AI in Treasury Management: Liquidity, FX and Payments

By Jane Ha · Updated Oct 11, 2026

AI in treasury management is the use of machine learning and language models to forecast cash and liquidity, quantify foreign exchange exposure, screen payments for fraud and tighten working capital, while treasury staff keep authority over funding, hedging and the release of money. Treasury is a good place for AI because the data is numeric and time-stamped. It's also the place where a wrong answer is hardest to undo, because cash that has left the bank is gone.

This guide is for treasurers, CFOs, controllers and finance directors deciding where AI belongs in the treasury function and which controls have to come with it. For the function-wide view, start with AI in finance and accounting and the planning hub, AI in FP&A. For the forecasting mechanics, read AI cash flow forecasting.

Key Facts

• In the 2026 AFP Payments Fraud and Control Survey (465 treasury practitioners, fielded January 2026), 76% of U.S. organizations experienced attempted or actual payments fraud in 2025, 74% were affected by business email compromise, and checks were the most targeted method at 58%, followed by ACH debits (30%) and wires (25%) (AFP).
• In the same AFP survey, only 17% of organizations used AI to combat payments fraud (AFP).
• The FBI's 2025 Internet Crime Report logged 24,768 business email compromise complaints with reported losses of $3,046,598,558, up from $2,770,151,146 in 2024 (FBI IC3).
• In AFP's 2026 Treasury Benchmarking Survey (425 responses, fielded May 2026), cash and liquidity forecasting was both the top treasury priority (63%) and the most challenging activity (49%) (AFP).
• In that same survey, AI and emerging technology policies scored 2.9 out of 5, the lowest of any policy area measured, and 38% named managing AI opportunities and risks as a significant challenge (AFP).
• Global foreign exchange trading reached $9.6 trillion per day in April 2025, up 28% from 2022, with the U.S. dollar on one side of 89% of trades (BIS).

What AI Does in Treasury

Treasury has four jobs that AI can touch: knowing how much cash you'll have, knowing what currency and rate risk you carry, getting payments out safely, and not tying up more working capital than you need. AI contributes in three ways. It forecasts from history (predictive analytics on bank and ledger data). It flags what looks wrong (anomaly detection on payments and balances). And it drafts or summarizes (a language model turning a position report into a morning briefing).

What it shouldn't do, at least not unsupervised, is move money or commit the company to a hedge. The AFP numbers show why the starting point matters: treasury teams rate their own AI policy maturity as the weakest of their policy areas, while fraud attempts hit three in four organizations. The tool arrives faster than the rulebook around it.

Where AI Helps by Treasury Function

Real Examples by Function

Liquidity forecasting. A treasurer running 14 bank accounts across five countries rebuilds a 13-week forecast by hand every Monday from spreadsheets sent by regional finance teams. A forecasting model ingests daily bank balances plus open receivables and payables, produces a baseline and a range, and shows which regions miss most often. The treasurer spends Monday on the misses, not on collecting files. AI cash flow forecasting covers the model choices and accuracy measurement in depth. Cash forecasting is the area treasury leaders themselves call hardest, so it's the natural first target.

FX exposure. A company invoices in euros, pays suppliers in yen and funds a subsidiary in Singapore dollars. The data to compute net exposure lives in three systems. An AI layer reads open invoices and purchase orders, maps them to currencies and settlement dates, and updates a net exposure report daily. Hedge decisions stay with the treasurer and the hedging policy. For scale, the dollar's role matters: it sat on one side of 89% of global FX trades in April 2025, so exposure to the dollar is rarely avoidable, only manageable.

Payment fraud screening. Business email compromise works by persuading someone to send a payment to a changed bank account. FBI complaint data for 2025 counts 24,768 such cases and over $3 billion in reported losses. A model that scores every outgoing payment against the vendor's history (new account number, unusual amount, odd timing) can hold the small share that deserves a phone call. It won't catch everything, and it should sit beside, not replace, the human call-back to a known number. Only 17% of AFP respondents used AI against payments fraud, so most firms are still leaning on manual review.

Working capital. An AI model reading payment history predicts which customers will pay late and which supplier payments can move a few days without breaking terms. The gains show up in working capital: fewer surprise shortfalls, better use of early-payment discounts. The inputs come from accounts receivable and accounts payable, which is why treasury should be at the table when those processes are automated.

