AI for treasury and cash management: where it earns its keep, where it doesn't
Published 30 April 2026
I have run multi-currency operations across more than twenty-five countries as a founder-CEO, and I have built and rebuilt the treasury layers of two international businesses. The treasury function in a mid-market business is rarely the most glamorous part of the operation, and it is almost always one of the highest-leverage. That is the lens I want to bring to the question of where AI in treasury actually earns its keep in 2026.
The honest position: treasury is one of the most natural homes for AI in the finance function, and the median deployment so far is shallow. This post is what is real, what is vendor marketing, and what a fractional CFO advising an Irish or UK mid-market business should be doing about it now.
What the data says about adoption
Three reference points worth knowing.
Citi Institute’s October 2025 GenAI in Treasury research: 82% of treasury teams are still experimenting with generative AI. 3% have scaled it across operations. (Source: Citi Institute, October 2025.) The gap between experimentation and scale is the largest in finance.
Deloitte’s 2024 Global Corporate Treasury Survey (200+ companies): approximately 50% of treasurers prioritise improving cash flow forecasting. Approximately 20% rate their current forecasting capability as above average. (Source: Deloitte 2024 Global Corporate Treasury Survey.) The demand-side intent is there. The supply-side capability is not.
AFP 2025 Cash Forecasting Survey: 59% of treasury teams cite data quality and availability as the primary forecasting challenge. 18% cite technology. (Source: AFP benchmarking.) AI is not the bottleneck. The data layer underneath the AI is the bottleneck. The CFO who installs an AI cash forecasting tool on top of bank-feed data that is three to seven days old has not improved the forecast.
Where AI is genuinely earning its keep
Four areas, in approximate order of maturity.
Cash flow forecasting. Vendor accuracy claims cluster at 88-95%. HighRadius claims 95% at a twelve-month horizon. Tesorio claims 95%+ at a thirteen-week horizon. Kyriba claims 90%. Industry commentary puts AI-assisted thirteen-week forecasts at 88-92% accuracy against a manual baseline of around 60%. (Sources: vendor product pages; AFP-derived industry commentary.) Two caveats. First, every number above is vendor-claimed and not independently audited. Second, accuracy at the thirteen-week horizon for a mid-market business with disciplined collections is materially different from accuracy at the same horizon for a business whose AR ageing is structurally rough. The tool helps. It does not substitute for the discipline.
The wave of AI cash forecasting tools that landed in 2025 is real. U.S. Bank’s AI cash forecasting tool, powered by Kyriba, launched in 2025. (Source: Finovate coverage.) Trovata raised $9m in July 2025 and shipped Trovata AI 2.0, with agentic features for variance explanation, anomaly detection, and reconciliation summaries. (Source: fintech.global, July 2025.) FIS launched Quantum Cloud Edition in April 2025 with embedded AI and a Treasury GPT assistant. (Source: BusinessWire, 29 April 2025.) The vendor side has shipped. The deployment side has not yet caught up.
Bank reconciliation and anomaly detection. Modern Treasury claims 90 to 100% item-level reconciliation rates on its AI-assisted bank reconciliation product, built on a combination of heuristics, deterministic algorithms, and LLMs, with human review by design. (Source: Modern Treasury announcement.) That is the kind of claim I take more seriously than the cash forecasting numbers, because bank reconciliation is a more bounded problem with a clearer ground truth. The reconciliation work that used to be a daily two-hour task is now a daily fifteen-minute review of exceptions. That is a real productivity gain for a small treasury team.
Payment fraud detection. The AFP 2025 Payments Fraud and Control Survey found 79% of organisations were hit by payments fraud in 2024. Recovery rates collapsed to 22% recovering more than 75% of funds lost, down from 41% the prior year. (Source: AFP 2025 via Truist.) The FBI’s IC3 attributed $893 million of 2025 losses specifically to AI-enabled crime. (Source: Nacha summary of FBI IC3.) Behavioural biometrics vendors (BioCatch, NICE Actimize, Trustpair) market AI as the answer. The honest position: AI in fraud defence is necessary because AI in fraud offence is already deployed. The arms race is real and the published false-positive-rate benchmarks are vendor-claimed.
