Predictive & Prescriptive Analytics – From Forecast to Recommended Action

Series: AI in Treasury | Article 3 of 7

Subtitle: 90% forecast accuracy – AI as a strategic sparring partner in treasury

Category: Data & decision support · Website version

Generative AI describes. Agentic AI orchestrates. Value creation in treasury happens where numbers become decisions: in forecasting and the resulting recommended action. Predictive analytics answers what is likely to happen next. Prescriptive analytics assesses, under uncertainty, which measure is better – including probabilities, alternatives and trade-offs.

This third article in the “AI in Treasury” series continues from the technology and workflow perspective of Article 2. The focus is on ML-based liquidity forecasts, anomaly detection and prescriptive optimization recommendations – and what they mean in practice for treasurers, CFOs and analysts. The guiding question: When is a forecast good enough to act on – and how does the human remain in control?

Benchmark first: what “90% accuracy” really means

In its analysis of AI-powered cash management, Capgemini notes that treasury functions using AI and real-time data can achieve forecast accuracy of up to 90 percent. Classic, fragmented processes often come with variance ranges above 20 percent and oversized liquidity buffers of 15 to 20 percent. What matters is the documented performance gap between Excel-driven models and data-driven ML approaches.

“Up to 90 percent” is not a guarantee for every company and every horizon. Achievable accuracy depends on data history, transaction volumes, seasonality, group structure and master-data quality. The order of magnitude is still enough to challenge the status quo: many treasuries still operate with forecasts in the 60 to 70 percent range and compensate for uncertainty with expensive buffers, short-term credit lines and reactive decisions.

The business consequence is immediate. Improving the forecast systematically reduces idle cash, lowers funding costs, surfaces shortfalls earlier and supports investment and hedging decisions with tighter confidence intervals. Predictive analytics therefore sits at the core of decision quality in treasury.

Predictive analytics: more precise forecasts – and why Excel hits structural limits

Classic Excel forecasts have grown historically in many organizations: bottom-up submissions from entities, manual aggregation, blanket uplifts, subjective expert estimates. In stable environments this works reasonably well. As volatility, external shocks or complex payment-behavior patterns increase, Excel hits structural limits. The models scale poorly, underestimate interdependencies and learn seasonal, calendar and behavioral patterns only to a limited extent.

Machine-learning models extract recurring patterns from historical payment data – seasonality, weekday and holiday effects, payment delays for specific customer groups, correlated outflows after production cycles – and combine them with external signals such as FX rates, interest rates or supply-chain indicators. What makes the difference is the systematic selection and calibration of suitable methods.

In practice, an ensemble of model families is used: classic time-series models (for example SARIMA/TBATS), structured Bayesian approaches, gradient boosting (e.g. XGBoost) or neural networks. Modern treasury platforms partly select models automatically, train them on their own histories and continuously compare forecast and actuals. The feedback loop of forecast versus actual, variance analysis and recalibration turns a one-off projection into a learning system.

ML liquidity forecasts: from position to driver analysis

The usual entry point is the cash-flow forecast on short and medium horizons – from the daily and weekly view through to the classic 13-week outlook. ML models use aggregated balances together with transaction and ledger data from ERP, banks and TMS. The approach becomes particularly effective when AR, AP and payroll patterns are modeled explicitly: When does which debtor typically pay? Which suppliers pull invoices forward? Which outflows are contractually fixed, and which behave stochastically?

The quality difference shows in hit rate and driver explainability at the same time. A good forecast reports “EUR 4 million short in three weeks” and makes visible whether delayed customer inflows, accelerated supplier payments, tax specials or FX effects sit behind it. For CFOs and analysts, that driver transparency is the real value: the forecast becomes a steering foundation.

Without a robust data history and consistent categorization, every model remains weak. Fragmented chart of accounts structures, inconsistent cash-flow codes and missing actuals reconciliations create systematic blind spots. Predictive analytics rewards discipline in data work – a point successful implementations confirm again and again.

