AI in FP&A: Driver-Based Forecasts, Scenarios and Variance Analysis
Turn this article into takeaways for your work.
Each assistant summarizes the article only for you and suggests best practices for your work.
AI in FP&A (financial planning and analysis) is the use of machine learning and language models to build baseline forecasts, test which business drivers move the numbers, run scenarios and draft variance commentary, while analysts own the assumptions, the story and the recommendation. It speeds up the parts of planning that are mechanical and repetitive. It doesn't replace the part where someone decides what the numbers mean and what to do next.
This is the hub for the planning and treasury section of the AI for Finance collection. It's written for CFOs, FP&A leads and finance managers who need to know where AI earns its place in the planning cycle and where it doesn't. For the whole-function view, start with AI in finance and accounting. From here, the detail on each job lives in a dedicated guide: AI cash flow forecasting, AI variance analysis, AI scenario planning, AI spend analytics and AI in treasury.
Key Facts
- The 2025 AFP FP&A Benchmarking Survey (362 FP&A practitioners worldwide) found 23% using AI daily, weekly or monthly, and another 40% testing it with plans to implement within a year (AFP).
- In the same AFP survey, 61% cited lack of data reliability and 60% cited lack of accessible data as challenges, and 96% still use spreadsheets for planning (AFP).
- Gartner's 2025 AI in Finance Survey (183 CFOs and senior finance leaders) found 59% of finance functions using AI, against 58% in 2024 and 37% in 2023, and 91% of respondents initially saw low or moderate impact (Gartner, via CPA Practice Advisor).
- In a PLOS ONE study of 1,045 monthly series from the M3 competition, eight statistical methods beat the machine learning methods tested on every accuracy measure and every forecast horizon (Makridakis, Spiliotis and Assimakopoulos, 2018).
- A 2024 study of three LLM-based forecasters found that removing the language model, or swapping it for a basic attention layer, did not degrade forecasting performance (Tan et al., arXiv).
What AI Changes in FP&A
FP&A has four recurring jobs: forecast the numbers, explain what moves them, test what-if cases and tell leadership what it all means. AI helps unevenly across those four.
Forecasting. Machine learning reads historical actuals, seasonality and outside signals, and produces a baseline faster than a spreadsheet model can be rebuilt each month. Think of predictive analytics applied to revenue, headcount cost or demand. The baseline is a starting point, not a verdict.
Driver modeling. A driver-based plan links the P&L to operational inputs: leads to pipeline to bookings to revenue, or units and price to margin. AI helps by testing which drivers actually explain the history. A model can show that churn explains more of your revenue swing than new-logo volume does, which tells you where to point the plan.
Scenario analysis. Once drivers are wired in, running ten cases takes minutes instead of days. That changes how often a team gets asked "what if" questions and says yes. The craft of choosing meaningful scenarios is covered in scenario planning, and the AI-specific mechanics in AI scenario planning.
Commentary. A language model can read the ledger detail behind a variance and draft the first paragraph of the explanation. An analyst still checks every claim, because the draft sounds equally confident when it's wrong.
Where AI Fits in the Planning Cycle
| Planning task | What AI contributes | What the analyst still owns | Risk level |
|---|---|---|---|
| Baseline forecast | Statistical or ML projection from history and seasonality | Whether history is a fair guide this year | Low to medium |
| Driver selection | Tests which inputs explain past results | Which drivers management can actually control | Medium |
| Scenario runs | Executes many cases quickly, flags sensitive inputs | Which scenarios are plausible and worth showing | Medium |
| Variance explanation | Drafts commentary from ledger and operational data | Truth of each claim and the "so what" | Medium |
| Cash forecast | Projects receipts and payments from payment behavior | Funding decisions and liquidity buffers | Medium to high |
| Spend review | Clusters and categorizes spend, finds outliers | Whether to renegotiate, consolidate or cut | Low |
| Board narrative | Drafts text from approved figures | Final wording and every external number | High |
Real Examples by Function
A SaaS finance team forecasts new bookings. Instead of a flat growth rate, the model uses pipeline by stage, historical stage-to-close rates and seasonality. The analyst compares the model's number to the sales leader's call each month. When they disagree by more than a set margin, that's the conversation worth having.
A retailer's merchandise finance group forecasts category demand using price, promotion calendar and weather. Granular history with many repeating patterns is where machine learning has the best chance. It doesn't win everywhere, which is why the team tests it against a simple baseline first.
A services firm wants a faster monthly variance package. A language model drafts "revenue is $400k under budget, driven by lower utilization in two practices" from the ledger and timesheet data. The FP&A manager confirms the utilization figures against the source before it goes to the CFO, and rewrites the recommendation herself.
A manufacturer's operations finance team feeds purchasing data into a spend-analytics tool that groups thousands of free-text line items into categories and highlights suppliers whose prices drift. The finance partner uses that to prepare for supplier talks. AI spend analytics goes deeper on how this works.
