AI Treasury Management 2026: Forecasting, TMS Picks & Governance

AI treasury management is mostly about one question: how much cash will we have, where, and when? Machine learning models now read ERP and bank data to project liquidity 13 weeks or 90 days out, flag unusual payments, and explain why actuals missed the forecast. Vendor pages describe this as treasury moving from the back office to a strategic role. That can be true, but only if the forecasts hold up on your own data.

This guide covers what the tools actually do, how the main platforms differ, how to test AI forecasting in Excel before you buy anything, how a smaller team can work out whether it pays, and the governance questions banks and regulators are now asking about AI agents.

1. What AI Treasury Management Actually Does

In plain terms, these tools turn transaction history, bank balances, and ERP records into a projection or an alert. Almost every feature fits one of six jobs:

Planning platforms such as OneStream sit next to this list. They extend generative and agentic AI to enterprise-wide planning, and treasury is one user among many. For the wider finance picture, see our guide to AI for finance teams.

Summary: name the job first. A forecasting engine, a payment control, and a chat assistant need different data and get judged by different tests.

2. How AI Cash Flow Forecasting Works

Providers including J.P. Morgan and Kyriba describe the same basic approach. The model learns patterns in historical transactions and in how specific customers and vendors behave, then projects cash 13 weeks or 90 days ahead. Each week it pulls fresh actuals from the ERP and updates the projection.

Two families of models come up most often:

• Random forests and other tree-based models: these work well on tabular data such as invoices, due dates, and payment terms. You can see which inputs drove a prediction, which helps when someone asks why the forecast moved.
• Neural networks: these can pick up non-linear and seasonal patterns across many series at once, which helps with large, multi-entity cash flows. They are harder to explain, and they need more history to beat a simpler model.

For most treasury teams, model choice matters less than three things: clean categories (payroll, tax, vendor payments, collections), enough history to cover seasonal cycles, and a habit of checking the forecast against actuals every week. Large one-off flows such as acquisitions, debt issuance, or a lawsuit settlement still need a person to enter them. No model can learn an event it has never seen.

Variance analysis is where AI often helps first. Comparing last week's forecast with actuals by category, and flagging the lines that moved most, is repetitive work that software does well. A person still decides whether a miss was a timing difference or a real change.

Summary: start with a model you can explain, measure it weekly by category, and keep people in charge of the large one-off flows.

3. AI Treasury Management Software Compared

The table below reflects how each platform describes itself in public material as of November 2026. It is not based on our own testing. Features vary by module, region, and contract, so confirm what is included in your quote.

FIS Neural Treasury vs Kyriba is one of the most searched comparisons. A fair way to settle it is a back-test: give each vendor the same 12 to 24 months of anonymized history, ask each to forecast a period you have hidden, and compare the errors by category. Also compare bank connectivity, how overrides are recorded, and the total cost over three years, not just the first-year license.

Summary: shortlist by architecture and bank coverage, then let a back-test on your own data pick the winner.

4. Starting in Excel Before You Buy a TMS

Many treasury teams still forecast in spreadsheets, and that is a reasonable place to start. A structured Excel model gives you the baseline that any AI tool has to beat. Detailed step-by-step guides are hard to find, so here is a vendor-neutral sequence:

1. Export 18 to 24 months of bank transactions: one row per transaction with date, amount, counterparty, and currency.
2. Categorize consistently: use the same ten to fifteen categories you report on, such as collections, payroll, tax, vendor payments, and debt service.
3. Build a 13-week rolling view: sum by week and category, then forecast each line with simple rules first, such as averages, payment terms, and known schedules.
4. Add a forecasting function: Excel's built-in forecast sheet, or Python in Excel, can add seasonality-aware projections to the lines with steady history.
5. Measure the error: each week, record the gap between forecast and actual as a percentage, by category. This is your baseline.
6. Use Copilot for review, not numbers: Microsoft Copilot can help explain variances and draft commentary, but check every figure it quotes against the sheet.

Teams on Dynamics 365 can also set up Azure AI predictive cash flow services inside the ERP. Whichever route you choose, keep the error log. It is the evidence you will need later to justify, or reject, a paid platform.

Summary: a measured spreadsheet forecast is your baseline. Buy software only when it beats that baseline on your data.

5. Cost vs Benefit for Mid-Size and Smaller Teams

Published ROI figures usually come from vendors or very large companies, so they rarely help a smaller treasury team. Work it out yourself instead. Count the license, implementation, bank connectivity fees, and internal time. On the benefit side, the two items you can usually measure are hours saved and cash you no longer need to hold as a buffer.

Annual benefit = (buffer released × yield gap) + (hours saved × loaded hourly cost)

An example with made-up numbers: a company holds a $2,000,000 cash buffer because its forecasts are unreliable. Better forecasts let it release $500,000 into an account or short-term paper earning 4% more. The team also saves 6 hours a week at a loaded cost of $60 an hour over 48 weeks.

