Demystifying The Hype: Understanding Agentic Systems In Cross-Border FX

Over the past few weeks I’ve written a lot about how to demystify agentic software. Many of my past pieces circle a core idea: Confusion exists about what genuinely constitutes agentic payment systems. As in any early technological stage, that confusion is where a lot of people are making money.

Last year, Gartner claimed that just 130 of “thousands” of agentic AI vendors are real. Without specifying what “real” means, it critiqued the practice of “agent washing”: When vendors claim that products like AI assistants, chatbots or RPA (robotic process automation) are agentic. By the end of 2027, more than 40% of agentic AI projects are projected to be cancelled.

A lot of what is being produced is theater, let's talk about what is going on backstage.

Rules vs. Intent: Redux

OpenFX is full of nerds; that nerdiness extends well beyond the confines of finance and crypto. Thanks to a couple of them, I’ve learned a lot about the commedia dell’arte, a theatrical form that emerged in the mid-16th century and took Europe by storm.

It’s basically people in masks, acting out dramas. But every person has a role to play—the servants, the masters, the braggart and the lovers. Within the prescribed framework of those roles, they improvise, resulting in a surprisingly supple show – different every time it’s performed. Contrast this with scripted theater, where every line is written in advance.

This comparison brings us back to something we discussed elsewhere when laying out the taxonomy of automated systems, rules versus intent. Rules-based automation consists of decision trees defined by a rule-writer. Software like this has existed for decades and is not agentic, but it is often marketed as if it is.

Often these systems have a certain degree of exception handling, designed to harden them against minor deviations in those rules. As an example, you pay an invoice to a regular supplier in another country, but this month the bank doesn’t clear the payment. Rules engine-based software may fail here, requiring that you manually troubleshoot the issue. A system with exceptions might try another bank on a prescribed list before tossing the ball back to your treasury department.

This is a scripted performance, the goal is to follow the rules as faithfully as possible, and the rules can be trusted to produce the outcomes the treasurer is looking for. The actors get on stage, read their lines, and everyone applauds. In the case of rules-based treasury systems, sometimes there’ll be a chatbot attached someplace to handle the applause.

There’s nothing wrong with a rules engine with a chatbot bolted on. That’s a useful way to build an automated treasury system – predictable and easy to audit. However, if what you think you’re getting is a payments agent, this can be highly misleading.

Which brings us back to the commedia dell’arte. Modern treasury systems are becoming increasingly complex, bound not by rules but by roles and constraints. They are designed to do specific jobs according to often flexible intent. Intent means the system isn’t following one rule after another, instead it is working toward a specific goal, in the case of our treasurer, getting money into the correct hands.

It’s at this point that treasury teams often invest in more complex machine learning systems, models that are able to make predictions based on available data and act accordingly. Predictive models of this kind have also been around for years, and while they also aren’t “agentic” in the sense usually described, they are far more autonomous, capable of flagging anomalies and proposing optimizations within the constraints of the data underpinning them.

So, if “true” agentic systems are neither rules-based nor simply predictive, what are they?

An agenetic system doesn’t just make recommendations, it is capable of acting on its own decisions autonomously within the confines of the role it has been given.

As I mentioned above, these sorts of systems are rare in general, but particularly in the payments space. To know whether what you’re being shown is actually meaningfully agentic, here are a few questions you can ask.

How to Check What’s Behind the Curtain

• Does the software decide, or does it propose? Does the system operate without need for manual approvals? If every major action requires oversight, this is a recommendation engine.

• Where does the liability live? Payment orders have to trace back to an authorized instruction, per the IMF. Someone has to have made the payment. How does the system handle this?

• What’s left when you take the chat window away? If you take the chatbot away and the product works the same through a dropdown menu or some other means, then the “agentic” features are largely decorative.

• What happens during exceptions? Ask your vendor to run a live exception that isn’t in their demo. See what happens with these three questions in mind: Who decided what to do? Why? Where can the record be found?

Cheat sheet: A scripted system might bounce exception handling back to human hands because it has no script for it. An agentic system might do the same thing, but it should be able to tell you which part of its mandate the case falls outside.

The View from the Top

In April, Gartner observed that, to date, only 17% of organizations have deployed AI agents, but over 60% plan to over the next two years. AI-driven agentic applications have “the most aggressive adoption curve among all emerging technologies” they measured. Gartner places them at the Peak of Inflated Expectations.

When you’re at the peak of something, the only available direction is down, but that does not mean that automation is likely to fall by the wayside. As we’ve shown here and elsewhere, automation is not the problem. For modern treasury, we believe that automation might be the only path forward. We invest heavily in AI systems and LLM-based processes, but that doesn’t mean that we believe in them uncritically, nor does it mean that you should accept a pitch just because it contains the word “agentic.”

The better you understand how automation actually works, the easier we believe it will be for you to make better decisions about integrating it into your own processes.