Exogenous factors: why internal history alone is often not enough

Many forecasting approaches train primarily on what already sits in the organization’s own systems: payments, balances, AR/AP behavior. That is necessary – and in calm phases often sufficient. In volatile markets, however, it also matters how early external impulses become visible: rate turns, inflation surges, commodity prices, demand indicators, trade or supply-chain shifts. Anyone who only notices these signals via delayed actuals in the ERP reacts late.

This is the strategic complement to classic transaction AI: driver- and signal-based planning that consciously embeds exogenous factors in liquidity and cash-flow models – and links scenarios to macroeconomic reality rather than only replaying them internally.

Early risk detection and anomaly detection

Beyond classic liquidity forecasting, predictive analytics also supports early risk detection. Models assess counterparty patterns, detect changes in payment behavior and simulate interest-rate scenarios or FX volatility based on historical and market-side signals. The value lies in early marking of deviations from the expected path – even where absolute shock prediction remains limited.

Anomaly detection in payments is a particularly practical use case. Systems learn what is “normal” for amounts, recipients, timing and approval patterns, and flag deviations before money leaves the house: unusually large payments, changed vendor bank details, atypical approval paths. Coupa and other platforms emphasize this behavioral approach; FIS Neural Treasury connects liquidity insights with threat monitoring across payments and networks. Fraud prevention and forecast quality grow from the same capability: pattern recognition under uncertainty.

Human approval remains essential. Anomaly flags create a need for decision, not autopilot release. The better the false-positive rate is controlled, the higher the acceptance in day-to-day operations – and the more AI relieves rather than generates alert noise.

The proven sequence: Reconciliation → Visibility → Forecasting → Prescription

Successful teams rarely start with the most complex use case. The sequence proven in practice builds capabilities step by step: first reconciliation and data integration, then cash visibility across accounts and entities, then ML-supported forecasting – and only afterwards prescriptive actions. Each step improves the data foundation for the next. Anyone who starts forecasting without clean actuals and integrated positioning trains models on noise.

This sequence is also organizationally smart. Early quick wins create trust. An improved reconciliation process or a consolidated liquidity view is tangible and politically easier to push than an immediate jump into autonomous hedging recommendations. Predictive analytics therefore follows a maturity path rather than a big-bang transformation.

Prescriptive analytics: from “what’s coming?” to “what should we do?”

The qualitative leap from predictive to prescriptive is decisive. Predictive says: “Cash shortfall in three weeks.” Prescriptive adds: “Activate the credit line with Bank X – or fund via an intercompany loan from entity Y; estimated probability of securing liquidity: 87 percent; opportunity cost of the alternatives: …”

This form of decision support changes the treasurer’s role. Instead of building scenarios manually for hours, the system evaluates options along predefined constraints: available lines, intercompany rules, FX limits, covenant thresholds, interest differentials, settlement times. The human remains the decision-maker – on a clearly richer option landscape.

Typical prescriptive fields in treasury include:

Optimizing payment timing: when to execute to capture FX advantages, cut-offs or working-capital targets
Dynamic hedging recommendations: adjusting FX hedges to volatility and risk profile – as a proposal with approval, without blind automation
Funding structure: guidance on tenors, debt/equity mix or internal versus external funding under liquidity and cost constraints
Cash deployment: where to park, pool or allocate idle cash internally – including expected interest income and availability risk

Prescriptive analytics is the logical continuation of a good forecast. Without a robust projection, every recommendation remains speculation. With a robust projection, the recommendation becomes a controllable decision product.

Real-time scenario simulation: what-if without an Excel marathon

A closely related benefit area is scenario simulation. M&A payments, currency crises, rate shocks or supply-chain disruptions can be run through in minutes rather than days in modern environments – across multiple horizons and entity structures. The value sits in iteration speed: treasurers can change assumptions, see sensitivities and make decision windows transparent to the CFO and board.

