Finance departments have no shortage of AI experiments. Somewhere in most mid-sized and large companies, an analyst has built a forecasting model in a notebook, a controller has trialled a tool that drafts variance commentary, or a treasury team has tested anomaly detection on bank transactions. The demos are usually impressive. Far fewer of these experiments ever make it into the monthly board pack.
The gap between a promising pilot and a trusted production tool is rarely technical. Models often perform well in testing. What stops them is a set of organisational questions nobody answered at the start: who owns the output, how it will be checked, what happens when it is wrong, and whether the people who sign off on reported numbers are willing to put their names next to something they cannot fully explain.
The Pilot Was Designed to Impress, Not to Survive
Many finance AI projects begin as proofs of concept chosen for their wow factor. A model that predicts next quarter’s revenue within a few percent makes a memorable presentation. But a pilot optimised for a single demo tends to rely on a cleaned, one-off dataset, a data scientist who knows its quirks, and a narrow time window where conditions happened to be stable.
Production is less forgiving. Source systems change, chart of accounts mappings drift, and an acquisition adds an entity with different reporting conventions. If the pilot was never built to handle those realities, it degrades quietly. The first time it produces a forecast that is clearly off, confidence collapses, and the project joins the long list of initiatives that were “interesting but not ready.”
Explainability Is the Real Approval Gate
Boards and audit committees do not approve algorithms. They approve numbers, and they expect whoever presents those numbers to explain how they were derived. A CFO can defend a forecast built on documented assumptions about pricing, volume and headcount. Defending a figure produced by a model whose reasoning is opaque is a much harder conversation.
This is why the tools that move from pilot to routine use are usually the ones that make their logic visible. Practical FP&A AI tools tend to succeed when they sit inside the planning workflow, show which drivers moved a forecast, and leave the final judgement to the analyst rather than replacing it. A tool that says “revenue is down because renewals in one region slipped by two weeks” earns trust faster than one that simply outputs a lower number, however accurate.
Standards and Skills Are Lagging Behind the Tools
Finance is not alone in hesitating. Across the investment industry, employers surveyed by CFA Institute overwhelmingly said the sector needs common standards and ethical guidelines for AI, with 85 percent agreeing, and the association reported that the absence of such standards is itself slowing adoption, alongside workforce anxiety about the technology and gaps in preparation, according to its 2024 survey. Corporate finance teams face a similar vacuum. Without an internal policy on acceptable uses, review requirements and data handling, every pilot has to renegotiate its legitimacy from scratch.
Skills are the other bottleneck. A controller who understands revenue recognition deeply may have no framework for judging whether a model’s error rate is acceptable. A data scientist who can tune that model may not know which reconciliations auditors will ask about. Pilots stall in the space between those two people. The fix is usually not more training for one side but a shared working rhythm: joint reviews of model output during the close, a common glossary for terms like accuracy and materiality, and a clear rule about who has the final word when the model and the analyst disagree.
What Regulators Expect From Serious AI Use
Supervisors have started spelling out what responsible deployment looks like, and their expectations translate well to corporate finance. A Federal Reserve governor speaking about the central bank’s own adoption described the need for clear guardrails on where AI is used, strong information security, rigorous model validation and human accountability for decisions, warning that these systems can amplify errors as quickly as they amplify efficiency, as the Board’s published text of his remarks makes clear. Replace “central bank” with “finance function” and the list reads like a checklist for getting a pilot through an audit committee.
The underlying point is that governance is not a tax on innovation. It is the mechanism that lets a board say yes. A pilot arriving with a named owner, a validation log, documented limitations and a fallback process is far easier to approve than one arriving with only an accuracy chart.
Designing Pilots That Graduate
A few design choices dramatically improve the odds of a pilot reaching production. Start with a use case where the model supports a human decision rather than replacing it, such as flagging unusual journal entries for review or drafting first-pass variance explanations. Define success before the pilot begins, including the error tolerance the business can accept and how results will be compared against the current process. Run the model in parallel with the existing method for at least a few close cycles, so the team builds evidence rather than relying on a single impressive month.
Ownership matters as much as design. Every pilot needs a finance owner, not just a technical one, who will answer for the output in the room where numbers are challenged. Budget for the unglamorous work too: data pipelines, access controls, documentation and training. These rarely appear in the pilot proposal and almost always determine whether it survives.
Earning a Seat at the Board Table
AI in finance does not need to be revolutionary to be valuable. The projects that reach the boardroom tend to be modest in scope, transparent in method and owned by people the board already trusts. They shorten the close by a day, catch a coding error before it reaches the ledger, or give the CFO a clearer view of which assumptions are driving the forecast.
The companies that scale AI fastest will not be the ones with the most pilots. They will be the ones whose pilots were built from the first week to answer the questions a sceptical audit committee will eventually ask.
