Every treasury platform now claims AI: real-time forecasts, smarter liquidity modeling, automation that takes work off your plate. On the surface it all sounds impressive, and much of it does not hold up — generic models, little auditability, and a quiet trade of control for convenience. For a CFO, the question is no longer whether to adopt AI in treasury but how to tell real capability from packaged noise. Here is the evaluation framework.
TL;DR
AI treasury management applies artificial intelligence across the treasury function: cash flow forecasting, liquidity management, payments, FX and risk, bank connectivity, fraud detection, and reconciliation. By 2026 nearly every treasury platform markets AI capabilities, so the CFO’s task is evaluation — separating solutions built for the work from generic models wrapped in AI language.
CFOs should evaluate AI treasury solutions across seven criteria. Scope and fit: does the AI address your actual treasury priorities (often the operational core of reconciliation, payment processing, and forecasting) rather than a flashy but narrow feature. Forecasting quality: not just the headline accuracy claim but how the AI handles the data quality that actually gates forecast accuracy. Data foundation: whether the solution produces or depends on clean, current, reconciled data, since treasury AI is only as good as the data beneath it. Bank and ERP connectivity: real-time visibility across all banks and entities. Fraud and risk: AI-based anomaly and fraud detection, increasingly essential. Transparency and auditability: whether the AI’s decisions can be explained and reconstructed, now a regulatory expectation. And control and governance: whether the CFO retains control and can govern the AI to the new standards.
The governance bar rose sharply in 2026. In March 2026 the US Treasury released the Financial Services AI Risk Management Framework, an operational framework with 230 control objectives across the AI lifecycle, adapted specifically to treasury, payments, fraud, and risk. For any organization using AI in treasury, this is now a reference point against which governance is measured, which makes transparency, auditability, and control evaluation criteria rather than nice-to-haves.
The throughline: the best AI treasury solutions are purpose-built, transparent, auditable, and keep the CFO in control, and crucially they rest on a clean data foundation, because the most sophisticated treasury AI still depends on the quality of the data it runs on. For the forecasting-tools comparison specifically, see The Top AI Cash Flow Forecasting Tools for Treasury Teams.
What AI treasury management actually covers
Treasury is broader than cash forecasting, though forecasting gets the most attention. A full AI treasury evaluation spans the whole function:
Cash flow forecasting predicts future cash positions across horizons (commonly 4-week, 13-week, and 26-week) so the CFO knows whether the business will have the liquidity to fund operations, service debt covenants, and execute on growth. Liquidity management optimizes cash positioning across accounts, banks, and entities. Payments execution moves money, increasingly through a centralized payment factory, with controls and fraud screening. FX and risk management hedges currency exposure and models financial risk. Bank connectivity aggregates balances and transactions across many banks (the largest platforms connect to thousands) and ERPs into a single real-time view. Fraud detection uses AI to catch anomalous transactions before they settle. And reconciliation matches and clears transactions so the cash position is accurate.
AI now touches all of these, and the depth varies enormously by platform and by capability. A CFO evaluating “AI treasury management” is really evaluating AI across this whole scope, which is why a single accuracy claim or a single flashy feature is a poor basis for a decision. The evaluation has to span the function and, underneath it, the data foundation everything depends on.
The governance shift CFOs cannot ignore
Before the criteria, one development reshapes all of them. In March 2026 the US Department of the Treasury released the Financial Services AI Risk Management Framework (FS AI RMF), alongside an AI Lexicon. It is not a policy statement but an operational framework, with a matrix of 230 control objectives mapped across the full AI lifecycle, adapting the NIST AI Risk Management Framework to the specific realities of treasury, payments, fraud detection, and risk management.
The significance for a CFO is direct: if your organization uses AI anywhere in the finance function, this framework is now a reference point against which your AI governance will be measured. That changes evaluation. Capabilities that might once have been treated as optional — explainability of AI decisions, auditability of AI outputs, control over what the AI does autonomously — are now governance expectations. An AI treasury solution that cannot be explained, audited, or controlled is not just operationally risky; it is increasingly out of step with the governance standard regulators and auditors are converging on.
