AI how-to guides Finance
How to do financial forecasting with AI
Feed a forecasting model the signals that move your numbers (sales, ERP, pipeline, FX, operations) so it recalibrates as new data arrives and explains which drivers changed the figure. Finance reviews, overrides and approves before anything is published. Start with one forecast that matters, test it against past cycles, then widen.
What AI does
- Recalculates the forecast whenever new actuals and signals arrive, instead of waiting for the next reforecast cycle.
- Gives a range, not just a single number, so you can see how confident the forecast is.
- Breaks each forecast down into its main drivers and shows what changed since the last version.
- Runs scenarios (FX, demand, pricing) on the same model so they stay consistent with the base case.
- Records the inputs, model version and reviewer behind every published figure.
What stays with people
- Approving the forecast that goes to leadership or the board, with a named owner.
- Overriding the model when they know something the data does not, such as a contract about to be signed or a price change.
- Choosing the assumptions behind board-facing scenarios.
- Deciding what to do with the forecast: budgets, hiring, investment.
How to put it in place
-
Pick the forecast and the decision it supports
Choose one forecast with real weight, such as revenue by business line or operating costs. Write down which decision it feeds, how often it should refresh and how far ahead it must look.
-
Gather and clean the history and the drivers
Pull actuals from the ERP and the signals that explain them: pipeline, volumes, prices, FX, headcount. Agree one definition per figure and fix gaps before modelling, because the model will learn whatever the data says.
-
Backtest against past cycles
Train the model on history and check how it would have forecast periods you already know, compared with your current method. Keep the approach only if it beats the baseline and its ranges are honest.
-
Design the review and override layer with finance
Decide who reviews each forecast, how overrides are recorded and who signs before publication. Show the forecast where finance already works, such as Excel or the BI tool, with its drivers and range.
-
Run in parallel, then switch over
Produce the AI forecast alongside the usual process for a few cycles and compare both with the actuals. Move over when finance trusts it, and keep measuring error so drift shows up early.
What you need
- Several years of clean actuals at the level you want to forecast.
- Access to the ERP, CRM and other systems that hold the drivers.
- Agreed definitions for each figure and each driver.
- A finance owner who reviews, overrides and signs off.
- A way to version models and record who approved what.
When it is not worth it
- Too little history, or history that no longer reflects how the business works.
- Figures driven by a handful of large, one-off deals where judgement beats any model.
- When data quality is so poor that the first job is to fix the data, not to model it.
Risks and how to control them
- Forecasts that look precise but are not: always publish a range and track error against actuals.
- Black-box numbers nobody can defend: require driver explanations and keep the model simple enough to explain.
- Breaks in the business (an acquisition, a new product, a shock) that history cannot predict: let finance override and record why.
- Lost audit trail: version inputs, models and approvals so any published figure can be traced back.
How Thinkia does it
We start by mapping the current forecast cycle, data sources, governance gaps and the decisions the forecast must support. Then we define the driver hierarchy, approval gates, sensitivity ranges and which scenarios are board-facing, and we validate models against your historical cycles.
Finance owns the override layer: assumptions can be locked and signed by a named reviewer, and every published forecast traces back to its inputs, model version and approval. It runs on Synapse and surfaces in the tools finance already uses.
The Thinkia solution behind this guide
- AI financial forecastingRolling forecasts your CFO can defend
- Treasury cash-flow intelligenceLive cash, FX, and liquidity in one governed view
Thinkia products involved
Frequently asked questions
Can AI predict revenue more accurately than a spreadsheet?
Often, when there is enough clean history and the drivers are in the data. The only way to know is to backtest it against your past cycles and compare it with your current method before relying on it.
Does AI replace the FP&A team?
No. It takes over the mechanical recalculation and the first explanation of variances. The team keeps the judgement: overrides, assumptions, scenarios and the conversation with the business.
How do you explain an AI forecast to the board?
By showing the range, the main drivers and what changed since the last version, plus who reviewed and approved it. If a figure cannot be explained that way, it should not go to the board.
What data do I need to start?
Actuals at the level you want to forecast and the main signals that move them, such as pipeline, volumes, prices and FX. Consistent definitions matter more than having every possible source.
Can I use generative AI or ChatGPT for forecasting?
Language models help to explain variances and draft commentary, but the forecast itself should come from statistical or machine learning models that can be backtested and audited.
Keep exploring
Related guides
Related decisions
- Fix the data first or start with AI now: which comes first?
- AI pilot vs production: what actually changes when you scale?