Why Finance Doesn't Trust Your AI Numbers

When organizations begin deploying AI at scale, the financial conversation changes as fundamentally as the technology does. Most leadership teams anticipate the technical challenges. Very few anticipate that the binding constraint on their AI program will be whether the finance organization believes the numbers it is being shown.

The pattern is consistent. A technology leader presents a business case for a production AI deployment. The analysis is thorough and the rationale is sound. Partway through, a finance executive asks whether the projected costs are on a cash or accrual basis, or how the sensitivity analysis was constructed, or how the figures reconcile against the current period's budget. The answer is technically accurate but addresses a different question than the one asked.

Nothing visibly goes wrong. But the finance executive has just learned that the person presenting is not fluent in the framework the decision will be evaluated against, and from that point every figure carries a higher burden of proof.

This is the problem FinOptik was built to solve, and the premise is worth stating plainly. We think this gap is best closed by people who came up through finance and crossed into technology, rather than the reverse. Twenty-seven years at that intersection, starting in FP&A, is why the CFO conversation moves differently.

Here is what we see most often, and what to do about it.

1: Finance Rigidity Is Not Skepticism About the Investment

Technology leaders frequently read finance scrutiny as opposition to the project. It almost never is.

Consider what a finance organization is managing. At the start of each period they collect results from dozens of teams, reconcile them, eliminate overlaps, confirm whether figures are cash or accrual, and produce a consolidated view for senior leaders who will make consequential decisions from it. All under a deadline that is externally imposed and non-negotiable.

Under those conditions, a number arriving in an expected format using an expected methodology requires almost no scrutiny. A number arriving otherwise requires a great deal of it. This is not a personality trait. It is a survival response, and practitioners who lack it do not last in the role.

So the axiom works like this: finance is not evaluating whether your AI investment is a good idea. Finance is evaluating whether it can defend your numbers to the people who will ask about them. Analysis that answers only the first question keeps getting sent back.

2: The Visible AI Costs Are the Smaller Number

We were recently in a discussion with an organization that had built what they considered a complete AI cost model. It covered token consumption, model hosting, and GPU compute. We asked what they had assumed for data preparation, embedding generation, fine-tuning, and retraining. Those were handled by existing internal teams, and so had not been included.

This is extraordinarily common. Costs that arrive as an invoice are easy to see and easy to model. Costs that take the form of redirected internal capacity generate no invoice, so they get treated as free. They are not free. That capacity was doing something else before, and what it is no longer doing has a cost even though no line item records it.

When those costs surface later, and they always surface later, finance does not conclude the estimate was conservative. It concludes the methodology was unreliable, and applies that conclusion to every subsequent projection from the same source.

Modeling only the invoice is modeling the visible portion of something whose larger portion is not visible. That is a reasonable description of how organizations end up investing millions on faith.

3: AI Is Harder to Model Than Cloud Was

Cloud eventually became tractable for finance, and it is worth remembering why. Resource-based billing has analogues in other operating expense. Commitment instruments resemble other purchase obligations. Consumption is variable in ways that model reasonably well after a few periods of history.

AI is harder on every one of those dimensions. Inference volumes shift with adoption patterns that are themselves difficult to forecast, and a deployment stable for two quarters can triple after a single product decision made elsewhere. Cost per inference and cost per token map onto nothing finance already tracks, and the terminology has not standardized across providers. And the benefits are frequently indirect, which means attributing them requires a baseline most organizations never established.

How FinOptik Closes the Gap

We are engaged in three ways, depending on where the organization sits.

Before the investment decision. We build the business case: full cost model including the categories that generate no invoice, risk-adjusted NPV and IRR, sensitivity analysis constructed the way a finance reviewer would construct it, and a board-ready presentation. Fixed scope, four to six weeks.

After the commitment. We build the financial operating model with your teams: tagging and allocation strategy, unit economics framework, cost per inference and cost per token definitions, chargeback and showback design, governance, and a Center of Excellence. Roughly twenty-six weeks for a working foundation.

On an ongoing basis. We run the continuous function on retainer: monthly unit economics tracking, anomaly review, model selection trade-off analysis, commitment optimization, and board-ready reporting.

Engagements have covered financial services, insurance, transportation, manufacturing, technology, and consumer goods, at organizations spending five million dollars or more annually on cloud. We have advised clients managing over six hundred million dollars in combined cloud spend, and we do not resell cloud, take rebates, or sell tooling, so nothing we earn changes when a customer's spend changes.

Fixed scope or monthly retainer, available through AWS Marketplace or direct, starting with a thirty-minute call at sales@finoptik.io. More detail on all three is on our FinOps for AI page. The hiring side of this problem is covered separately in The AI FinOps Leader: What the Job Actually Demands.

Do These Challenges Sound Familiar? Here Is What to Do Next

1. Build the model on a cash basis and state the basis explicitly. Where sunk costs or existing internal capacity are involved, a TCO framing produces a number finance cannot use.

2. Include the categories that generate no invoice. Data preparation, embedding generation, prompt engineering, fine-tuning, retraining, integration, orchestration. Estimate them, caveat them, put them in. A disclosed estimate is defensible. A discovered omission is not.

3. Establish the baseline before you deploy. If the case rests on productivity improvement, measure the current state first. This gets skipped routinely, which is precisely why so many AI benefit claims cannot survive scrutiny.

4. Invest modestly in mutual education. A working understanding of accrual accounting on one side and of how inference costs behave on the other is generally sufficient, and it changes every subsequent conversation.

5. Give the translation function an owner. Internal hire, existing FinOps practitioner with an expanded remit, or outside help. In our experience it does not happen otherwise.

One client reduced its workload justification cycle from roughly three months to three weeks, not by changing the underlying analysis but by establishing a methodology both organizations recognized as sound. Approvals that had required several review cycles began clearing in one.

That compression is what this capability actually buys. The savings follow, but they are rarely the main event.

Rich Hoyer is CEO and co-founder of FinOptik. He served on the FinOps Foundation Governing Board and Technical Advisory Council and co-authored the Google Cloud FinOps global operating model. FinOptik builds AI and cloud financial management practices inside large enterprises. Inquiries: sales@finoptik.io

Published August 2026 | Rich Hoyer, CEO, FinOptik