In addition, some solutions analyze bank relationships and pricing historically: Which bank performs on which product? Where do fee clusters arise? Where does concentration or diversification pay off? This is systematic evaluation of large transaction histories – and exactly that evaluation was often uneconomical when done manually.

Who offers this today? Vendor landscape 2025/2026

The market for predictive and prescriptive capabilities in treasury is broader and more mature than many decision-makers assume. It can roughly be grouped into five categories: specialized European and global TMS/cash platforms, forecasting specialists (partly acquired), ERP-native assistants, enterprise planning/FP&A platforms with exogenous signals, and bank and consulting offerings.

TMS and cash platforms with native AI

In April 2025, Nomentia significantly expanded its AI cash-flow forecasting. The solution focuses on automated data connectivity, anomaly correction, seasonality detection and a range of forecasting models – including Bayesian structural time series, TBATS, SARIMA, Prophet, XGBoost as well as neural and rule-based approaches. Models are selected and trained on historical data; forecast and actuals are compared continuously. Nomentia positions the offering explicitly for treasurers and reports measurable accuracy gains with substantial time savings from customer projects – for example with Karl Mayer. For European mid- and large-cap groups, this is a relevant, practical provider at the forecasting core.

Kyriba combines Advanced Liquidity Planning with embedded machine learning for trends, seasonality and calendar effects – with a simultaneous focus on explainability, governance and auditability. The ambition is a controllable planning framework across horizons, including scenarios and plan-to-cash logic.

FIS Neural Treasury addresses liquidity insights and flexible cash forecasting as well as anomaly/threat monitoring in payments. Generative support (Treasury GPT) for setup and usage is added – an indication that predictive capabilities are increasingly embedded in broader AI suites.

Coupa Treasury (on a BELLIN foundation) uses AI for liquidity forecasts from historical patterns and obligations as well as anomaly detection in payments and approvals. The tight link between spend, payment and treasury data is the structural advantage here.

Ripple Treasury (formerly GTreasury) has strongly expanded forecasting and variance analysis with the acquisition of CashAnalytics and the GSmart stack: GSmart Ledger learns from invoice and ledger patterns; Forecast Insights compares forecast and actuals, flags anomalies and delivers actionable explanations. Vendors communicate accuracy improvements in the order of more than 30 percent – depending on starting level and data quality.

TIS integrates cash forecasting and working-capital analytics – historically also via Cashforce – and now emphasizes predictive cash forecasting with continuous recalibration as well as explainable AI assistants for treasury-specific questions. The focus sits strongly on payments, bank connectivity and cash data as the training base.

SAP increasingly addresses the topic through the Joule Cash and Treasury assistant and specialized agents: predictive forecasting based on actual payment behavior, cash-positioning agents for early detection of funding gaps and concentration risks, and support on allocation and financing questions. For SAP-centric landscapes, this is a natural integration path.

Forecasting specialists and horizontal finance AI

HighRadius positions an agentic-AI-based cash-forecasting solution with a claim of very high inflow accuracy (vendor-stated up to 95 percent) and specialized agents for data intake, modeling, variance analysis and overrides. The approach is an AI-native finance platform with a strong treasury/working-capital focus – beyond the classic TMS cut.

Specialized forecasting layers also remain relevant – either standalone or, after acquisition, integrated into larger suites. The consolidation wave (CashAnalytics into Ripple Treasury/GTreasury, Cashforce in the TIS/banking environment) shows: pure point-tool forecasting is increasingly embedded in broader treasury ecosystems.

Enterprise planning with exogenous signals: Board

Board (Board International) sets a distinct accent – more an enterprise planning/FP&A platform than a classic TMS, yet clearly relevant for cash and liquidity planning. Through Board Foresight and integrated external signals, exogenous factors can be systematically embedded in forecasts and scenarios: macroeconomic indicators, interest rates, inflation, commodity prices, trade and demand impulses. The value is that liquidity planning does not only learn from internal transaction history, but also models drivers outside the organization’s own books – and makes deviations versus those external signals explainable. Board thereby closes a gap that many TMS-native forecasting modules cover only to a limited extent.

