What AI actually changes in FP&A: a practitioner's note

Most of what is written about AI in finance is written by people selling something. This note is from the other side: I run AI inside a live P&L: Claude, Claude Code, Claude in Excel and ChatGPT, in daily use across analysis, forecasting, modelling and the tooling that carries them, and have done since the tools became capable enough to trust with real work.

In my experience, four things genuinely change, and one thing must not. To see why, start with what a forecast actually is.

A forecast is a driver tree, not a spreadsheet

A revenue line is not a forecast; it is the answer at the bottom of one. A credible model works in demand units first (customers, patients, orders) and builds the number from the drivers that cause it. In practice that tree has five branches, and each one has to be modelled on its own terms before AI can do anything useful with any of them.

Seasonality, decomposed rather than smoothed. Selling into the travel market, my demand concentrates around booking windows and holiday peaks; a seasonality curve fitted before 2020 was quietly wrong for years afterwards, and averages hide exactly that kind of drift. The pattern has to be separated from trend and re-fitted against current actuals, or every downstream number inherits the error.

Lifecycle stage, per product, not per portfolio. Every product sits somewhere on its curve: launch, growth, maturity, decline. In pharma the whole commercial model turns on this: adoption curves on the way up, patent cliffs on the way down, and the patient-based forecasting I built for valuations at GSK was, at its core, lifecycle modelling done properly. Consumer products trace the same shape, faster. The old frameworks still earn their keep here: BCG's growth-share logic and the product lifecycle are not slide decoration, they are the checklist for where growth assumptions deserve suspicion. A mature product forecast to grow like a launch product is how a plan fails quietly, a year before anyone notices.

Marketing as a demand driver, not a cost line. Acquisition spend has its own economics: cost per acquired customer, value by channel, diminishing returns by audience. I first built this view at O2 in the early 2000s, running customer lifetime-value analysis by tariff, device and sales channel over raw usage data, and the output redirected real marketing spend toward the customers worth acquiring. Two decades on, the same discipline runs my own paid acquisition. If marketing sits in the model as a single expense line, the forecast cannot answer the only question the board actually has about it: the marginal return on the next pound of spend.

The supply side, moving on its own clocks. Landed cost is not one number; it is unit cost plus freight plus duty, all of it restated through the exchange rate, and each of those reprices on its own cycle. Freight rates swing, currencies move, and a renegotiated factory price reaches the margin line months later, when the inventory it bought actually sells through. I carry these in my own model because the cash consequences are mine, and any business that imports what it sells should demand the same of its forecast.

Price and mix across all of it, because growth that arrives through discounting and growth that arrives through volume are different businesses wearing the same revenue line.

Drivers first, tools second. AI does not change that order. It changes everything about what happens next.

1. The assumption hunt gets fast

Every forecast is a stack of assumptions, and the dangerous ones are the assumptions you have stopped seeing. Rebuilding a demand forecast last year, I put the model through Claude in Excel and asked it to surface every embedded assumption it could find. It returned several I had carried, unexamined, for two planning cycles, including a seasonality weighting that no longer matched the post-COVID demand pattern I could see in the actuals.

The value was not that the machine was right. It was that it made me re-justify what I had stopped questioning, in minutes, not days. AI compresses the time between "the number looks fine" and "here is why the number is wrong," and for an FP&A team that compression is worth more than any reporting refresh.

2. Sensitivity analysis stops being a luxury

A forecast with forty assumptions usually turns on three. The senior work has always been knowing which three, and until recently that knowledge was expensive. Presenting licensing scenarios to internal and legal-entity boards at GSK taught me what that depth of work used to cost: analyst weeks, reserved for the decisions that justified them. Most planning cycles never got it.

That constraint has gone. Ranking the driver tree by what actually moves the year, testing what a ten per cent freight increase does to gross margin by quarter, or what a two-week supplier delay does to peak-season availability, is now a conversation with the model rather than a project. Testing the volume response an assumed elasticity implies, a question that once waited for the next planning round, is answered while the meeting that raised it is still running. The discipline this rewards is old, not new: modelling work I led took forecast error against actuals from roughly fifteen per cent to four, and that was earned with patient-based models and hard graft long before AI. The tools simply make that standard affordable on every cycle, for every business size.

