AI in the Finance Workflow

Most of what is written about AI in finance is written by people who have never closed a book. The demos are impressive, the promises are large, and the workflows are imaginary.

This guide takes the opposite approach. I spent more than fifteen years inside finance organizations — running FP&A cycles, owning trial balances across entities, reconciling ERP data against consolidation layers, drafting the variance commentary that lands on a CFO's desk. I did, by hand and in Excel, almost everything AI now proposes to automate.

Knowing the manual reality of a workflow tells you precisely where a model adds value, where it merely adds noise — and where letting it act unsupervised would be professionally negligent. One color code runs through every graphic that follows: teal for what the model should do, amber for what a human must own. If you only skim the pictures, you'll still leave with the argument.

Every finance professional knows the rhythm. Close ends, the numbers land, and someone has to explain them — actuals versus budget, versus forecast, versus prior year, by entity and product line. The analysis is half the work; writing it in language a business leader will read is the other half. In most teams I've led, commentary consumes two to four days of every cycle, much of it mechanical.

The pattern that works

The two constraints in bold are the entire game: don't guess drivers, and flag what needs confirmation. An unconstrained model invents plausible causes — "driven by higher freight costs" — that nobody verified. Plausible and wrong is the most dangerous combination in financial reporting.

This is where FP&A either earns its seat or stays a reporting function. Commentary that only narrates the numbers is a rear-view mirror. The value is in what comes next: if this business unit is dropping, what compensates it — and how much do we need to recover, per month, to close the gap?

The mitigation logic follows a structure every experienced FP&A professional carries in their head: quantify the gap, identify the offset candidates, compute the required run-rate, test feasibility. The first three are mechanical — which means a model can draft them. The fourth is judgment, and it stays human.

What the model drafts — and what it never decides

Notice what the constraint does: the model produces the options and the math — the part that used to consume an analyst's afternoon — and explicitly refuses the feasibility call. Whether BU B's growth can really be accelerated, whether the freight increase is structural or spot, whether a price action survives contact with customers: that judgment is why FP&A has a seat at the table. The model prepares the decision. It does not make it.

Scenario work in most organizations is spreadsheet archaeology. Someone builds a driver model, someone inherits it, and by year three nobody fully trusts the links. When leadership asks what happens to margin if input costs rise eight percent and volume softens, the true answer arrives days later.

I use LLMs in three distinct roles here — and keeping them separate is the discipline:

For years I built monthly business review packages — the layer where consolidated data becomes a story leadership can act on in a ten-minute read. High-value work with a highly repetitive core: every month, the same structure, new data, a fresh narrative. The narrative is where the hours go — which makes it the single most automatable text in the finance calendar.

The honest result is qualitative as much as quantitative: when producing the narrative stops consuming the team's energy, the content improves — because people finally have time to think about what the numbers mean before writing about them. Automation here doesn't replace the analysis. It funds it.

Beyond FP&A, four operational areas dominate my map of where AI creates value fastest — ranked by conviction earned the hard way: I have personally been the manual version of each.

Mapping validation deserves a special word because it is unglamorous and extremely valuable: anyone who has migrated a consolidation platform knows account mapping is where migrations quietly go wrong. A single bad mapping surfaces weeks later as an unexplainable variance. AI-assisted match suggestion and anomaly flagging — "this account maps differently from 94% of similarly named accounts" — would have materially reduced risk in every migration I've been part of.

This is the section most AI-in-finance content omits — and the one that determines whether any of the above survives contact with an audit. I come from a controls background: external audit, then designing SOX controls inside operating companies. That makes me conservative in specific places, and the conservatism is a feature.