Why TAI intelligence compounds every month

Everyone in treasury software is selling AI right now. That's worth pausing on, because if every vendor has AI, then AI isn't your advantage. The question is what's underneath it.

The large language model (LLM) itself, the engine that understands language, reasons through problems, generates answers, isn't Kyriba's differentiator. Every vendor in this market has access to the same generation of frontier models. OpenAI, Anthropic, Google: they sell the same engine to everyone. If you've heard a competitor describe their product as "AI-native," they're almost certainly running on one of those same engines.

What differs is everything wrapped around that engine. In engineering, that collection of surrounding pieces is called the harness. Essentially, everything in the agent that isn't the LLM belongs here: the prompts and instructions, the tools and integrations, the memory, and the orchestration. Think of it less as a technical term and more as what it actually does: the onboarding, the guardrails, and the judgment layer that turn a brilliant new hire into someone you'd actually trust with your company's treasury operations.

The model is table stakes. The harness is the moat.

A frontier AI model, on its own, is like hiring the smartest person you've ever interviewed, but who has never worked a day in treasury and knows nothing about your company. They understand finance in the abstract. They don't know your entity structure, your bank relationships, your approval hierarchy, or the dozen unwritten rules your team follows because of something that went wrong three years ago.

The harness is what closes that gap. For TAI (Kyriba’s Trusted AI), it's four things working together:

• Domain knowledge built from decades of Kyriba working with treasury teams. This knowledge reflects how treasury works under pressure, at scale, across jurisdictions

• RBAC-governed access to your real data. Role-based access control to your treasury is set up in real time. TAI never has to ask you to explain what should already be obvious from your own system.

• Guardrails and audit trails. Every action is explainable, traceable, and reversible, with a human approving anything that matters.

• Rules and checks refined across deployments. As Kyriba's product team encounters new edge cases in the real world, those become permanent logic inside the harness, shipped to every customer.

That last point is where the compounding happens.

What this looks like in practice

In a recent session with customers, I asked the room: "One of your team members is leaving next month. What do you need to make sure gets handed over to their replacement?"

I asked TAI. It pulled together the accounts they managed, the approvals they were responsible for, the recurring exceptions they'd been handling manually, even a pattern of intercompany transfers they processed every month that wasn't documented anywhere. Nobody had to explain the entity structure. Nobody had to upload a spreadsheet. TAI already knew, because it's connected directly to the treasury data itself.

Then we asked about a supplier whose bank details had recently changed. TAI flagged it, cross-referenced the change against the vendor master and payment history, and surfaced exactly the kind of pattern that fraud teams look for.

The same happened when a user asked if a bank statement had been successfully received today: TAI was able to respond by showing which accounts received the statement and which had not. When asked about cash balance, TAI returned the information by including the forecast, using value data, commenting on closed accounts that contain cash, or excluding intercompany accounts.

In those examples the harness played a critical role. It explains how users are managed in the application, the logic on approval and roles, how suppliers are modeled in the system, when a report needs to be executed, and treasury management best practices.

These aren't hypotheticals. They're the kinds of patterns that get built into the harness one real deployment at a time and that become part of what every Kyriba customer benefits from. A vendor that shipped AI features last year starts from zero. Every edge case they haven't encountered yet is a blind spot.

By the end, the room understood why TAI brings value to them. Not because I'd explained the technology better. Because they'd seen what it does when it's actually plugged into their business, on questions that mattered to them specifically.

One important clarification: when we say the harness gets stronger over time, that means the rules, checks, and guardrails get stronger. Your treasury data stays in your environment. It is never pooled with other customers' data and never used to train a shared model. What travels is logic: Kyriba's product team identifies a pattern worth catching, writes it into the harness as a new rule or safeguard, and ships it as an update. Everyone's harness gets sharper. No one's data goes anywhere it shouldn't.

Three questions to ask any treasury vendor

If you're evaluating AI capabilities across treasury vendors, these three questions cut through the noise quickly:

1. How many real treasury deployments has the vendor carried out, especially with the complexity of global operations? The answer reveals how much real-world judgment is in the harness.
2. When your team encounters an edge case the model hasn't seen, what's the process for that becoming a permanent safeguard and how long does it take? A vendor with a thin harness will describe that process as a future roadmap item. A vendor with a mature one will describe a process that's already running.
3. When the underlying AI model improves, what changes in your product? What doesn't? The answer to the second part matters most. If the only thing improving is the model, you're relying entirely on the AI labs. A strong harness means your product improves on two tracks simultaneously, and the compounding is yours to keep.

Where this goes next

Better models don't shrink this advantage but move it up the stack. As models get more capable, they unlock harder workflows: the kind of multi-step judgment calls that used to require a senior treasury analyst. Harder workflows expose new edge cases. New edge cases get built into the harness. The harness keeps growing to meet whatever the model can now attempt. With hundreds of customers already using TAI in production, and more joining every month, thousands of new questions are compounding TAI intelligence exponentially, month over month and quarter over quarter.

The same logic is about to extend beyond Kyriba's own interface. The next chapter is orchestration, where the harness becomes the treasury intelligence layer inside whatever AI tool your organization already uses, whether that's Kyriba's own interface or something else entirely. The AI asking the question changes. The trusted treasury intelligence underneath it doesn't.

The vendors who will matter in treasury AI aren't the ones who shipped AI features first. They're the ones who've been building the judgment layer for decades: one deployment, one edge case, one guardrail at a time.

That's not something you bolt on in a quarter.

Curious what this looks like with your own treasury data? Get a demo or talk to your account team.