AI in FP&A: A Practical Guide for Finance Teams

Consider a typical FP&A team of three in a business of around 200 people. The ledger closes on working day five, and the next four days go on building the pack: refreshing the model, writing commentary for a dozen cost centres, answering the CFO's follow-up questions. The work the business actually wants from FP&A, deciding what to do about the numbers, gets whatever time is left.

Research consistently shows finance professionals spend roughly 75% of their time gathering, cleaning and reconciling data, leaving around 25% for analysis. That is where most teams find AI useful first. Not in replacing the model, but in shortening the work around it. This guide covers where AI fits in the FP&A cycle, a worked budget vs actual example, what to expect from AI forecasting, which tools to use and how to make the case for it. For AI across every finance process, see our guide to AI use cases in finance.

Where AI Fits in the FP&A Cycle

It helps to separate two kinds of AI that often get lumped together. The first is the generative assistant: Copilot, ChatGPT, Claude, Gemini. It reads and writes text, works with tables you give it, writes formulas and code, and answers questions about a model. The second is statistical or machine-learning forecasting, usually built into a planning platform, which finds patterns in your history and projects them forward. Most disappointment with AI in FP&A comes from asking the first kind to do the second kind's job.

Walk through a normal month and the fit becomes clear:

• Data preparation. Mostly not an AI problem. Power Query, ERP reports and a stable chart of accounts do more here than any assistant, although an assistant can write the query for you.
• Finding the variances. An assistant can recalculate, rank by materiality and flag lines that look wrong. Useful as a check, quick to set up.
• Commentary. The clearest win. A first draft from figures plus context notes, reviewed by the analyst.
• Forecasting. Statistical baselines belong in a planning tool. The assistant helps with driver logic, assumptions and documentation.
• Scenarios and questions. Asking "what happens to cash if debtor days rise by ten?" in plain English, then checking the formula it builds.

Day to day, the change is that the analyst spends less time producing first drafts and more time reviewing them. That only works if review is a deliberate step with a method, not a quick read.

AI for Variance Analysis: A Budget vs Actual Example

Take a sample month, in £000. Revenue is £1,240k against a budget of £1,310k, £70k adverse. Payroll is £15k over budget, marketing £17k under, and EBITDA lands at £179k against £234k. Paste that table into an assistant and ask for commentary, and you will get fluent paragraphs back. The question is whether the reasons in them are true.

The model does not know that two customer go-lives slipped by a month, that payroll includes two contractors covering vacancies, or that the marketing campaign simply moved. Without those facts it will supply plausible ones. So the method is to give it the facts as short context notes, tell it to use only those, and have it write [DRIVER NEEDED] where a material variance has no explanation.

Some explanation can be automatic. If you give the assistant volume and price by product, it can split a revenue variance into price, volume and mix itself, because that is arithmetic rather than knowledge. The why behind the volume drop still has to come from the business.

To see this on a real team, read our forecast commentary case study, covering how one finance team reduced their commentary time from three days to four hours. The full prompts, the eight-point sign-off checklist and a downloadable Excel kit with this sample month are in our guide to checking AI variance commentary before it reaches the board. For ready-made prompts covering revenue, cost, price-volume and headcount variances, see our variance commentary prompts.

AI for Forecasting: What It Can and Cannot Do

Do not ask a chat assistant to produce forecast numbers from a description of your business. It will give you numbers, and they will be guesses dressed as analysis. Where assistants earn their place in forecasting is around the model rather than inside it.

Good uses for an assistant:

• Reviewing a forecast model's logic and pointing out hard-coded numbers, broken links or circular references.
• Turning a list of drivers into formulas, then explaining each formula back to you.
• Writing the assumptions log that nobody has time to write.
• Running a simple statistical baseline from your history, where your tool can execute code, as a comparison point rather than the answer.

For a statistical forecast you rely on, planning platforms such as Workday Adaptive Planning and Anaplan include machine-learning forecasting features that work on your actual history inside the model. They need several years of clean, consistently coded data. If your history has been restructured twice, fix that first, because a confident forecast built on messy data is more dangerous than a manual one. Whatever you use, judge it by backtesting on months you have already closed. Our AI budgeting and forecasting guide covers the annual budget side.

Scenario Modelling With a Conversational Assistant

Conversational AI changes how scenario questions get asked more than how they get answered. Instead of rebuilding a sheet, the analyst asks the assistant inside the workbook to add the scenario, and then checks what it built. A prompt like this works in Copilot in Excel or Claude for Excel:

Add a scenario column to the cash flow sheet where debtor days rise from 45 to 55 from next month. Keep the base case untouched. Show the change in closing cash for each of the next six months, and list every cell you changed.

The last line matters. Asking the assistant to list what it changed turns a black box into something you can review in a minute. Check one month by hand. If the arithmetic holds, you can run three or four scenarios in the time one used to take, and spend the saved time on the conversation with the business about which one to plan for.

Treat every scenario output as a draft model. Copy the logic into your controlled model only after a person has checked it, and keep the prompt with the file so someone else can see how the numbers were produced.

AI Tools for FP&A: What to Use

Pick tools by job, starting with what your organisation has already approved for financial data. A new platform you cannot get through IT is not an option, however good the demo.

For most teams the first investment is the assistant layer, used well. A planning platform earns its place once the planning model has outgrown Excel. Our FP&A AI tools ranking compares the platforms, and our Claude vs ChatGPT vs Copilot comparison covers the assistants task by task.

Building a Business Case for AI in FP&A

A business case for AI in FP&A should rest on hours you have measured, not on percentages from a vendor deck. The calculation is simple: hours spent on the task each month, times the share AI removes in your pilot, times the loaded hourly cost, times twelve, minus the licence and set-up cost.

Here is how the arithmetic works with illustrative inputs. Suppose the team logs 24 hours a month on commentary, and a four-week pilot on closed months shows drafting time falls by half. At a loaded cost of £45 an hour, that is 12 hours times £45 times 12 months, or £6,480 a year of analyst time redirected, before licence costs. Your numbers will differ. The point is that every input is one you measured.

Two things make the case stronger than the hours alone. Faster commentary can move the pack earlier, which has value to the board. And a documented checking method answers the question the CFO will ask first, which is what stops a wrong number reaching the board. Our AI business case template for finance sets this out in full.

Frequently Asked Questions

Our AI for Finance Leaders course teaches these FP&A workflows with marked practice, and the first two modules are free to try. For your own team, our AI audit and assessment measures where AI would save the most time, and our AI consulting team can run the pilot with you.