These figures illustrate the method and are not benchmarks. Notice how sensitive the result is: if the buffer cannot actually be released, the same tool loses money. Count a benefit only once your error log shows the forecast has improved over several months. Our guide to AI ROI metrics goes further into baselines and attribution.

Summary: the benefit usually comes from released cash and saved hours, so measure both before you sign a multi-year contract.

6. Fraud Detection and Payment Controls

Anomaly detection is the second big use of AI in treasury. Models learn what normal payments look like for each vendor and flag the ones that do not fit. Typical signals include:

• Changed bank details: a vendor's account number changes shortly before a large payment.
• Unusual amounts or timing: a payment far outside the usual range, or released outside normal hours.
• New counterparties: a first payment to a beneficiary in an unfamiliar country or bank.
• Duplicates: the same invoice paid twice under slightly different references.
• Approval shortcuts: the same person creating and approving a payment, or approvals that skip a step.

An alert is a lead, not a verdict. Decide in advance who reviews flags, how fast, and what happens to the payment while it waits. Then track false positives. A control that flags everything gets ignored, and that is when fraud gets through.

Summary: use AI to rank payments by risk, and keep a named person responsible for releasing anything it flags.

7. Governance, Unauthorized AI Agents, and Regulators

As AI agents gain the ability to log in and act, banks have started paying attention. News coverage in 2026, including from Reuters, reports U.S. banks working to detect unauthorized third-party AI agents accessing customer accounts, alongside new regulatory guidance for examiners reviewing banks' use of AI. For a corporate treasury, the practical risk is simple: a tool connected to bank portals with credentials nobody approved.

Treasury-specific governance frameworks are still rare, so here is a short checklist to start from:

• Inventory: list every AI tool that touches treasury data, including personal chat tools and browser extensions. Our shadow AI guide covers how to find them.
• No shared credentials: never give bank portal logins or payment rights to an AI tool outside your approved TMS and bank channels.
• Read vs act: separate tools that read and summarize from tools that can initiate payments, and require human approval for anything in the second group.
• Explainability: for any AI-suggested hedge, investment, or funding move, someone must be able to say why, in writing.
• Logging: keep a record of the forecast, any override, who made it, and the outcome.
• Data rules: define which data may go into general-purpose AI tools. Bank balances, account numbers, and counterparty details usually should not.

Regulated institutions face stricter requirements, which we cover in our banking AI compliance guide.

Summary: let AI read and suggest, keep payment authority with people and approved systems, and write down every override.

8. Migrating From a Legacy ERP or TMS

Moving from an older ERP or TMS to an AI-capable platform is mainly a data project. A phased approach lowers the risk:

1. Clean the history: fix categories, remove duplicates, and map old account codes to the new structure. Forecasting models are only as good as the actuals they learn from, the same principle behind our AI-ready data guide.
2. Connect banks first: reliable daily balance and transaction feeds come before any AI feature.
3. Run in parallel: keep the old forecast running for at least one full quarter and compare both against actuals.
4. Switch category by category: hand over the predictable lines first, such as payroll and recurring vendors, and keep manual control of volatile ones.
5. Retire the old process: only once the new forecast has beaten the old one for several consecutive months.

Summary: clean data, connect banks, run both systems side by side, and let the error log decide when to switch.

9. Courses and Certifications for Treasury Teams

Searches for an AI in treasury management course are growing. The Association for Financial Professionals (AFP) offers a digital badge on applying large language models such as ChatGPT to treasury tasks, and practical courses increasingly use Microsoft Copilot in their exercises. When you compare options, check:

• Hands-on work: whether you build a forecast or a variance analysis on realistic data, not just watch slides.
• Controls content: whether it covers data rules, model risk, and approval workflows, not only prompting.
• Currency: the syllabus date, because tools and guidance change quickly.
• Recognition: whether your employer or professional body counts it toward continuing education.

Summary: choose courses with practical exercises and a controls module, and confirm the syllabus is current.

10. Where TheBar Fits (and Where It Does Not)

TheBar is not a treasury management system. It does not connect to your banks or ERP, forecast cash, or move money on your behalf. It is a desktop app for chat, documents, slides, websites, and research, and it helps with the review, creation, and delivery work around the tools above.

• Research: search the live web to compare TMS vendors, follow banking AI news, and check current course details, with sources you can verify.
• Documents: draft an RFP, a back-test plan, or a treasury AI policy from non-confidential notes, and export it as a PDF.
• Slides: turn summary forecast results into a board or CFO update. Our guide to AI board reporting covers how to structure it.
• Websites: build a simple internal page that explains your forecasting process or new payment controls to the wider finance team.

Summary: use treasury platforms for cash and payments, and use TheBar for the research, writing, and presentations around them.