This is why the criteria below weight transparency, auditability, and control heavily. The era when “AI-powered” was a selling point on its own is over; in 2026 the AI also has to be governable, and the CFO is accountable for governing it. For the broader documentation standard, see AI Audit Trail Requirements: A 2026 Checklist.
The seven criteria CFOs should evaluate
1. Scope and fit: is it built for your actual priorities?
The first question is whether the AI addresses the work you actually need done, not whether it has impressive features. Survey data suggests the highest-ROI starting points are often the operational core — bank reconciliation, payment processing, and transaction categorization — which are high-volume, rules-based processes where AI delivers relatively rapid and measurable return, building the organizational confidence to tackle more complex applications like forecasting and risk.
How to evaluate: identify where your treasury actually loses time and carries risk, then assess whether the AI is built for that, rather than being dazzled by a capability you will rarely use. A solution purpose-built for your priorities beats a broader one whose AI is generic. As one treasury-AI evaluation framework puts it, the question is not whether AI is coming but whether the solution being pitched is built for the work you actually do.
2. Forecasting quality: beyond the headline accuracy claim
Forecasting is treasury’s most consequential AI application, and the most over-claimed. Vendors cite high accuracy figures, but the number alone is misleading.
How to evaluate: probe how the AI achieves accuracy and, critically, how it handles data quality, because forecast accuracy is gated more by the quality of the underlying data than by the sophistication of the model. Ask how the solution performs on your messy, real-world data rather than clean demo data, how it handles the unapplied AR and inconsistent actuals that degrade real forecasts, and how it performs across multiple horizons rather than a single best-case number. A forecasting AI that assumes clean data is solving the easy version of the problem. This is covered in depth in The Top AI Cash Flow Forecasting Tools for Treasury Teams.
3. Data foundation: the criterion most CFOs underweight
This is the criterion that most determines whether treasury AI delivers, and the one most often overlooked. Every treasury AI capability — forecasting, liquidity optimization, fraud detection — runs on financial data, and its output is only as good as that data. Stale actuals, unapplied receivables, and balances that disagree across systems degrade every AI output downstream.
How to evaluate: assess not just the AI but the data layer feeding it. Does the solution produce or depend on clean, current, reconciled data? How does it handle the reconciliation and actuals work that determines data quality? A CFO who evaluates the AI capability without evaluating the data foundation beneath it is evaluating half the system, and usually the easier half. The teams that get the most from treasury AI treat the data-and-execution layer as seriously as the analytics on top of it, because that layer is where accuracy is actually decided. See also The Best AI Reconciliation Software for Mid-Market Finance Teams.
4. Bank and ERP connectivity: real-time visibility
Treasury AI needs current data from across the banking and ERP landscape to be useful. Real-time visibility into cash across all accounts, banks, and entities is foundational, and the leading platforms connect to thousands of banks (some to nearly ten thousand) plus the major ERPs.
How to evaluate: assess the breadth and the latency of connectivity, whether it spans all your banks and entities, and whether the data is genuinely real-time or batch. For organizations with many banking relationships across entities, this connectivity layer is what consolidates the group cash position into a single view, which is especially valuable for multi-entity and PE-backed structures. Weak connectivity undermines every AI capability above it, because they all depend on current, complete data.
5. Fraud and risk: AI-based protection
Payment fraud and financial risk are growing threats, and AI-based fraud detection is increasingly essential rather than optional. The leading platforms include AI and machine-learning-based fraud detection that flags anomalous transactions before they settle.
How to evaluate: assess the AI fraud and anomaly detection capability — how it learns your normal patterns, how it flags exceptions, and how it integrates with payment controls. Given that the new US Treasury framework specifically addresses fraud detection, the quality and governability of this capability is both an operational and a compliance consideration.