Banks and consulting: AI-as-a-service and transformation

JPMorgan offers Cash Flow Intelligence as an AI-as-a-service proposition around pattern recognition, forecasting and liquidity visibility – fed by payment and transaction data. Reports such as Bloomberg’s on a tangible reduction of manual cash-flow work (in the order of around 90 percent in selected customer workflows) underline the point: banks now treat forecasting as a service product.

Capgemini and KPMG act primarily as transformation and implementation partners: from maturity assessment via data harmonization through to the introduction of predictive models and change management. Capgemini’s “up to 90 percent” benchmark and KPMG’s focus on predictive cash visibility frame the business case. The product decision for TMS, bank platform, planning platform or specialist solution remains separate.

What the vendor landscape means for decision-makers

Predictive and prescriptive capabilities are no longer laboratory topics in 2026. Differences between offerings sit in the data model, explainability, integration path and governance – far beyond the marketing buzzword “AI”. Anyone selecting should at least check:

Which data sources are natively connected (bank, ERP, AR/AP, TMS) – and how high is the integration effort?
Are exogenous factors (macro, rates, commodities, demand impulses) systematically included – or only internal history?
Which model families are used, and how transparent are forecast drivers and variances?
Is there a productive forecast-versus-actuals loop with continuous learning?
Where does prediction end and prescription begin – and which approval gates are in place?
How audit-proof are assumptions, overrides and recommendations documented?

In this field, Nomentia clearly belongs among the providers that treat forecasting as a core competence. Board complements the spectrum where FP&A and treasury need to steer exogenous drivers together. Kyriba, FIS, Coupa, Ripple Treasury, TIS, SAP, HighRadius and bank offerings such as JPMorgan CFI together form a market in which build-versus-buy and TMS-native versus best-of-breed must be reassessed.

What this changes for treasury decisions

For numbers-oriented professionals, the decision rhythm changes. In place of periodic, often already outdated forecast packages comes a continuous steering loop: signal → forecast → options → decision → learning outcome. CFOs receive tighter uncertainty bands. Analysts shift time from data gathering to variance interpretation. Treasurers develop from Excel operators into portfolio managers of liquidity risks and action options.

At the same time, demands on judgment rise. A recommendation with an “87 percent success probability” is useful only when it is clear what the probability refers to, which assumptions apply and which tail risks remain out of scope. Explainable AI and clear decision rights are prerequisites for acceptance.

Limits and prerequisites – without illusions

Predictive and prescriptive analytics rarely fail due to missing algorithms. They fail due to weak data, unclear accountabilities and unrealistic expectations. Models need history, clean categories and stable as-is processes. Organizations need override rules, escalation paths and the willingness to actively steer forecast quality.

From a regulatory and governance perspective as well: the closer recommendations move toward automated execution, the more important audit trails, model monitoring and human approval become. Prescriptive analytics may increase decision capability and must preserve accountability.

Conclusion and outlook to Article 4

Predictive analytics makes uncertainty more measurable. Prescriptive analytics makes options decidable. Together they shift treasury from retrospective control to forward-looking steering – provided data quality, model transparency and human judgment grow with them. Documented practice in leading organizations shows: accuracy levels toward 90 percent are achievable. The path there is a maturity journey with clear intermediate steps.

The next article in the series covers the convergence of TradFi and DeFi as an ongoing transformation with concrete practice examples. Because anyone who steers liquidity and risk on data will sooner or later also have to assess the market infrastructure on which that liquidity is moved and tokenized.

Whether AI can forecast is now answered. What remains open is whether treasury consistently turns better forecasts into better decisions.

Series: AI in Treasury | Article 3 of 7 | For: numbers-oriented treasury professionals, CFOs, analysts