A treasury team forecasts weekly cash from historical payment timing by customer, then compares the forecast to actuals every Monday. AI cash flow forecasting and AI in treasury cover the method and the controls.
How to Implement AI in FP&A, Step by Step
- Start with a baseline you can beat. Pick one forecast line, say monthly revenue or cash receipts, and record how accurate your current method is. Without that number you can't prove anything improved.
- Check the data before the tool. Look at how consistent your chart of accounts, product hierarchy and customer records are across periods. The AFP survey's top obstacles were data reliability and accessibility, not the tools themselves. The checklist in finance data readiness for AI is a good first pass.
- Test a simple statistical method alongside the AI one. Run a seasonal naive forecast or a basic exponential smoothing model next to the machine learning model on the same history. If the complex model doesn't beat the simple one out of sample, don't deploy it.
- Build the driver tree with the business. Sit down with sales, operations and HR to agree which three to six drivers matter, who owns each and where the data comes from. A model without owners for its inputs becomes a black box nobody trusts.
- Backtest honestly. Hold out the last 6 to 12 months, forecast them blind and measure the error. Keep the test period out of training so the result isn't flattered.
- Run in parallel for a full cycle. Produce the AI forecast and the human forecast side by side for at least one quarter, and review the gaps. Document where the model was better and where it missed.
- Define the review step. For every output, name who checks it, what they compare it to and what they can override. Treat it as real human-in-the-loop review, not a click-through.
- Track accuracy continuously. Measure forecast error each period, not once. Models drift when the business changes, and a pricing change or new product can break last year's relationships overnight.
- Then consider agents. Once the manual workflow is stable, an AI forecasting agent can run it on a schedule and raise exceptions for review.
Adoption: What the Evidence Shows
Adoption in FP&A specifically is earlier than the overall finance numbers suggest. The AFP's 2025 survey found only 23% of respondents using AI daily, weekly or monthly, with 40% testing and planning to roll out within a year. That's a profession at the experimentation stage, not the scaling one.
The broader Gartner series shows the same pattern from the CFO side. Finance AI use rose from 37% in 2023 to 58% in 2024 and 59% in 2025, and 91% of respondents reported low or moderate impact at first. Organizations further along were nearly three times as likely to report high impact. The pattern suggests value arrives after the pilot, once the data and processes are fixed.
Those two surveys measure different populations (FP&A practitioners versus CFOs and senior finance leaders) with different definitions of "using AI," so don't line their percentages up as if they were one trend.
The Limits of AI in FP&A
Complex doesn't mean more accurate. In the PLOS ONE study of M3 data, the machine learning methods were dominated by simple statistical ones across every accuracy measure and horizon, and they needed considerably more computation. The authors caution that results may depend on the data, such as its length and business-oriented nature. Treat that as a reason to test on your own history. Always keep a simple benchmark in the test.
Language models aren't forecasters. A 2024 study of LLM-based time series forecasters found that stripping out the language model, or replacing it with a basic attention layer, did not hurt accuracy. Use language models for what they're good at, reading and drafting, and use time series models for projecting numbers.
History isn't always a guide. Every statistical forecast assumes the future resembles the past. A price change, a lost customer, a regulatory shift or a supply shock breaks that. This is where analyst judgment and scenario work matter most.
Hallucinated commentary. A model can write a persuasive variance explanation that cites a driver that didn't change. The AI hallucination problem is real in finance Q&A: FinanceBench, a test of 10,231 questions about public companies, found GPT-4 Turbo with a retrieval system answered incorrectly or refused 81%. That was a 2023 test of specific models, and newer ones do better, but it justifies requiring every number in a draft to point to a source cell.
Data quality. Both the AFP and Gartner surveys put data issues ahead of technology as the main obstacle. Models trained on inconsistent mappings produce confident, wrong baselines.
Over-trust. A polished forecast that nobody questions is the failure to watch for. Keep a person responsible for each number, and keep comparing predictions to actuals.
What to Measure
- Forecast accuracy: mean absolute percentage error by line and horizon, versus your pre-AI baseline.
- Bias: whether errors lean consistently high or low, which is often more costly than random error.
- Cycle time: days from period end to forecast and variance package.
- Analyst time: hours spent on assembly versus analysis.
- Decision use: how often forecasts and scenarios changed a decision, not just a slide.
Where to Go Next
If you do one thing this quarter, pick a single forecast line, record today's accuracy and run a simple statistical model next to any AI tool for a full cycle. Then work through the guides in this section: start with AI variance analysis if commentary eats your close week, AI cash flow forecasting if liquidity is the worry, and AI scenario planning if leadership keeps asking "what if." The best result is a forecast your team trusts because it has a track record, not because it came from a model.