Read the guide online
The full guide, chapter by chapter. Start at the foundation or jump straight to the part you need.
- Part IThe FoundationThe groundwork for AI forecasting: the business case in euros, 24 months of clean history, bank connectivity, transaction categorisation and ERP data.
- Part IIThe AIWhich AI models fit which cash flow categories, the agentic approach, human-in-the-loop autonomy, the LLM as orchestrator, and AI security and governance.
- Part IIIThe OrganisationMulti-entity and multi-currency forecasting, intercompany eliminations, central versus local forecasting, AI as the quality layer, and common pitfalls.
- Part IVThe ValueTurning forecasts into decisions: excess cash, shortages, FX exposure, early warnings, actuals versus forecast, forecast versions and scenario analysis.
- Part VThe FutureFrom recommendations to execution: how agentic AI turns treasury into a real-time, continuously optimising function, with humans setting the boundaries.
Who this guide is for
Treasury and finance teams that want a reliable cash flow forecast without weeks of chasing subsidiaries and fixing spreadsheets. It works whether you build the system yourself, run it in spreadsheets or use a platform: the principles are the same.
The guide follows the order in which the work actually happens: first clean bank and ERP data and consistent categories, then the AI that fills the gaps, then the organisation around it, and finally the decisions a good forecast lets you make.
Frequently asked questions
How much historical data do you need for AI cash flow forecasting?
Aim for at least 24 months of bank transactions, so the model sees full annual cycles: seasonality, tax months and quarter-end patterns. With 36 months or more, it gets easier to tell real patterns from one-off events.
Which AI models work best for cash flow forecasting?
It depends on the cash flow category. Time-series models such as Prophet, ARIMA or Holt-Winters suit categories with many transactions and clear seasonality. Gradient boosting can add external drivers such as payment terms. Simple statistical models often win for stable or sparse categories like rent. The best results come from choosing a model per category, every time the forecast runs.
Should the AI fill in the forecast on its own?
Treat autonomy as a dial. Start advisory, with a person approving every AI entry. Move to semi-autonomous once you trust it, where the AI fills high-confidence categories and flags the rest. Only then give it high autonomy, with escalation when something unusual appears.
How do you improve forecast accuracy over time?
Compare actuals with the forecast continuously, per category and per entity. Look for systematic errors, such as always over-forecasting supplier payments, and feed them back into the next forecast. Each cycle makes the next forecast better.
How often should you create a new forecast version?
Many treasury teams create one every week. Companies with stable cash flows often work monthly, in line with their reporting cycle, and some add versions before events like a board meeting. Lock each version once it is final, so you can always compare what you forecast with what happened.