Should treasury trust an AI forecast over human judgement?
Not automatically. Statistical forecasts are useful as an independent reference, but treasury should override them when it has credible business information the historical data cannot contain, especially during structural changes or unusual events.
Cash flow forecasting works, but it carries a high manual cost
Cash flow forecasting is not fundamentally broken. Treasury teams have produced useful forecasts for years. The weakness is that the process often depends on time, local knowledge and the people available in each reporting cycle.
Subsidiary forecasts can arrive late. Quality can vary when the person responsible is busy or absent. Experienced employees may understand the quirks behind a cash flow that are never written down, and that knowledge can leave with them.
AI cash flow forecasting is useful when it reduces that dependency. The strongest business case is often straightforward: spend less time creating and checking forecasts, and more time investigating the risks that could create the next liquidity surprise.
Forecast quality is both a data problem and a modelling problem
Poorly classified cash flows will limit any forecasting method. Treasury needs consistent categories and a usable history before a model can learn anything meaningful.
Good data does not make every cash flow easy to predict. Project-based businesses, irregular customer payments and structural changes can remain difficult even when the underlying transactions are correctly classified.
That is why AI forecasting should not be reduced to 'garbage in, garbage out'. Data quality is one part of the problem. The technology still has to deal with seasonality, outliers, different payment patterns and time series that behave very differently from one another.
Several years of history help separate patterns from accidents
Johannes Pöschl describes three to four years of historical data as a useful target for identifying recurring patterns and seasonality. With only a short history, it can be difficult to know whether a high or low month represents a genuine seasonal effect or a one-off event.
The forecast process also needs a test period. In Nomentia's approach, multiple models compete for each cash flow series against a period where the actual outcome is already known.
Over time, repeated testing gives treasury more evidence that a selected model performs well for that specific type of cash flow rather than simply winning once by chance.
A good model should show where its confidence comes from
Explainability matters because treasury remains accountable for the forecast. A prediction should not arrive as a number with no supporting context.
Useful supporting detail includes detected outliers, seasonality, trend changes and recurring payment days. Those signals help the user understand why the automated forecast looks different from the manual expectation.
The same standard should apply to inherited spreadsheets. Familiarity is not the same as explainability. Treasury should understand the logic behind any forecast it relies on, regardless of whether the calculation comes from Excel or an AI model.
The treasurer should overrule the model when business context changes the answer
Historical models are strongest where past behaviour remains relevant. They are weaker when the underlying business changes abruptly or when the decisive information has never appeared in historical data.
A major acquisition, a structural change in the business or a genuine crisis can require information that only people inside the company possess. In those situations, treasury judgement should carry more weight.
The useful model is therefore not one that asks the treasurer to surrender ownership. It provides a second view that can standardise repetitive forecasting while leaving room for human adjustment when new information changes the outlook.
Bias is another reason to compare human and statistical forecasts
Human forecasting can include behavioural bias. Teams may be consistently optimistic, conservative or influenced by the latest event. Those tendencies can differ between people, subsidiaries and regions.
A statistical forecast provides a common reference point. If the business forecast and the AI-generated view differ materially, the variance becomes a question to investigate rather than a reason to assume one side is automatically correct.
That comparison can improve forecast discipline because treasury has to explain what the business knows that the historical data does not, or what the historical pattern may be capturing that the manual process has missed.
The strongest business case is time and consistency
Treasury teams are usually lean. The value of AI cash flow forecasting is therefore less about replacing the person responsible for liquidity and more about reducing repetitive forecasting work.
A reliable automated forecast available on a predictable schedule can free time for counterparty risk, funding questions, unusual exposures and other issues that require judgement.
The size of the company is not the only factor. Even a smaller finance team can benefit if one person spends a meaningful part of the week building and checking liquidity forecasts.
AI forecasting is useful precisely because it is not a promise of certainty
No forecast can know every future event. Environmental shocks can defeat both human judgement and statistical models.
Treasury should therefore evaluate AI forecasting on the quality of the process: the data used, the testing approach, the supporting analysis, the ability to compare predictions with actual outcomes and the ease with which a user can challenge the result.
The practical objective is a better forecasting process with less repetitive effort and a clearer view of where the numbers deserve confidence.
About the interview
This article is based on Nomentia's interview with Johannes Pöschl, Data Scientist at Nomentia. Pöschl has spent seven years developing predictive forecasting capabilities and helping establish Nomentia's AI function, with a focus on data science and AI for financial forecasting.
