Measuring AI value: from cost tracking to budget decisions

A gap between tracking spend and measuring value

Tracking AI costs is progressing in organizations. The ability to say what those costs produce remains much rarer. This is the contrast that emerges from reading the week's FinOps publications, starting with a figure that stands out: 12%.

According to KPMG, only 12% of respondents systematically assess the value of AI against its costs. This isolated figure does not mean organizations measure nothing, but it suggests that value measurement remains the exception rather than the rule.

What the KPMG figures say

KPMG's survey — Global AI Pulse Q3 2026 — covers 2,131 senior executives in 20 countries. It indicates that 59% of respondents say they monitor spending during operations, and 61% review costs at the time a project is approved.

These practices are self-reported: they reflect what executives say they do, not necessarily what systems make it possible to verify. The gap between 59% or 61% on one side and 12% on the other nevertheless deserves attention. It suggests that budget control often stops at the point when the question becomes the return obtained.

The practical question this raises is simple: can the data already present in a cost dashboard be used to decide which use case deserves more budget? As things stand, nothing guarantees that this data is structured to answer that question.

Agents in production: partial visibility

The issue becomes more complicated with agents. Open Future Forum reports that, among 60 respondents with agents in production who also answered the cost question, only 30 have complete visibility into real-time costs.

The sample is small and does not represent all companies. It nevertheless invites checking concrete points internally: can the owner of each agent, its budget and its actual output be identified? Without these three elements, attributing a cost to an outcome becomes difficult, and budget arbitration relies on estimates.

Model routing: a promise to verify

On optimization, Weave reports that routed traffic was billed about 44% less than the models initially requested would have been. Its analysis covers 1.62 million requests across 52 installations, and 82% of requests were served by a model different from the one requested at the outset.

This result makes routing worth testing. It should not, however, be entered as 44% in a business case as is: it is a vendor conclusion, from its own platform. Several costs must be included before concluding: output quality, the cost of routing itself, that of evaluation and that of any rework. A billing gain becomes an economic gain only if quality remains at the expected level.

A concrete avenue: cost per accepted business outcome

To make the budget discussion more useful, I would start from two use cases and a single measure: cost per accepted business outcome. The example often cited is that of the resolved support ticket, including failed attempts and human review.

This approach forces us to define what 'resolved' really means before comparing costs. It shifts the conversation from the volume of tokens consumed to what the organization recognizes as a useful outcome. This is a suggested method, not an observed result: its relevance depends on the ability to define the outcome stably and attribute it unambiguously.

Practical questions before the next budget arbitration

What business outcome can be attached today to the AI bill? The answer determines what follows. If no outcome is identifiable, the budget discussion remains a discussion of costs.

Other questions deserve to be asked: who owns each agent, what budget is associated with it, what actual output is attributed to it? How should failed attempts and human review be handled in the calculation? And how can two use cases that do not produce the same type of outcome be compared?

These questions do not provide a ready-made answer, but they shift attention toward what is most often missing: an explicit link between the expense and what it makes it possible to obtain.