3. The model becomes a living one

This is the largest opportunity, and the one I see least discussed honestly. The pieces above (demand with its seasonality and lifecycle staging, marketing economics, supply costs, currency) usually live in separate spreadsheets owned by separate people, reconciled quarterly, stale by the time they meet. Integrated planning platforms have promised the joined-up version for a decade; what AI changes is the cost and the speed of getting there. One driver model, connected to actuals, answering what-if questions on demand. Change the freight assumption and watch landed margin, contribution and the cash position move together. Reprice a product and watch the volume response flow through to the same three lines. Ask it which driver moved against plan this month and get the bridge, with its workings, on the spot.

What has actually changed is the cost of building this. A model of that kind used to be a project measured in quarters, with a change-request queue behind it. I now build the model logic myself: Claude Code writes and rewrites the tooling as fast as the design conversation moves, and Claude in Excel interrogates the model where the finance team actually lives. The honest caveat is the data: connecting clean, reconciled actuals from the source systems is still the real work, and how long it takes depends on the estate. The logic, though, now moves in sessions rather than months. I run my own company this way. On a mandate, it is the difference between a monthly pack that describes last month and a model the board can question live, in the meeting, and get answers that show their assumptions.

4. First-pass analysis stops being a bottleneck

The organisational change follows from the rest. First-pass work (variance narratives, scenario drafts, bridge explanations) used to queue behind whoever had the model open. It now happens conversationally, at the speed of the question. The finance leader's job shifts accordingly: less time producing the first pass, more time interrogating it. Judgement becomes the scarce resource, which is where it always should have been.

What must not change: the control environment

Everything above is speed, and none of it is safe without control. This is where finance leaders should feel at home, because governing a powerful, occasionally wrong instrument is what the profession already does. Five controls run through my own practice, and I would install the same five on any mandate.

No naked numbers. Every AI-assisted output must state the assumptions it made and the data it relied on. If it cannot show its workings, it does not enter the pack. That is not an AI nicety. An auditor asks the same of a journal entry.

Cross-examination before presentation. Material analysis is pressure-tested by a second model, briefed to attack the first one's reasoning. On decisions that matter I go further: a panel of models from different vendors works the same question independently, a separate model judges their answers blind, and the synthesis has to survive all of them. Assumptions get the same treatment as answers. Where models agree for different reasons, I look harder; where they disagree, a human decides. Finance has a name for this instinct already: no preparer approves their own work.

Validation against source. Nothing AI-assisted informs a decision until it has been checked against source data, the same discipline as reviewing a junior analyst's work, applied without exception. The tools are fast, tireless, and occasionally confidently wrong. The finance leader who treats AI output as an answer rather than a draft has not understood the technology; they have delegated their judgement to it.

The learning loop. When an output misses, the miss is logged and the workflow is amended, so the same failure cannot arrive twice. This is how a finance leader develops an analyst, applied to a tool, at a pace no human team could match: continuous critique reins the model in rather than letting it drift, and every correction builds into the practice. The seasonality catch earlier in this piece is an example: found in review, verified against source, and an assumption audit has been standing practice on every cycle since. This is what separates a practice that compounds from a team that is perpetually experimenting.

Where the data goes. Accuracy is not the only exposure. On client work, confidential data is processed only within whatever AI arrangements the client has approved: enterprise controls, private deployments, or not at all. The fluency transfers; the data governance is the client's call, and it is agreed before the first prompt, not after.

The same control instincts that run a listed-company close, applied to a new class of tool.

Where I would start on a live P&L

Make no mistake about the scale of the change: it is transformative. Finance lifted as business partner, as safeguard of the business's assets, and as the function that reports and controls, with roles changing around it. But starting does not require a transformation programme. On a mandate, the first month has a simple shape. Week one: map the driver tree with the people who own the numbers, demand, price, mix, marketing, supply and cash, and agree what actually causes what. Week two: the assumption audit, with AI surfacing every embedded assumption in the current model and the team re-justifying or retiring each one. Week three: the sensitivity pass, ranking the drivers and agreeing the three that matter this year and the watch-list behind them. Week four: stand up a controlled working prototype around the highest-value driver set, with the controls wrapped around it from day one. The calendar stretches with complexity, product count, data readiness and the team's other commitments; it is the sequence that holds, not the timetable.

Boards are now asking their finance leaders for an AI position. The credible answer is not a slide about the future. It is a working description of your own practice: which tools, on which workflows, governed how. That answer can only be earned by use.