FX and exposure management. Bank of England data from the BIS Triennial Survey, April 2025, puts the UK FX market at $4.745 trillion daily turnover. (Source: Bank of England, BIS Triennial Survey.) Citi has reported using generative AI to detect FX trading fraud. Kyriba’s FX Cash Flow module forecasts currency exposure pre-hedge. The Bank of England’s Financial Stability in Focus report, April 2025, raises concentration risk and model risk as systemic concerns for AI in financial markets. (Source: BoE FSiF April 2025.) For a mid-market multi-currency business, the value of AI in FX is in exposure forecasting, not in trading. The trading is still better done by a banker who knows the market.
What is changing underneath the AI layer
The plumbing under treasury changed substantially in late 2025 and continues to change in 2026, and AI’s value depends on the plumbing.
ISO 20022 coexistence ended 22 November 2025 for SWIFT cross-border payments. (Source: Kyriba migration FAQ.) The richer remittance data in ISO 20022 messages is what makes downstream AI reconciliation work better. SWIFT itself has built an AI address-structuring model to help banks meet structured-address requirements. (Source: Trade Treasury Payments coverage.)
SEPA Instant became mandatory across the EU in October 2025. Ten-second settlement, 24/7, at pricing parity with standard SEPA Credit Transfer. (Source: Treasury Management International.)
FedNow’s transaction limit increased from $1 million to $10 million in November 2025. (Source: Federal Reserve Financial Services.)
The treasury function’s operating cadence is shifting from batch to continuous. Overnight liquidity buffers are smaller. Fraud exposure is sharper. AI is one of the levers that lets a small treasury team handle 24/7 payment flows without quadrupling headcount. The plumbing change is the precondition. The AI is the response.
What does not change
Two things, both worth being explicit about.
The data layer is still the bottleneck. The AFP 59% figure on data quality is the single most important number in this post. A treasury team running AI cash forecasting on bank-feed data that is several days old will get plausible-looking forecasts that are systematically wrong. The work of building real-time bank connectivity, clean accounts payable and receivable data, and a chart of accounts the AI tool can read is the work that determines whether the AI tool helps or hurts. The data quality post is the long version. The treasury-specific version: get the bank feeds and the AR/AP cadence right, then layer the AI on top.
Hedge effectiveness testing under IFRS 9 still requires an audit trail. The treasury function that lets AI choose hedge ratios without documented human review of the methodology is the function whose audit committee asks an uncomfortable question. AI can support the analysis. The hedge designation decision is still a human one, and the documentation has to show that it was.
What a fractional CFO should advise this quarter
Three concrete recommendations.
For businesses at €5m to €20m turnover. A rolling thirteen-week cash forecast, maintained weekly with discipline, is worth more than an AI cash forecasting tool deployed on rough data. NetSuite OneWorld with the cash management module, or Sage Intacct with its February 2025 AI predictive analytics addition (source: Big Bang 360 product comparison), will cover most of the function’s needs. Wise Business or Airwallex for multi-currency holding and conversion. Airwallex Copilot’s AI automation layer is worth piloting if multi-currency is core to the business.
For businesses at €20m to €100m turnover. A dedicated treasury management system is now the right answer. Kyriba, FIS Quantum, GTreasury, ION are the enterprise options. Trovata and Modern Treasury are the agile alternatives. The vendor selection should not be driven by the AI features. It should be driven by the bank connectivity, the ERP integration, and the reporting your audit committee actually wants. The AI features are now a baseline expectation, not a differentiator.
Across the range. Adopt the payment fraud defences seriously. Trustpair for supplier verification, behavioural biometrics for high-value approvals, multi-factor authentication on payment release. The AI in fraud offence has gotten better in the last twelve months. The fraud defences your business put in place in 2022 are not adequate to the 2026 threat landscape.
Where this lands
Treasury is one of the more legible places for AI in finance, partly because the work has historically been heavy on the data and process layers that AI most reliably amplifies, and partly because the failure modes are bounded enough to be governable. The vendor market has caught up to that potential. The deployment side, on the evidence, has not.
The treasury function that adopts thoughtfully will reduce the cost of running treasury, reduce the cost of fraud, and improve the quality of the forecast the board uses to make decisions. The treasury function that adopts because the vendor promised 95% forecast accuracy will discover that the 95% claim referred to a different business with different data, and the forecast variance the board sees will not be lower than it was in 2024.
The honest line for a fractional CFO is the one most boards already know. Discipline first. Data second. AI third. In that order, the tools work. In the wrong order, the tools amplify the problem they were supposed to solve.
Maebh Collins is a Fellow Chartered Accountant (FCA, ICAEW) with Big 4 training and twenty years of operational experience as a founder and senior finance leader. She has run multi-currency operations across more than twenty-five countries as a founder-CEO.