Why Your AI Business Case Fails Finance Review

CFOs reject AI investment proposals not because the technology is unproven, but because the submitted case leaves out costs and names no one responsible for results.

Businesses have spent between $35 billion and $40 billion on generative AI. MIT NANDA's 2025 report found that 95% of those pilots failed to produce measurable business impact. The money went in. The results did not come out. And the most common explanation — bad tools, wrong vendor, weak data — misses where the problem starts.

The case fails before the project does

CFOs reject AI proposals for a specific reason: the document describes a future benefit with no credible path to realising it. No measured baseline. No named person responsible for the result after go-live. No line item for what it costs to change how people work or keep the model running after launch.

These are not oversights. They are signals. A finance team reading a case without those elements draws one conclusion: the team has not thought past the demo.

[PERSON_NAME]'s 2025 analysis of failed finance reviews identifies the same three gaps across rejected cases. Change cost is absent. Ongoing model upkeep is treated as someone else's budget. Opportunity cost — what the team gives up to run this project — does not appear at all. A CFO reads each missing line as a question the submitting team chose not to answer.

What the opposing view gets right

The Fortune coverage of MIT NANDA's report points to a different explanation: specialized vendors succeed roughly 67% of the time, while internal builds succeed about one-third as often. That gap is real. It reflects execution decisions — tool choice, workflow fit, internal capability — not document quality. Practitioners who have watched these projects collapse are not wrong to focus there.

The distinction matters, though. The vendor gap describes what happens after a project is funded and running. It does not explain why the case was rejected before funding was granted. The two failure modes are sequential. A well-structured case does not fix a mismatched tool. A better tool does not get funded if the case never clears finance review.

The four questions a CFO always asks

Every finance review of an AI proposal comes down to four questions. What is the current baseline, measured and dated? What does full ownership cost, including change management, retraining, and model monitoring? Who is accountable for the result after launch, by name and role? What does success look like at 90 days, not just at year three?

Most submitted cases answer none of these directly. They project a benefit figure, attach a technology cost, and stop. The NIST AI Risk Management Framework, published in 2023, treats ongoing monitoring and governance as non-negotiable from day one — not as post-launch additions. Finance teams increasingly read proposals through that lens.

Bharadwaj et al.'s 2013 research on digital business value showed this pattern predates generative AI by more than a decade. Firms have consistently failed to turn technology spend into business gains when the case treats the technology as separate from the work it is supposed to change. The tool is not the investment. The changed workflow is.

Build the case before the project, not after it stalls

The timing problem is structural. Most teams build the business case after a pilot stalls, which means they are asking finance to approve a sunk cost. The baseline is gone. The change costs have already been absorbed informally. No one owns the result because the project has been running without a named owner for months.

Building the case before the project starts forces the decisions that make the case credible. You have to name a baseline before you can claim an improvement. You have to scope the change work before you can cost it. You have to name an owner before the project begins, or the role never gets filled.

I have seen teams spend six weeks building a benefits model for a project that had already been running for four months with no financial model attached. The CFO declined it in one meeting. Not because the number was wrong, but because there was no way to verify it against anything that had been tracked.

The MIT NANDA data suggests execution model matters independently of case quality. Pick the wrong tool and the project fails regardless. But a project built on a vendor relationship with a 67% success rate still needs a finance-literate case to get funded. The case is not the whole answer. It is the prerequisite.