Can your treasury team rely on AI to run FX risk management?
FX risk management depends on three things working together: clean exposure data, a sound hedging policy and experienced judgment about how the pieces connect. AI can speed up the work at each step but doesn’t supply any of those foundations.
If you work in treasury, somebody has probably asked what you are doing about AI. Maybe the CFO, maybe the board, maybe a vendor who got twenty minutes on your calendar. It is a reasonable question, and the honest answer has less to do with the technology than with how your FX process works today.
What parts of FX work does AI handle well?
The repetitive, rules-based work that eats your month.
Where AI earns its place
- Manual, repetitive daily tasks, such as pulling rates, matching data and formatting reports
- Rules-based work that changes slowly, such as hedge accounting documentation, regulatory reporting and standard month-end packs
- First drafts of analysis and variance commentary
Treasury teams are moving quickly on this. In PwC’s 2025 Global Treasury Survey of 350 treasurers, 74% said they were either expanding or actively using AI, while only 26% rated their AI capabilities as moderately or very mature. That gap between adoption and maturity is worth watching.
What happens when you automate a weak FX process?
If an FX risk management process is weak, automation will just make it run faster. People often say “garbage in, garbage out” about data, and it applies to FX risk management processes, too.
A TD Bank survey of 246 finance leaders at AFP 2025 found that nearly 80% of treasury professionals still rely on manual or fragmented systems. For many teams, the honest starting point is inherited spreadsheets, disconnected sources and a process nobody has redesigned in years.
If that sounds familiar, check out these ten warning signs that spreadsheets may be increasing your FX risk.
Put AI on top of that, and you get faster output, not better output. An AI tool will produce a confident-looking analysis from whatever exposure data you give it. It cannot tell on its own whether that data is complete, whether an entity was missed, .
What can’t AI judge about your FX program?
Whether the process underneath it is any good. These are judgment calls; they sit upstream of everything AI handles well, and they decide whether that work produces something you can trust:
- Whether the underlying process is right in the first place
- The strategic side: your FX policy, what to hedge and why, and how results are explained to leadership
- Connecting data, process and results end to end, so a variance can be traced back to its cause
- Knowing what good looks like, which comes from years of running FX programs
Take a simple example: AI cannot draft a clear explanation of why last month’s FX loss was larger than forecast. It cannot tell you whether the loss came from a forecasting miss, an exposure that was never captured, slippage in the relevant rates, or a booking error that goes unnoticed. Someone has to know where to look, and the process has to make that answer easy to find.
How do you know whether AI forecasting is actually working?
You measure it by flow against what you were doing before. GE HealthCare did not take the improvement on faith when it adopted AtlasFX AI forecasting. The team back-tested AI accuracy against its manual forecasts on a mean average error basis, across 253 flows in 2024 and 383 in 2025. Roughly 80% of P&L flows improved, and the team estimates 120 hours of annual time savings in the forecasting process.
That is the standard of proof to hold any AI tool to, including one your team assembled itself. The question is not whether it produces a number, but whether anyone has measured it against the old way, by flow, over a period long enough to mean something.
What should you look for when weighing build vs. buy FX risk management software?
Look at whether the process underneath the software is sound, not whether the product has AI features. It comes down to three things: how exposure data is captured from your ERPs and kept current as entities and accounts change; whether the corner cases are understood and accounted for that are crucial towards creating meaningful analytics; and whether the people behind the product have run an FX program themselves.
Automation reinforces whatever process it sits on. Here is how AtlasFX handles each of the three:
- Exposure data you can trust. AtlasFX delivers and maintains the ERP scripts that identify your balance sheet exposures, and the scripts update as entities and GL accounts are added.
- Built by practitioners. AtlasFX’s leadership team includes people who ran FX cash flow and balance sheet hedging programs at Hewlett-Packard, so the platform reflects how those programs work in practice.
- Configured to you unique situation. There isn’t a one-size-fits-all work flow and analytical package when it comes to FX risk management, and having the flexibility to adapt to each customer’s configuration is crucial in getting the results analytics right.
7 questions that show where your FX process stands
Note: a “no” to one of these questions is not disqualifying on its own. Three or four together tell you the process needs work before any automation goes near it.
- Is your FX exposure data pulled directly from your ERP without manual adjustment?
- Would you know if a new entity or GL account was missing from your exposure data?
- Could you explain last quarter’s FX variance to your treasurer in under ten minutes?
- Are your hedge accounting documentation and your exposure reports built from the same data?
- Can your team tell a forecast error from a hedging error when reviewing results?
- Can you easily find booking errors or remeasurement/revaluation inconsistencies?
- If AI drafted your monthly FX commentary today, would you trust the data underneath enough to sign it?
So, can AI run FX risk management?
AI cannot run FX risk management on its own, and that isn’t the useful way to frame the choice anyway. It’s very good at the work that fills your month, and very bad at knowing whether that work rests on anything solid.
The teams getting real value from AI are the ones who fixed the process first, which means the question in front of you is about your process rather than about the technology.