Control Requirements: What Has to Be True Before AI Touches Cash

Treasury's classic controls exist because one person with access to money is a risk. AI adds a new kind of actor that can also be wrong, or be manipulated. Keep these in place.

Segregation of duties. The person or system that prepares a payment can't be the one that approves it, and neither should be able to change vendor bank details alone. An AI agent that drafts payment batches is a preparer. Give it no approval rights, and don't let one service account both create and release.

Payment release approvals. Every payment above your threshold needs a human approver with the information to say no: the amount, the beneficiary, the change history on the account and the AI's risk score with its reasons. An approver who sees only a green tick learns to click it. This is human-in-the-loop control, and it works only when the person has real time and real authority.

Bank detail changes. Treat any change to a beneficiary account as high risk. Require independent verification using contact details already on file, never those in the request. This is the defense against business email compromise, and AI makes the fraudulent emails more convincing, so the control matters more.

Model oversight. Give each model an owner, a documented purpose, the data it uses, test results and a change log. Backtest forecasts against actual cash on a schedule and set a tolerance that triggers a review. Record what the model recommended and what the human did.

Access and data limits. A model needs read access to balances and transactions. It rarely needs write access to a banking portal. Keep API credentials for payment initiation out of any tool that also reads email or documents, and keep sensitive bank data out of unapproved public chatbots.

Fallback. Know how treasury operates if the tool is wrong or unavailable: a manual forecast, a manual screening checklist, a named person who can pause automation.

How to Roll Out AI in Treasury, Step by Step

1. Start with read-only work. Forecasting, exposure reporting and daily position summaries can't move money. They build trust and expose data problems cheaply.
2. Map the data. List bank feeds, ERP modules, TMS data and spreadsheets. Fix mismatched account codes, duplicate counterparties and stale vendor records first; finance data readiness goes through what to check.
3. Baseline today's accuracy. Record how far your current forecast misses by horizon (next week, next month, next quarter). Without a baseline you can't tell if the model helps.
4. Run in parallel. Let the model forecast alongside the existing process for at least two or three cycles and compare line by line.
5. Add screening as a flagger, not a blocker. Let the fraud model raise holds for review before you let it auto-hold anything. Track false positives, because a team that's flooded with alerts stops reading them.
6. Write the approval rules down. Who approves what, at which threshold, with which information on screen. Get internal audit to read it before launch.
7. Pilot one entity or one region. Measure forecast error, alerts raised, confirmed fraud caught, hours saved and incidents.
8. Review quarterly. Backtest, re-check access rights and confirm the humans are still doing the reviewing.

Common Pitfalls

• Treating a forecast as a fact. A model gives a central estimate and a range. Plan funding to the range.
• Garbage bank data in. Cash that's misclassified in the ledger produces forecasts that are consistently wrong in the same direction.
• Giving the agent the keys. Payment-initiation rights on an AI service account collapse segregation of duties into a single credential.
• Alert fatigue. Screening that flags a large share of payments gets ignored. Tune thresholds with the reviewers.
• No policy. AFP's benchmark found AI policy the least mature area in treasury. Write the rules before the pilot, not after the first incident.
• Skipping model monitoring. Customer payment behavior shifts after a rate move or a market shock, and a model trained on calm months drifts. Rerun the backtest on a schedule.
• Hallucinated narratives. A language model summarizing the cash position can state a wrong number fluently. Feed it the figures, and have it quote them, rather than ask it to recall them.

What to Measure

• Forecast quality: error by horizon, bias (consistently high or low) and the share of weeks inside tolerance.
• Liquidity outcomes: idle cash balances, use of credit lines and short-term borrowing cost.
• Fraud and control: payments held, confirmed prevented losses, false-positive rate and time to resolve a hold.
• Process: hours spent assembling the cash position and the share of exceptions closed on the same day.

Where to Go Next

If you do one thing this quarter, baseline your current cash forecast accuracy and put a model beside it in read-only mode. Then read AI cash flow forecasting for the build, AI in FP&A to connect treasury to planning, and the AI in finance overview for the wider governance picture. Write the approval and bank-change rules down before any tool gets near a payment run.