There is a paradox at the heart of corporate treasury. The function responsible for keeping a business financially alive – managing its cash, hedging its risks, funding its growth – has historically operated with less real-time information than almost any other part of the organisation. A company’s marketing team knows what happened on its website this morning. Its treasury team, in most cases, is working from yesterday’s bank statements. AI is closing that gap and the speed at which it is doing so is something worth paying attention to. 

The conversations happening inside treasury functions right now are unlike anything from the past decade. Not because the technology arrived – artificial intelligence has been discussed in finance for years. But because it is now being applied to the specific, unglamorous problems that treasury teams wrestle with every day: the inaccurate forecast, the unmatched payment, the exceptions queue that never empties. And those applying it are seeing results that are genuinely hard to dismiss. 

What Treasury Does – and Why the Stakes Are High 

Corporate treasury is not accounting. The distinction matters more than most people realise. Accounting records what has happened. Treasury manages what is happening right now and positions the organisation for what comes next. It sits at the centre of three responsibilities: 

• Liquidity. Ensuring the company always has enough cash, in the right place, at the right time – because a profitable business can still become insolvent if liquidity runs out at the wrong moment. 

• Risk management. Protecting the business from financial volatility – currency swings, interest rate movements, counterparty failures – that can quietly erode margins even when trading is strong. 

• Capital optimisation. Deciding how the organisation raises and deploys capital – minimising the cost of debt, and ensuring surplus cash earns a return rather than sitting idle across dozens of accounts. 

The evidence of structural pressure on this function is consistent across the most credible research available. PwC’s 2025 Global Treasury Survey*, drawing on insights from 350 treasurers worldwide, found that both CFOs and treasurers ranked cash and liquidity management as their top priority – while simultaneously flagging that the challenges of poor data quality (cited by 76% of respondents), lack of effective tools (53%), and limited incentives for business units to contribute (46%) are near-universal. 

AI is touching nearly every corner of the treasury function – hedging, fraud detection, counterparty risk monitoring, regulatory reporting. Each area has its own story worth telling. This piece focuses on two: cash forecasting and cash reconciliation. Not because they are the only areas that matter, but because they are where the data problem is most acute, most daily, and most directly visible in the decisions treasury makes every morning. Fix the information quality here, and the ripple effect across the function is immediate. 

↗ PwC 2025 Global Treasury Survey 

Cash Forecasting: When Your Most Important Output Is Also Your Least Reliable

The AFP Treasury Benchmarking Survey 2025, drawing on responses from over 500 practitioners, found that over 60% of treasury professionals identify cash or liquidity forecasting as their single most challenging task. Behind that number is a process that has not fundamentally changed in decades. In most organisations, cash forecasting is a manual collection exercise. Treasury sends spreadsheet templates to business units across dozens of countries. It chases submissions. It consolidates responses in Excel, applies judgment-based adjustments, and publishes a forecast that is, by the time it lands, already several days stale. The inputs carry all the biases and inconsistencies of human estimation – optimistic sales pipelines, conservative payables assumptions, and timing guesses that rarely survive contact with the actual payment behaviour of real customers and suppliers. 

What AI Changes:

Machine learning approaches forecasting differently. Rather than collecting estimates from humans, AI ingests data directly – from ERP systems, bank feeds, accounts receivable ageing reports, payroll schedules, and accounts payable due dates. It trains on years of historical behaviour: how specific customers actually pay, how seasonality affects inflows in each market, how macro signals correlate with collections patterns. The forecast is not a human estimate. It is a pattern-detected inference, updated continuously as new data arrives.  

↗ AFP Treasury Benchmarking Survey 2025 

Cash Reconciliation: The Bottleneck That Hides in Plain Sight

Every day, treasury and finance teams must match thousands of incoming payments against open invoices in their ERP systems. When a payment arrives with complete remittance information – the right invoice number, the correct amount, the right entity – the match is automatic. But payment data is rarely that clean. References are missing, truncated, or formatted inconsistently. A customer pays five invoices with a single wire that does not identify which ones are covered. A bulk payment file contains 800 transactions with non-standard descriptions. 

Every unmatched payment enters an exceptions queue. In most treasury and finance departments, that queue never fully empties. 

What AI Changes: 

AI-powered reconciliation engines learn the payment behaviour of each counterparty over time. They recognise that a particular customer always pays on the 28th of the month with a reference in a specific format. They match partial payments to the most likely open invoice based on amount, timing, and historical patterns – even when remittance information is incomplete. They handle complex many-to-many matching scenarios that defeat rule-based systems. And they do it continuously, not at month-end. 

The result is a fundamental shift in where human attention goes. Instead of manually investigating hundreds of exceptions from scratch, treasury teams review AI-suggested matches, confirm or adjust, and move on. The exceptions queue shrinks from hundreds to tens. Month-end close happens on schedule rather than when the queue finally clears. 

Why the Right Partner Changes Everything

Here is the part that gets lost in the AI conversation: the technology itself is not the hard part. The hard part is knowing what treasury data actually means – how it flows between banking systems, ERP platforms, and payment infrastructure; where the edge cases hide; how remittance formats vary by geography and payment rail; why a reconciliation break in one entity can cascade into a liquidity misjudgement in another. General-purpose automation tools applied to treasury without that underlying domain knowledge produce outputs that look plausible but require extensive expert review to trust. 

Opus’s Smart Cash App, an agentic B2B cash application, is one example of this approach in practice. It addresses both the forecasting data quality problem and the reconciliation matching challenge within a single integrated workflow. The outcomes align with what the research suggests is achievable: material reductions in manual cash application effort, near-perfect auto-match rates, and month-end close that happens on schedule. 

The distinction between domain-led and technology-led AI is visible in deployment timelines, exception rates, and whether the output can be trusted without rebuilding it from scratch. It is also visible in whether treasury teams actually adopt the solution – or quietly shelve it after six months. 

The Real Return: Five Things That Change When AI Works

For years, the treasury function has attracted some of the sharpest analytical minds in any organisation. The persistent frustration is that those minds spend most of their time assembling data rather than interpreting it. AI changes that equation – by shifting treasury’s energy from chasing data to acting on it. 

When it works, five things change: 

• Treasury knows what it has. Cash positions are accurate before the first meeting of the day – not reconstructed from yesterday’s statements. 

• Treasury knows what is coming. Forecasts reflect live data and real payment patterns – not optimistic submissions from business units. 

• Problems surface before they become losses. Anomalies in payments and reconciliation are flagged in real time, not discovered in the quarterly audit. 

• The function scales without adding headcount. As payment volumes grow, AI absorbs the matching and exception workload that would otherwise require more people. 

• The treasurer becomes a strategist. Time reclaimed from data assembly is redirected to scenario planning, risk decisions, and CFO-level conversations that actually move the needle. 

The organisations that move first will not just operate more efficiently. They will make better decisions, and that advantage compounds.