Read the front page of results for AI in treasury management and you will notice they all price the same benefit. Your analysts get their week back. The forecast comparison report that ate half of Tuesday now takes seconds. One vendor page puts the recovered time at four to eight hours a week per analyst.
That is a real benefit if you employ treasury analysts. Most companies asking this question do not. If you run a group of four or ten or thirty entities across a few currencies, your treasury function is one finance lead, an accountant on contract, and you. There are no analyst hours to reclaim, so the entire value case in the SERP fails to transfer, and it fails quietly, which is worse.
So here is the version for that reader. EX FI provides multi-currency payment accounts, not a treasury management system, so we have nothing to sell you at the end of this. What follows is what AI genuinely does in treasury work, which of those jobs survive contact with a company that has no treasury department, and the one prerequisite that decides whether any of it is worth paying for.
What AI for treasury management actually means
It is not one product. It is AI applied at whatever points in the treasury workflow the work is repetitive, data heavy, or hard to keep current by hand, which is why the capability map runs across eight separate areas from cash visibility to reconciliation to reporting.
Three different technologies get sold under the same two letters, and the difference matters when you are reading a pitch:
- •Machine learning finds patterns in history and predicts from them. It learns that one customer pays on day 30 and another always drifts to day 45, then forecasts accordingly instead of trusting the stated terms.
- •Generative AI drafts. It writes the cash position summary, explains a variance in plain language, and produces the board paragraph.
- •Agentic AI runs a recurring job end to end. It assembles the daily position or reviews the payment run on a schedule, then surfaces what needs attention.
The honest framing for a small group is that none of these save you an analyst salary you were not paying. What they can save you is a category of error, and what they can find you is cash you already own but cannot currently see. Measure any tool against those two, not against reclaimed hours.
The five jobs AI does in treasury, ranked for a company with no treasury team
1. Reconciliation and transaction matching
The strongest case, and the one to start with. Matching engines apply fuzzy logic to bank statements against open invoices even when the payment reference is truncated or mistyped, then surface only the genuine exceptions. It works because it runs on data you already hold, and it shows a result in weeks rather than quarters. Practitioner confidence tracks that: the share expecting AI to ease manual reconciliation rose from 55 to 62 percent in a single year, per Strategic Treasurer research summarised alongside the wider adoption picture.
2. Anomaly and payment fraud detection
The second strongest, and the one small groups most underrate, because a four person finance team is exactly the environment where an unusual beneficiary passes unchallenged. Rule based screening cannot keep up with attacks that adapt, so the model learns the normal cadence, amount and destination of your outbound payments and flags the outlier before the wire leaves. The same AFP research finds more than three quarters of surveyed US organisations faced attempted or actual payments fraud in 2025, names treasury as the department most likely to catch an attempt at 83 percent, and puts the share of organisations using AI against fraud at just 17 percent.
3. Short-term cash forecasting
Genuinely useful, routinely oversold. More than 60 percent of treasury professionals in AFP's survey rank cash and liquidity forecasting as their hardest task, and the share expecting AI to improve it climbed from 65 percent in 2024 to 76 percent in 2025. Read that carefully. It measures expectation, not delivered accuracy. The widely quoted figure of a 30 percent or better improvement in forecast accuracy is GTreasury's claim about its own product, not an independent finding, and it is measured on enterprises with years of clean history to train on. If your group's payment history lives in four banking portals and a spreadsheet, there is nothing yet to learn from.
4. Reporting and variance narrative
Real time saved, small absolute number. If nobody currently writes a weekly treasury report, automating it produces a report nobody was waiting for. Worth having once it is bundled with something you already bought, rarely worth buying alone.
5. Liquidity and FX scenario modelling
Last, and for most small groups out of scope. Modelling three rate environments before a board meeting is valuable when your positions are large enough that hedging them is a decision. Below that threshold it is an interesting screen. Note the mechanism the enterprise vendors point at, though, because it is the one that scales down: better forecast accuracy is what reduces borrowing costs and puts idle balances back to work. That is an operational effect of knowing where your cash is, not a return, and nobody should sell it to you as one.
The prerequisite nobody puts in the demo
Across all five, the constraint is almost never the model. It is the data underneath it. Treasury information sits scattered across banking portals, payment tools and accounting systems, and by the time a figure reaches a report it is frequently already stale. Point any model at incomplete or out of date numbers and the output inherits every flaw, however good the model is.
The causes are structural rather than technical. Fragmented systems, inconsistent data ownership and different assumptions held by different functions drive more forecasting error than analytical limitations do. And the sharpest observed divide in treasury today is not between companies with better or worse models. It is between organisations still working manually and those whose data already sits on a platform that a model can read.
