#124 - AI as The Decision Layer

How AI is impacting cross-border payments. A guest post by Ola Oyetayo, Co-founder & CEO, Verto

Hi all - This is the 124th edition of Frontier Fintech. A big thanks to my regular readers and subscribers. To those who are yet to subscribe, hit the subscribe button below and share with your colleagues and friends.

---

This is a guest post by Ola Oyetayo, CEO and Co-Founder at Verto

---

Before a single dollar moves across a border, roughly thirty decisions have already been made.

These decisions involve various questions: Is this business who it claims to be? Do their documents satisfy the requirements of both originating and destination jurisdictions? Which payment rail offers the best combination of speed, cost, and reliability for this corridor today, at this exact moment? Is the transaction volume consistent with their history, or does this spike warrant a flag? Who has the authority to approve it, and what happens if they’re unavailable?

For most of the history of cross-border finance, those thirty decisions required people. Compliance analysts. Risk officers. Ops teams working on queues. The payment was almost incidental - a final, simple instruction that followed a long chain of human judgment. Moving the money, in the end, is simple. The hard part was everything before that.

This hard part is what AI is now learning to do.

Most fintechs deploying AI today have caught on to the possibilities of automating the visible, customer-facing aspects of the process. AI can automate up to 70-90 percent of repetitive data processing, reducing manual errors in this area drastically. However, the companies that will define the next decade of cross-border payments are building AI into the decision layer itself, the invisible infrastructure underneath every transaction.

At Verto, we’ve spent seven years moving money across 49 currencies for businesses operating in some of the hardest regions in the world to move money in and out of. These are largely frontier markets. Last year, we processed over $25 billion in volume. That journey has taught us something that doesn’t show up cleanly in pitch decks: the friction in cross-border payments was never just about the payment rails. It was about the decisions layered on top of them. Get those decisions right, fast, and consistently - and the economics of the whole business transforms.

Here is where we’ve learned that most concretely.

Onboarding as the first test

For businesses in emerging markets, opening a cross-border payments account has historically been a bureaucratic endurance test. Document requests arrived in batches - submit these five things, wait two weeks, discover you also need three more, and wait again. For many of Verto’s target customers, the average journey from application to first transaction runs fourteen to twenty-eight days. In that window, businesses would find workarounds. Some never came back.

We built an AI-powered onboarding assistant that changes the structure of that process. When a customer uploads their documents, the system analyses them in real time against the requirements of their jurisdiction, their business type, and our compliance framework. If something is missing or insufficient, the customer knows immediately. We close the feedback loop in seconds, not in a week, or after a manual review.

The impact is not primarily speed, though the speed is significant. The deeper effect is on drop-off. Forty percent of onboarding applications previously stalled mid-process, often because customers didn’t know what was wrong or what to fix. When the system tells you exactly what’s needed and why, completion rates climb. The AI assistant hasn’t replaced our compliance team - it has made their work trackable. They review decisions, handle exceptions, and manage edge cases. The machine handles the volume.

Routing as a continuous optimisation problem

Every cross-border payment travels a path. That path - which rails it crosses, which correspondents it touches - determines cost, settlement speed, and whether the payment arrives at all. Historically, routing decisions were made once, by a human, based on what was true at the time the rule was written. Rail A is preferred for this corridor. Rail B is the fallback.

The problem is that corridors are not static. Liquidity conditions shift. Rails go down. Settlement times vary by time of day. A rule written on a Tuesday performs differently on a Friday before a public holiday in the destination country.

Our AI routing engine treats this as a continuous optimisation problem rather than a static rule set. It evaluates transaction-level variables - corridor, currency pair, time of day, counterparty, payment purpose - against live performance data and routes accordingly. The effect is measurable: higher acceptance rates, lower transaction costs for our customers, and a multi-acquirer strategy that means no single point of failure. When one rail degrades, traffic moves before a human has had the chance to notice.

Transaction monitoring operates on the same principle. The behaviours that signal risk don’t announce themselves in advance. They emerge as patterns - a volume spike on an unusual corridor, a timing anomaly, a counterparty relationship that doesn’t quite fit the customer’s stated business profile. Machine learning models trained on transaction data flag these patterns faster and more consistently than rules ever did, and they improve as the data compounds.

The invisible workforce

Less visible than either of the above, but in some ways more consequential, is what AI is doing inside Verto.

We have built an internal LLM tool we call Ask Frank - a conversational AI trained on Verto’s product documentation, compliance frameworks, pricing logic, and operational procedures. An account manager handling a complex customer query doesn’t need to search through documentation or wait for a specialist. They ask Frank. A new hire learning the nuances of a particular corridor asks Frank. The institutional knowledge that used to live in the heads of the most experienced people is now accessible to everyone.

We also use AI agents for research, competitor analysis, and internal workflow automation - tasks that previously consumed significant analyst time. This is not about headcount reduction. It is about leverage. A team of 200 people operating with AI assistance can do analytical work that would have required a team of 600.

Agentic payments: the architecture, not the ambition

All three of these applications share a structural similarity: AI as an accelerant of human decisions. The compliance analyst is faster. The routing decision is smarter. The account manager is better informed.

What looks, from the outside, like three separate AI initiatives is actually one architecture becoming legible.

By creating specialised agents that each own a distinct decision domain across the customer lifecycle, we are in fact building a central decision engine - a unified layer that coordinates approvals, risk tiers, pricing logic, and limits across the entire system.

Agentic payments are the logical conclusion of an architecture we have already assembled. When a customer completes onboarding through our AI assistant, that journey has already been pre-simulated - transaction flow modelled, corridor viability assessed, likely margin impact calculated - before the first document was submitted. When a payment routes itself to the optimal rail, that decision relies on the same risk profile. The monitoring agent then updates in real time. When a growth agent identifies that a customer’s corridor usage suggests appetite for a new product, it is reading from the same data the re-risking agent used to update their limits last week.

The agents are already talking to each other. The customer doesn’t see that - but they feel it. Lead-to-first-transaction times that used to run fourteen to twenty-eight days now run two to three days in total. That compression isn’t the product of a single AI feature. It is what happens when every decision in the sequence is made by a system that never sleeps, never loses context, and never has to wait for a colleague to come out of a meeting.

The point was never to automate ourselves. The point was to understand, from the inside, what autonomous workflows actually require - and to build customer-grade infrastructure on the back of that experience.

The clearing house never disappeared

What AI is doing now is not different in kind from what happened when electronic clearing replaced paper. Instead of disappearing, the clearing house has just become invisible. Layer by layer, the human judgment that sat at the centre of every financial transaction was abstracted away - into systems, into rules, into code.

The fintechs that recognise this early - that understand they are not building AI features but an AI decision layer - will look, in ten years, like the obvious winners in retrospect. The ones that treat AI as a customer-experience enhancement will wonder what happened to their margins.

For us at Verto, the bet is specific: the business that wins cross-border payments in emerging markets will not necessarily have the best rails. It will have the best judgment sitting above them.

We are building for that.