6. Transparency and auditability: can you explain its decisions?
This criterion rose from optional to essential in 2026. Many AI treasury solutions are built on generic models that offer little auditability and operate as black boxes. Under the new governance standard, that is a problem.
How to evaluate: assess whether the AI’s decisions can be explained and reconstructed. When the AI makes a forecast, flags a transaction, or moves money, can you see why? Can you reconstruct the decision for an auditor? Solutions built on transparent, explainable approaches meet the governance bar; opaque generic models increasingly do not. This is where the architecture of the AI matters: deterministic, explainable systems whose decisions can be traced and reconstructed are auditable by design, while probabilistic black-box models are not. For treasury feeding audited financial reporting, auditability is now a requirement. (On why a confidence score is not an audit trail, see When Confidence Scores Lie.)
7. Control and governance: do you stay in charge?
The final criterion is whether the CFO retains control. Many AI solutions ask teams to give up control in exchange for convenience — autonomous actions you cannot easily constrain or oversee. Under the FS AI RMF’s 230 control objectives, retaining governable control is now an expectation.
How to evaluate: assess whether you can define what the AI does autonomously versus what requires human approval, whether you can constrain and oversee it, and whether the solution supports the controls the governance framework expects. The goal is AI that augments the treasury team under the CFO’s control, not autonomous action the CFO cannot govern. The best solutions keep the human in charge and the AI accountable.
How the criteria fit together
These seven criteria are not a checklist to score equally; they are layered. Connectivity and the data foundation (criteria 3 and 4) are the base, because every AI capability depends on current, clean, complete data. The functional capabilities (forecasting, fraud, and the operational core — criteria 1, 2, and 5) sit on that base. And transparency, auditability, and control (criteria 6 and 7) wrap all of it, because under the 2026 governance standard, capability that cannot be explained, audited, or controlled is a liability regardless of how good it is.
A useful way for a CFO to apply this: start at the base. If the connectivity and data foundation are weak, the AI capabilities on top will underdeliver no matter how impressive they look in a demo, because they will be running on incomplete or stale data. If the foundation is solid, evaluate the functional capabilities against your actual priorities. And throughout, treat transparency, auditability, and control as gating requirements rather than bonus features, because the governance standard now treats them that way.
This layered view also explains the most common treasury AI disappointment: a CFO buys an impressive forecasting or analytics capability, deploys it on a weak data foundation, and gets mediocre results, then concludes the AI failed when the data layer was the actual problem. Evaluating the foundation first prevents that. For the broader question of how to measure return on these investments, see The CFO’s Guide to Measuring ROI on Finance AI.
Where Kognitos fits in a treasury AI evaluation
A note on scope, in the honest spirit this framework demands. Kognitos is not a treasury management system. It does not forecast cash, manage liquidity, execute payments, or run FX, and it does not replace Kyriba, GTreasury, HighRadius, or the other TMS and forecasting platforms a treasury function runs on. For those capabilities, a CFO should evaluate the treasury platforms directly.
Where Kognitos is relevant to a treasury AI evaluation is the data foundation (criterion 3) and the transparency, auditability, and control criteria (6 and 7). Kognitos is a deterministic, neurosymbolic, agentic platform that produces clean, current, reconciled data — the cash application, reconciliation, and cross-system consolidation that determine whether the treasury AI on top has good data to work with. And because it executes deterministically in plain English (English as Code) with every decision logged and explained, it aligns with the auditability and control standards the new governance framework expects.
In other words, Kognitos is relevant to the layer beneath the treasury platform — the data foundation that most determines whether treasury AI delivers — and to the governability standard, rather than as a treasury system itself. For many treasury teams, that data foundation is the underweighted criterion that decides whether their treasury AI investment pays off. The point is not to add another treasury tool; it is that evaluating the data layer is part of evaluating treasury AI, and it is the part most often missed. Related: The Top AI Tools for Accounts Receivable Automation and Cash Application and Finance & Accounting Automation Solutions.
Book a working session with a Kognitos solutions engineer → Try Kognitos free →