AI in FP&A: Where the Real Leverage Hides

The marketing pitch for AI in finance is a CFO asking natural-language questions of the general ledger and receiving instant strategic advice. The reality, for the next two to three years, is duller and more valuable: a finance team that spends forty percent less time on reconciliation, attribution, and scenario refresh, and reinvests that time in judgement-heavy work that the AI is not yet competent to do. The gains are real, but they live in the middle of the stack, not at the executive surface.

The pattern we see in clients who deploy AI in FP&A well is consistent: start with the dullest, most repetitive workflow in the team; build a tight, verifiable automation around it; measure the time saved over one quarter; repeat. The pattern in clients who deploy poorly is also consistent: start with the most visible workflow (the board pack), build something that looks impressive in a demo, and discover that the trust required to actually rely on it in production is six to twelve months further out than the demo suggested.

Bank reconciliation, intercompany reconciliation, and the matching of subledger to general ledger are the workflows where AI delivers the fastest, cleanest payback in a finance team. The work is repetitive, the inputs are structured, the right answer is verifiable, and the time cost is large — a typical senior analyst spends six to ten hours a month on reconciliations that an LLM-driven workflow can close in under thirty minutes with human review.

The architecture is straightforward: pull the two ledgers being reconciled into a structured form, prompt the model to identify probable matches and unmatched items with reasoning, and present the unmatched items to a human for adjudication. The model is not making decisions; it is doing the visual scan that takes a human three hours and producing a structured handoff. The verification step — the human review of unmatched items — is what makes the workflow safe to deploy in production.

The variance attribution workflow is where AI delivers the most strategic value, because the work is currently either skipped or done badly. The standard FP&A variance report attributes movements at the consolidated level — 'revenue down £400k versus budget' — because attribution at the contract, SKU, or customer level requires either a data warehouse the company has not built or a week of analyst time that no one has.

An LLM paired with the right structured prompts can produce contract-level or SKU-level attribution from the operational data the company already has, in minutes. The output is not perfect — the model will sometimes attribute a movement to the wrong driver — but the output is reviewable, and the review takes far less time than the original attribution would have taken to produce manually. The shift is from 'no attribution' to 'attribution that the analyst can verify and refine,' which is a transformation in the quality of variance commentary the board receives.

The other workflow operators ask us to 'put AI on' is scenario refresh: re-running the financial model with one or two assumptions changed. This is almost always the wrong tool for the job. Scenario refresh is a templating problem — the model already knows how to recalculate when an input changes — and the right solution is a well-structured spreadsheet with named scenarios, not an LLM. Deploying AI for this workflow tends to introduce variability where the business needs determinism.

The right test for whether a workflow is a candidate for AI is whether the human work it replaces is 'recognise patterns in unstructured input and produce structured output.' Reconciliation passes. Attribution passes. Scenario recalculation does not. Knowing the difference is the discipline that separates teams who get value from AI deployment from teams who spend a quarter building tooling and abandon it.

The risk profile of AI in finance is not what the popular discourse suggests. The model is not going to invent a number that destroys the close. The actual risk is more subtle: the model will produce a confidently-stated answer based on an assumption the analyst did not realise was embedded in the prompt, and the answer will be wrong in a way that is hard to detect. 'Customer acquisition cost was £420 in Q3' may be correctly calculated against the wrong cost base, and the wrongness will not be visible until someone notices that the answer does not reconcile with the cash flow.

The mitigation is to build verification into every workflow before it goes live. Every AI-produced number should be traceable back to a specific source, and every prompt should be version-controlled and reviewed. The teams that treat AI as a junior analyst — output reviewed, assumptions challenged, work product owned by a named human — get the productivity gains without the silent errors. The teams that treat AI as an oracle do not.

Our finance automation engagements typically start with a two-week audit: which workflows in the team are repetitive, structured, and verifiable; which are judgement-heavy; which are a mix. The repetitive-and-verifiable workflows are sequenced for automation, the judgement-heavy workflows are left to senior humans, and the mixed workflows are decomposed into the parts that can be automated and the parts that cannot.

Over a single quarter, a typical eight-person finance team can move forty percent of its reconciliation and attribution work to AI-assisted workflows, free roughly fifteen hours per analyst per month, and redirect that time to the planning, analysis, and business-partnering work that the team has been underweight on for years. The headcount does not fall. The output of the team does. This is the actual transformation, and it is available to any finance team with the discipline to start at the dull middle rather than the visible top.