AI costs: who accepts the outcome and who pays for the attempts?

A timely question, not a new obligation

On 10 September 2026, Tammy Burnitt and Luis Lazo published a FinOps Foundation perspective on model selection asking who sets minimum quality and cost per outcome. This is guidance attributed to its authors, not a market measurement or legislation.

Dean Oliver's agent analysis, dated 18 August, proposes policy-bounded authority and spending and step limits. The framework's Allocation page organises attribution of direct and shared costs. Reviewed on 5 October, these three pieces frame Velqory's question: can an SME explain the bill for a completed task?

A workload needs an owner

Velqory recommendation: give the operation a stable identifier, for example support-reply preparation, and name the person accountable for its usefulness. That owner defines an accepted outcome; the technical team owns configuration; finance confirms allocation.

Link each execution to task identifier, environment, version and cost centre. Do not record sensitive content simply to explain expenditure. If attribution is missing, keep the amount in an explicit unresolved-cost category with an owner and deadline. Do not silently distribute it among whichever projects are easiest to identify.

Count completions, not activity

We propose this metric: complete cost allocated in the period divided by the number of unique outcomes accepted in that period. The numerator includes initial and repeated consumption, tools, infrastructure, human review and the agreed share of common costs. Disclose exclusions and whether costs are invoiced or estimated.

The denominator must correspond to a work unit defined before measurement. A task may generate several calls and versions but only one accepted completion. Report failures and refusals; their costs remain in the numerator even without an accepted outcome. With zero acceptances, division is undefined: report cost and causes, not zero euros per outcome.

Fictional example: a bill that includes rework

An entirely fictional arithmetic example created for this article: every euro amount and volume is invented, not vendor pricing, client data or observed results. Suppose there are 100 tasks in the same period and 140 calls: 100 initial calls and 40 retries. At close, 80 tasks have been accepted once each; 20 ended without acceptance.

Imagine €90 for initial consumption, €30 for retries, €20 for tools and infrastructure, €100 for human review and €40 for shared costs. Total: €280. Cost per accepted outcome: €280 ÷ 80 = €3.50. Retry and failure costs are already included in these components; do not add a duplicate category.

Dividing the same total by 100 tasks would give €2.80, but answer a different question: cost per started task. It would not be cost per accepted completion. This example does not calculate savings, productivity or returns; it only demonstrates how the denominator changes the interpretation.

A closing matrix that makes the bill reconstructible

The following matrix is an original Velqory proposal. A simple register should allow the calculation to be reproduced and identify who resolves discrepancies. It does not require buying a new platform.

Compare alternatives without hiding human work

Before switching models, keep the task, corpus and criteria constant. Compare complete cost, acceptance and deadlines, including the effort to correct outputs. Separate simple operations from cases requiring specialist review; a global average can hide the group that stops working.

Document the allocation rule before looking at the winner. If a common service remains entirely within a central budget, disclose that treatment and the portion excluded from the comparison. Distinguish recurring operating expenditure from initial setup. An apparently cheaper alternative has not been demonstrated while the comparison bases differ.

What to decide at the next meeting

Choose one workload and ask for its cost per accepted outcome, failure count and named decision-maker. For agents investigating costs, Velqory proposes starting with read access and evidence preparation. Changing resources requires separately delegated authority; access to an account does not define that mandate.

Set ceilings for attempts, expenditure and duration, with referral to a person when reached. Reopen the decision when criteria, contract or configuration changes. October's objective is an explainable bill and an accountable decision, not an announcement of efficiency without evidence.