AI for FP&A: Start With the Commentary, Not the Model
An FP&A team should start using AI on the monthly variance commentary, because it takes 2 to 4 analyst days a month, it follows rules you can write down, and every sentence can be checked against the ledger. Leave the forecast model alone for the first 90 days, and move the agent on to budget holder questions and scenarios only once your controller keeps 8 of every 10 lines it drafts without an edit.
Why FP&A teams start in the wrong place
Most finance teams I meet start with the model. Someone asks an AI tool to rebuild the 3 statement model or predict next quarter, and 6 weeks later the CFO has a forecast nobody can trace and a team that trusts AI less than before. The model is the one place where a wrong number goes straight into the board pack and the bank covenant test.
Commentary is the opposite. It's repetitive: the same 30 to 60 lines explained every month. It has a right answer you can check: the revenue line moved because 2 named customers renewed late, and the CRM says so. And it eats the week. In a 100 to 300 person company I usually find the team spends working days 3 to 6 writing sentences about numbers that were final on day 2. That's the work to hand over first.
The five steps I use inside an FP&A team
1. Time one month end before you build anything
For one cycle, every analyst logs their hours in half day blocks against 5 buckets: pulling data, reconciling, writing commentary, answering budget holders, and modelling. It costs nothing and it settles the argument. If commentary and budget holder questions together take more than 3 days, that's your first build. If pulling data takes longer than both, fix the data feed first, because an agent can't write about numbers it can't reach.
2. Write the commentary rulebook
This is a 2 page document your best analyst could hand to a new joiner. It holds the materiality line (I use any variance over £10,000 or 5% of budget, whichever is bigger), the list of acceptable causes (timing, volume, price, one off, reclass, error), the house words for each, and the sources that prove a cause: CRM for revenue, HR for payroll, purchase orders for suppliers. The agent loads this every run. When a sentence is wrong, you fix the rulebook, not the sentence.
3. Calculate first, then let the agent write
The variances are worked out in the ledger or the spreadsheet, the way they are today. The agent never does arithmetic. It reads the finished variance table, picks every line over the materiality line, looks up the cause in the source systems, and writes one or 2 sentences per line. Every sentence carries a tag back to the rows it read, so the controller can click from "marketing is £18,400 over because of the March conference" to the 3 invoices. I set the draft to land by 08:00 on working day 3.
4. Measure the keep rate, and gate on it
The controller marks each drafted line keep, edit or delete. That one column is the whole performance review. In the first month I'd expect 5 or 6 of 10 lines kept. Every edit goes back into the rulebook as a new rule or a new cause. The draft goes into the board pack only after 2 months in a row at 8 of 10 or better. Most teams get there by month 3.
5. Then move outward, one stage at a time
Once commentary passes, the same agent and rulebook take the next job. Each stage has a gate, and nothing skips one.
| Stage | What the agent does | Gate to move on | Typical timing |
|---|---|---|---|
| Variance commentary | Drafts one sentence per material line, source tagged | 8 of 10 lines kept, 2 months running | Months 1 to 3 |
| Budget holder questions | Answers "why is my cost centre over" from the same data, read only | Under 1 in 10 answers escalated to an analyst | Months 3 to 4 |
| Pack narrative | Drafts the 1 page summary the CFO opens the pack with | CFO rewrites under a third of it | Months 4 to 5 |
| Scenarios | Changes inputs on a copy of the model, never formulas | Every run logged with who asked and what changed | Month 6 onward |
Budget holder questions are the quiet win. A head of sales who can ask at 21:00 why their travel line is £6,000 over, and get an answer with the 4 expense claims attached, stops emailing your analyst the next morning. The forecast drivers come last, and I cover how to build them in the AI financial forecasting guide.
From my operating seat
I'm Hayat Amin, an AI-native CFO, AI operator and AI coach for CFOs who teaches finance leaders how to build AI into finance operations. I spent twenty years in the C-suite, with three exits and three FT100 listings. Now I build and run AI agent systems inside companies, across finance, sales pipeline, onboarding and compliance evidence, and I coach CFOs one to one on the same work.
When I ran FP&A myself, the commentary was the part I dreaded and the part the board read most closely. Buyers in due diligence read it too. They line up 24 months of it and check whether the stories agree with each other. Inside one client today the agent drafts commentary overnight from the ledger, the CRM and the HR system, and the same agent keeps a live cash runway figure that the CFO sees every morning. The analysts spend working day 3 arguing about the 4 or 5 lines that matter. I'd rather they did that than type the other 40.
Can AI do variance analysis?
Yes, if the maths is done before the model sees it. Calculate the variances in the ledger or a spreadsheet, then give the agent the numbers, your materiality line and the source systems that explain causes: CRM, HR, purchase orders. It writes one sentence per line over the threshold, each tagged to the rows it read. A person still signs it. The agent should never calculate a variance in its head.
What are the best AI use cases in FP&A?
In the order I build them: variance commentary, answers to budget holder questions, the first draft of the monthly pack narrative, scenario runs on a copy of the model, and collecting budget inputs. Each one reads data you already hold and produces something a person checks. Forecast drivers come later, and only the ones with 24 months of history and a written rule.
Will AI replace FP&A analysts?
It replaces the part of the job that is copying numbers into sentences. In the teams I run, that's about 2 to 4 days of every month. The analyst keeps the judgement: which variance matters, what the budget holder isn't saying, what the board will ask. A 4 person FP&A team usually stays at 4 and spends the freed days on pricing, unit economics and the next raise.
Where I come in
This is the work I do with CFOs one to one and inside FP&A teams. In the first session we time your last month end and write the commentary rulebook together, and by your next close the agent is drafting beside your analysts. For the wider build order see the first five finance agents, and for the tool see Claude for finance teams. The map of the whole topic is at AI for CFOs, the end state is at the AI-native CFO, and the sessions are at coaching for CFOs. How many days did your team spend on commentary last month? Start at meethayat.com and tell me.