In most finance organizations right now, two documents sit a few clicks apart. One is the business case that justified the AI investment. The other is the bill. They don’t agree, and the gap is widening in a way nobody modeled.
According to Gartner, AI cost estimates in many organizations have been off by 500 to 1,000%, and more than half of AI pilots are expected to be abandoned by 2028 because of unexpected cost overruns. A miss of that size points past procurement errors to the cost structure itself.
Why the cost curve breaks the model
Experimentation costs are bounded. Rollout costs are bounded. Ongoing usage — tokens, compute, data pipelines, retraining — grows with adoption, and success makes it grow faster.
That inverts the relationship between usage and unit cost that finance is used to underwriting. A licensed application gets cheaper per user as more people use it. An AI deployment that reasons through the same analytical question on every request gets more expensive per answer as adoption spreads.
Gartner’s warning to CFOs applying the old technology playbook is direct: the 2% to 3% annual labor cost increases you planned for get replaced by 8% to 10% increases in cloud and vendor costs.
Most of that spend is defensible on its own terms. Someone asked a reasonable question and got a fluent answer. Nothing failed. The cost simply accumulated without an owner.
Why the return is so hard to see
Now put the second document next to the first. Gartner finds that 84% of CFOs struggle to measure AI ROI, and fewer than one in six finance leaders have a method that works repeatably. BCG’s 2025 survey of more than 280 experienced finance executives puts median ROI in finance at 10%, with roughly a third reporting limited or no gains. Deloitte’s Finance Trends 2026 research sharpens it: 63% of surveyed finance leaders have fully deployed AI, while 21% report clear, measurable return.
The usual read is that measurement practice needs to catch up with deployment. There’s also a structural reason, and it’s the same reason the cost curve misbehaves.
You can only measure what repeats. When an AI system reconstructs the analysis every time it’s asked, there’s no stable unit of work to baseline against. The output shifts with the phrasing of the question. Two people asking the same thing in the same week get answers that differ enough to matter, and nobody can say which one was right or what changed between them. Comparing this quarter to last quarter under those conditions produces a story, not a number.
Bain’s analysis of where AI is already working in finance arrives at the same constraint from another direction: the limitation is the data that provides context rather than the price of the compute. Bain’s guidance on evaluating results is equally pointed. Measure impact on business outcomes, not on prompts used or hours saved.
What changes when the logic is governed
Business logic is defined by finance, not inferred by AI. When the reconciliation rule, the allocation methodology, the materiality threshold, and the exception logic live in governed workflows, each AI request calls logic that already exists and has already been validated.
Two things follow. Complex analytics execute inside the workflow rather than in the model, which replaces token spend that scales with organizational curiosity with predictable, governed execution. And the answers become comparable, because the same approved logic produces the same governed answer on current data however the question was phrased.
Comparability is what a baseline requires, and it arrives with the audit trail attached: which data was used, which rules were applied, who approved the result. A productivity claim built that way is one a controller will sign.
That is the practical content of the standard worth holding AI to in finance. Visible, understandable, repeatable, auditable — what we call VURA. In a measurement conversation, repeatable carries the weight. It separates a pilot that produced a good week from a workflow that produced a number you can defend in a board meeting.
Scaling is a portfolio decision
Portfolio management works only when the units can be compared to each other. Governed workflows compare cleanly. Ad hoc AI usage produces a different unit of work every cycle, which is why the organizations with the most AI activity are often the ones least able to say what any of it returned.
So the question for the next AI review isn’t a count of live use cases or a total of dollars committed. Ask what the second workflow costs to run once the first one is working, and whether the answer it produced this month can be compared to the answer it produced last month. Two clean answers mean you have something to scale. Two murky ones mean more deployment will deepen the measurement problem, on a cost curve that rewards nobody.
Finance teams that get this right end up somewhere genuinely different: a function where the close and the forecast run continuously, because the logic underneath them is stable enough to trust without being re-derived each cycle. That destination is reachable. Spending faster is what delays it.