” before dividing anything

Choose one workflow: for example, preparing a source-backed research brief for a customer reviewer. Define the output boundary, intended use, required evidence, review method, and acceptance endpoint. “Generated,” “delivered,” “accepted,” and “billed” need not describe the same event.

For this calculation, count each distinct final output once when the defined reviewer accepts it. Corrections and reruns of the same requested output remain linked to that output; they do not create extra denominator units. A genuinely new customer task needs its own identity.

Decide in advance what happens if a customer later rejects an output counted as accepted, and apply that rule consistently. The acceptance endpoint is the one you define and disclose.

Confirm who accepts the quality standard, links resources to jobs, records review effort and approves the cost basis. Finance must be able to reconcile the inclusions and allocation; a business sponsor decides what to change. Do not substitute proposed role labels for that agreement.

If acceptance is unknown, the accepted count is unknown—not the number of successful API responses. If no output is accepted, cost per accepted output is **undefined**, not zero. Report incurred cost and the zero count separately.

## Use a matched cohort and a complete numerator

Select a cohort of requested outputs with one workflow/version and a defined start window. Follow its work through the review cutoff. Attribute costs to that cohort even when attempts or review occur later. Record final dispositions and pending work; do not silently exclude difficult jobs.

For a closed cohort:

> **Cost per accepted workflow result = all in-scope costs incurred for the cohort ÷ distinct outputs accepted under the agreed standard.**

The numerator includes accepted and rejected work. Keep the following categories visible rather than blending them into “inference”:

| Cost category | Include in the defined scope |
| --- | --- |
| Model execution | Every actual attempt, retry, fallback model, intermediate generation, and charged failure—not just the selected answer |
| Tools and data preparation | Retrieval, extraction/OCR, search, external tools, ingestion/index updates, and task preparation attributable to the cohort |
| Infrastructure and services | Serving compute, idle/reserved capacity where relevant, storage, network, queues, monitoring, licenses, and shared services under explicit allocation rules |
| Human work | Preparation, evidence checking, correction, repeat review, escalation, and final acceptance, including time spent on rejected outputs |
| Operation and support | Attributable troubleshooting, failed-job recovery, and customer/workflow support |
| Lifecycle allocation | Agreed share of implementation, evaluation, integration, training, and maintenance work over a stated useful period or volume |

Avoid double-counting: a provider charge already in a managed-service invoice is not another separate compute cost; one correction session is not two blocks of labor because it carries two labels. Record credits and adjustments consistently with finance's chosen cost basis.

Show at least two views when useful: **operating cost per accepted output** and **fully loaded cost per accepted output**, with the latter including agreed lifecycle allocations. Neither is automatically marginal cost. State remaining exclusions and unknowns; a partial observed total is not a complete-cost claim.

The [FinOps Foundation's Unit Economics capability](https://www.finops.org/framework/capabilities/unit-economics/) distinguishes resource-efficiency metrics, such as cost per token, from business-unit metrics, such as cost per case resolved. That distinction is the point: keep token efficiency as a diagnostic measure, while making the business denominator explicit.

*Figure: All in-scope cohort costs, including rejected work and agreed lifecycle allocations, divided by distinct accepted outputs under the same quality standard. Unknown cost is not zero; with zero accepted outputs the ratio is undefined. — Operating and fully loaded views use the same denominator; neither automatically represents marginal cost.*

## Invented arithmetic: 100 requests do not mean 100 accepted results

*Figure: Invented cohort: 100 requests, 150 attempts, 80 accepted and 20 rejected briefs, none pending. The 480 CU fully loaded cost is 6.00 per accepted, 4.80 per requested or 3.20 per attempt. Operating cost is 380 CU, or 4.75 per accepted brief.*

Everything in this example is invented: workload, costs, acceptance counts, and allocations. **CU means fictional cost units, not dollars, current provider prices, or a client result.** No jobs or tests were performed.

Assume a closed cohort of **100 requested briefs** under one review standard. It requires **150 actual model attempts**: 100 initial attempts and 50 additional attempts across the cohort. At the cutoff, **80 distinct briefs are accepted and 20 are rejected**, with none pending. All preparation, attempts, review, and disposition work is included below.

| Fictional cohort cost | CU |
| --- | --- |
| All model attempts, including retries and rejected work | 75 |
| Tools and data preparation | 30 |
| Infrastructure, services, and license allocation | 20 |
| Human preparation, review, correction, and repeat review | 225 |
| Attributable operation/support | 30 |
| **Operating total** | **380** |
| Agreed lifecycle allocation for this cohort | 100 |
| **Fully loaded total in this example's scope** | **480** |

The category split is arbitrary, chosen to make the arithmetic easy to follow. It says nothing about the share of review, inference or any other category in a real workflow.

The useful calculations are:

- **Operating cost per accepted brief:** 380 ÷ 80 = **4.75 CU**.
- **Fully loaded cost per accepted brief:** 480 ÷ 80 = **6.00 CU**.
- **Cost per requested brief:** 480 ÷ 100 = **4.80 CU**. That is a different denominator, not a cheaper accepted result.
- **Cost per model attempt:** 480 ÷ 150 = **3.20 CU** across the same total cost scope. It is not a token rate or a cost per accepted brief.