What a model reads from, concretely, is an interface: real time balance and payment status APIs, ISO 20022 messaging, pre-validation of account details before a payment goes out. Batch files and overnight statements cannot feed anything that claims to work intraday.
This is where our own structure is relevant, and we will state it as structure rather than as a claim. An EX FI account holds balances in GBP, EUR, USD and more than 16 currencies with dedicated IBANs on one set of rails, SWIFT, SEPA, FPS, BACS and CHAPS, with internal transfers between your own entities free and Xero reconciliation built in. That does not make it an AI product. It means the multi-entity, multi-currency position exists in one place, which is the thing you would otherwise spend the first year of any AI treasury project assembling. Getting that right is also the cheapest half of whether treasury management software is worth buying at your size, and it changes what each internal transfer between your own entities really costs.
One caveat on scope, since this is regulated territory. EX FI accounts are payment accounts, not bank accounts. Regulated payment services are provided by Gemba Finance Limited, FCA FRN 804853, for UK customers; EX Financial Solutions Ltd is not itself authorised by the FCA and acts as a distributor. Funds are not covered by the FSCS and are instead safeguarded in segregated accounts under the Payment Services Regulations 2017.
Where AI in treasury still fails
Four limits, all of which survive better tooling.
It is not a crystal ball. Models still learn from history, so they break on the unprecedented: a sudden market shift, a geopolitical event, a policy change. They improve the quality of a forecast. They do not remove uncertainty from it.
It cannot own the exceptions. Structured, rule shaped work it handles at high accuracy. The messy exception that needs someone to weigh context is precisely what gets escalated to a human, and in a small team that human is you.
It cannot be accountable. A recommendation nobody can trace back to real data is indefensible to an auditor or a board, which is why explainability and auditability are treated as requirements in treasury rather than nice extras. Using a model does not move responsibility onto the model.
It should not move money by itself. The boundary that makes agents usable in treasury at all is that an agent prepares and proposes while a person approves anything that actually moves funds. Any vendor blurring that line is selling you a liability.
Worth adding the vendor caveat too, since the efficiency percentages travel far. The frequently cited 20 to 30 percent reduction in treasury operating costs traces back to BCG research on digitisation generally, quoted in vendor marketing alongside a 40 percent improvement in suspicious activity identification at large institutions. Neither describes a ten person company.
What to do first if you run a small multi-entity group
A sequence, not advice on your finances:
- •Make the position visible before you evaluate any model. Every entity, every currency, one view, refreshed without anyone exporting anything. This step is unglamorous and it is most of the value.
- •Pilot exactly one workflow. Reconciliation or short-term forecasting, because both run on data you already hold and both show a result fast enough to judge.
- •Keep a person on anything that moves money. Treat early output as a draft to review, not a verdict to execute.
- •Measure time saved and errors caught, then expand on that evidence rather than on the roadmap.
If that sounds slow, note the field you are actually competing in. In a survey of more than 100 firms across the US, Europe and Asia, fewer than 10 percent of treasury teams use AI for core functions and half have not started at all. Being deliberate here is normal. Buying a model to sit on top of data you cannot yet trust is the expensive mistake, and it is the one the current wave of marketing is most likely to sell you. When you do get to the shortlist, it is worth comparing treasury management tools on their bank connectivity before their AI features, and if your interest is autonomous systems paying each other, that is a different question covered in the payment rails AI agents actually settle on.
FAQ
Can AI replace a treasury team? No. It absorbs repetitive analysis and pattern spotting. Strategy, banking relationships and the final call on any flagged payment stay with people, and the working boundary in treasury is that a system prepares and proposes while a person approves anything that moves money.
Which treasury task should AI take on first? Reconciliation or short-term cash forecasting. Both run on data your team already holds, so neither requires a data project before it can produce anything, and both give you a visible result quickly enough to judge whether to continue.
How accurate is AI cash flow forecasting? More accurate than run-rate models where patterns are stable, and no better than the data it reads. It cannot anticipate an unprecedented event, and headline accuracy improvements are usually vendor figures measured on enterprise datasets rather than independent findings.
Do you need a data scientist to use AI in treasury? Purpose-built tooling that plugs into an existing system does not require one, though that assurance generally comes from the vendors selling it. For a small group the scarce input is connected, current data rather than modelling skill.
How many treasury teams actually use AI today? Fewer than 10 percent for core treasury functions in a survey of more than 100 firms across the US, Europe and Asia, with roughly half not having started at all. Adoption trails the coverage by a wide margin.