Liquidity Planning with AI – Best Practice Approaches and What Is Truly Different Today

In December, we took a close look at the main challenges treasury departments face when planning liquidity requirements: fragmented data landscapes, predominantly manual processes, limited transparency and the pronounced volatility of central influencing factors shape day-to-day operations. These issues have persisted for years and continue to accompany many treasury organizations.

In this article, we shift the focus from “what should we do?” to “how can we do it?”: Specifically: How can innovative artificial intelligence approaches effectively address these challenges today? How do modern methods differ from classical forecasting techniques and where do clear limitations remain? Since the topic has been present for years, the value today lies less in the technology itself and more in how it is implemented.

The foundation of any AI‑supported liquidity planning process is still structured financial data – such as historical cashflows grouped by planning categories, for example customer inflows or supplier‑related outflows for operating materials – sourced from ERP and treasury systems. These data sets are then purposefully enriched with external structured or unstructured information. Examples include macroeconomic publications, market and interest rate data, or text‑based early indicators. Thanks to advances in automated text and signal analysis, such information can now be converted into quantifiable drivers and integrated into forecasting models in a controlled way. The goal is not to expand the data universe indiscriminately but to select the drivers that truly matter, avoiding the introduction of unnecessary volatility into the forecast. Today, this selection is automated using high‑performance AI models.

With the expanded data foundation, data cleansing and preparation methods have also evolved. While manual plausibility checks once dominated, modern approaches combine rule‑based logic with statistical and machine learning techniques. Outliers, anomalies, or structural breaks can be identified and assessed automatically. The focus therefore shifts from one‑off cleansing to continuous assurance of forecast quality, where changes in data structures or business models are addressed through model adjustments or re‑training.

The biggest progress today lies less in new data sources and more in the type of modeling. Established AI methods such as gradient boosting models and newer deep‑learning time‑series architectures capture nonlinear relationships and complex temporal dependencies far more robustly than classical regression or history‑based forecasts. Complementary probabilistic forecasting techniques make it possible to represent ranges and scenarios instead of single‑point estimates – improving risk management. Another major development is the use of explainable models. Explainable AI techniques show which variables influence the forecasts and to what extent. This increases transparency, supports governance requirements, and simplifies business validation in the treasury context, which is a clear evolution from earlier black‑box approaches.

At the methodological level, no single approach is universally superior. Instead, different model classes are used in parallel, objectively validated and compared. Alongside classical time-series methods such as ARIMA or exponential smoothing and established machine-learning models, deep-learning solutions are increasingly applied because they capture complex interactions, non-linear effects and temporal dynamics more effectively. The substantial performance gains of modern models stem not from individual algorithms but from the combination of greater computing power, better data availability, automated model selection, continuous re-training and advanced feature engineering. As a result, the focus shifts from building a model once to managing a continuous and adaptive forecasting process.

Another key difference from earlier planning methods lies in how uncertainty is handled. AI-based forecasts are no longer treated as static point estimates but as a starting point for systematic scenario and sensitivity analyses. By varying key drivers, alternative developments can be simulated and their impact on liquidity made transparent. AI does not replace scenario thinking but makes it more scalable, reproducible and data-driven. Decision spaces become clearer, as does the sensitivity of forecasts to individual drivers.

The underlying data modeling and architecture of modern solutions increasingly follow modular principles. Data, features, model and decision layers are clearly separated, which improves scalability, improves transparency and simplifies integration into existing treasury processes, for example when feeding forecasts into financial risk management for foreign exchange, interest rates or counterparty exposures. Forecasting models therefore evolve from isolated calculation logics into steerable and verifiable decision tools. This not only streamlines the ongoing development of individual models but also supports their integration into operational decision processes.

Despite these advances, AI is not a cure-all. Highly irregular cashflows, one-time special effects or abrupt structural shifts remain difficult to forecast with precision. More complex models can also reduce transparency. Especially in the treasury environment, professional plausibility checks, governance and human responsibility remain indispensable. AI supports decision-making but does not make the decisions.

In light of these developments, KPMG has invested in its own solution for AI-based liquidity forecasting that is fully aligned with the principles outlined above. The focus is on a modular data architecture, the parallel use of different forecasting methods, continuous monitoring of model quality and systematic integration of scenario and sensitivity analyses. The goal is to embed modern forecasting approaches into existing treasury environments in a practical and controlled way, with a clear understanding that AI supports decision-making but does not replace it.

In practice treasury leaders must determine how much control should shift toward probabilistic accuracy and process efficiency. The earlier AI is tested and evaluated under real conditions, the more confidence can be built and the easier it becomes to define suitable application areas. This also requires a shift in mindset within finance and treasury management.