Short answer
Mostly in the hours before analysis. Pulling and checking data, drafting variance commentary, catching model errors and running baseline forecasts. The judgment calls stay with your team.
Key takeaways
- Spreadsheet-based FP&A teams spend about half their time collecting and checking data. That is where AI pays back first.
- Variance commentary and model error checks are the next best uses, and Copilot and Claude now do both inside Excel.
- AI forecasting earns its place on lines with long, clean history. Run it beside your current forecast for three cycles before you trust it.
- Fix the data feed before buying a platform. Bad or inaccessible data is the top obstacle FP&A teams report.
For an FP&A team that lives in Excel, AI helps most with the work before the analysis: pulling actuals and checking them, drafting variance commentary, catching errors in the model, and producing a baseline forecast for lines with long, clean history. It helps least with the parts that need judgment, like pricing decisions, new customers, hiring plans and the story you tell the board.
The reason is where the hours go. In the 2025 FP&A Trends Survey, FP&A teams spent 46% of their time on data collection and validation and 31% on insight and action. Teams that rely mainly on spreadsheets spent 50% on collecting and validating data and 28% on insight. Most of that 50% is the first place AI earns its keep.
Where AI helps, in order of payoff
| Work | What AI does | What it needs |
|---|---|---|
| Data pulls and checks | Loads actuals, maps accounts, ties totals back to the ERP, flags breaks | Access to the ERP, CRM and HRIS, and a stable chart of accounts |
| Variance analysis and commentary | Compares actuals to plan, finds the drivers, drafts the first explanation | Transaction detail behind the totals, not only the summary tab |
| Model review | Traces assumptions, explains formulas, finds broken links and hardcodes | The workbook, plus someone who knows what it should do |
| Baseline forecasts | Statistical or machine learning forecast for volume, revenue by product, collections | Two or more years of consistent history per line |
| Questions from the business | Answers “why is T&E over budget” from the actual numbers | Clean data and access controls |
Gartner’s 2026 survey of 160 finance leaders found that data extraction, AP and AR automation, and report creation generally pay back within 9 to 10 months. Data management, insight generation and forecasting take longer to deliver value. That matches the order above.
What the tools in Excel can do now
You may not need a new platform to start. Microsoft says Copilot in Excel’s finance skills are generally available, and its newer accounting capabilities, in early access through its Frontier program, analyze performance against plan, identify the key drivers and prepare reporting that explains what changed and why. It also describes consolidating budget inputs into a forecast-review brief.
Claude for Excel, an add-in from Anthropic, answers questions about a workbook with cell-level citations, traces how a number was derived, finds errors and their root causes, and changes assumptions while keeping formulas intact.
Excel has also had statistical forecasting for years. FORECAST.ETS predicts future values from history using exponential smoothing with seasonality, so a baseline forecast doesn’t require a new system.
Model review matters more than most teams admit. In field audits collected by spreadsheet researcher Ray Panko, Coopers and Lybrand found errors in 91% of the 23 accounting spreadsheets it audited, counting only errors that moved a result by 5% or more. An AI pass over every formula before the board pack goes out is cheap insurance.
Where AI won’t help much yet
- Assumptions that live in people’s heads. A model can’t forecast a price increase, a lost customer or a hiring freeze that hasn’t been decided.
- Short or broken history. After an acquisition or a chart of accounts rebuild, there may not be enough consistent data for a statistical forecast.
- Bad source data. In AFP’s 2025 FP&A benchmarking survey, 61% of practitioners said unreliable data held back their technology, and 60% said data they couldn’t get to did. AI doesn’t fix either.
- The board narrative. AI can draft it. The CFO still has to decide what it says.
How to start, in order
- Time one monthly cycle. Log the hours for actuals load, reconciliation to the ERP, variance work, forecast refresh and the board or lender pack. Start with the biggest block.
- Fix the data feed first. One reliable connection from the ERP, CRM and HRIS into the workbook removes the largest block of manual work. Excel-native tools such as Cube, Datarails and Vena do this while keeping your models in Excel.
- Use the AI already in Excel for review and commentary. Have it check formulas and draft variance explanations, and have an analyst edit rather than write from scratch. Use it only on trusted files. Anthropic warns that workbooks from outside sources can contain hidden instructions.
- Run an AI baseline forecast beside your own for three cycles on two or three lines with good history. Compare the error each cycle and keep whichever is better. In the FP&A Trends survey, the few teams already using AI or machine learning rated their forecasts great or good 73% of the time, against 42% for teams that don’t.
- Move to drivers where you can measure them. In the FP&A Trends survey, 77% of teams using driver-based models rated their forecasts accurate, against 38% of the rest. Among teams with fully or dynamic driver-based models, 20% produce a forecast in under two days, while 38% of teams on static models need more than 10 days.
- Keep a person’s sign-off on anything that leaves finance, with the source of each number recorded.
What goes wrong
- Buying a platform to fix a data problem. If the ERP extract is wrong, a new tool just presents the wrong number faster.
- Commentary that describes but doesn’t explain. “T&E is 12% over plan” isn’t an explanation. Without transaction detail, AI can only restate the variance.
- Expecting forecasting to pay back first. Gartner’s data puts forecasting among the slower use cases. Start with data and reporting and let forecasting follow.
- Treating an AI forecast as finished. Gartner predicted that by 2028, half of organizations will have replaced time-consuming bottom-up forecasting with AI. The ones that get there still need analysts who can say why the model moved before it reaches the CFO.