The rejected briefs still incurred cost. Removing their work from the numerator would answer “what did successful jobs consume?” rather than “what did this workflow spend to produce accepted outputs?” Conversely, counting every retry as another result would inflate the denominator.

## Attribute the same costs and outcomes to tenants

*Figure: Invented tenant A has 270 CU and 60 accepted briefs; B has 210 CU and 20. Combine costs and counts: 480 divided by 80 equals 6 CU per accepted brief, not the unweighted rate average of 7.50. Why B differs is unknown. — Inspect case mix, attempts, review and allocation before explaining a tenant difference; the table establishes no cause.*

For each job, retain tenant ID, workflow/version, output ID, attempt/tool IDs, outcome and acceptance timestamp, cost references, review effort, and allocation method. Keep prompts, source documents, and personal data out of the cost worksheet; use controlled references.

Take model charges from the provider's usage report or invoice and associate each call with a job ID. A charge that cannot be tied to a job stays in an explicit **unassigned** line rather than disappearing or being guessed onto a tenant.

Directly attribute identifiable consumption. Allocate shared costs using a finance-approved driver that fits the resource—such as measured usage or reserved capacity—and disclose the method. Keep unassigned spend explicit. Equal splitting is not a factual tenant attribution merely because it reconciles a total.

Suppose the invented 480 CU is completely attributed as follows, including each tenant's unsuccessful work and allocations:

| Fictional tenant | Requested briefs | Accepted briefs | Fully loaded CU | CU per accepted brief |
| --- | --- | --- | --- | --- |
| A | 70 | 60 | 270 | 4.50 |
| B | 30 | 20 | 210 | 10.50 |
| **Combined** | **100** | **80** | **480** | **6.00** |

The combined rate is **480 ÷ 80**, not the unweighted average of 4.50 and 10.50, which is 7.50. The table does not explain why B differs. Case complexity, attempts, review, and allocations need inspection before concluding anything about customer value or behavior.

## Compare quality and cost without changing the question

*Figure: Invented comparison on the same 100 tasks and unchanged standard: 480 CU for 80 accepted briefs is 6 CU per accepted; 420 CU for 60 accepted is 7 CU per accepted despite lower total spending. — Invented arithmetic, not measured performance or a recommended choice. A better-looking ratio is not sufficient acceptance evidence.*

A proposed cheaper route might use a different model, fewer tool steps, caching, or a simpler non-AI process. Compare it on the same permitted tasks and acceptance standard. Include the route's own retries, rejected work, full review, support, and lifecycle allocation. Retain access, freshness, and domain-review conditions; “accepted” must not become easier just to improve the ratio.

As a second **invented arithmetic variant**, suppose a candidate spends 420 CU on the same 100 tasks but produces only 60 accepted briefs under that unchanged standard. Its fully loaded cost per accepted brief is **420 ÷ 60 = 7 CU**, above the first scenario's 6 CU despite lower total spending. This demonstrates denominator sensitivity, not measured model performance or a recommended choice. The larger unresolved/rejected workload and quality findings still matter.

Report distributions by relevant task type, tenant, route, attempts, and review burden. A pooled average can hide an expensive exception path. Compare like with like; a changed case mix or review standard needs a separate explanation.

For an open cohort, report cost to date, accepted/rejected/pending counts, and the cutoff. A provisional ratio is not the final unit cost: later review may add cost or accepted outputs. Do not pair this month's spend with an unrelated set of outputs accepted from earlier work.

## Calculate cost per accepted workflow result

Complete this short record for one cohort before choosing an optimization:

1. **Unit and quality:** workflow/output, reviewer standard/version, acceptance evidence, distinct-output rule.
2. **Boundary:** cohort start window, review cutoff, currency/cost basis, cost categories, lifecycle-allocation basis, exclusions and unknowns.
3. **Work and outcomes:** requested jobs, all attempts/tools, accepted/rejected/pending outputs, human effort, support.
4. **Attribution:** direct tenant/job costs, shared-cost allocation, unassigned amount, source reconciliation and owner.
5. **Calculation:** operating and fully loaded totals divided by the matched accepted count; denominator and uncertainty shown beside each number.
6. **Next finding:** the cost category, quality issue, or missing measurement that most changes the decision—and who will resolve it.

Use existing records or a small approved worksheet if they suffice. This calculation does not establish profitability, the right pricing plan, billable units or return on investment; those require additional commercial and financial evidence. If the missing measurement needs cross-system technical decisions, [discuss the question with the Grand Rounds curator](https://grandrounds.tech/talk-with-the-curator). No token-price reduction or margin improvement is promised.

If revenue and tenant records do not join to cost, start with Your AI Feature Is Popular, but You Cannot Explain Its Margin. For limits and metering, see AI Usage-Based Billing.

## Boundaries

Grand Rounds is an initiative of Dinarys, and a conversation may lead to cooperation with Dinarys.
