Every payments company talks about AI right now. Most of them are talking about chatbots, prompt wrappers, or bolting a large language model onto an existing dashboard and calling it "intelligent."
That is not what we are doing at Payrails.
I lead the Data Intelligence team here, and I want to share how we are actually building AI into the platform: not as a feature announcement, but as a fundamental shift in how we think about payments infrastructure.
The foundation nobody wants to talk about
Our team is called "Data Intelligence," but if I am being honest, a more accurate name for what we spend a large portion of our time doing would be "Data Janitorial Services."
Before you can build any meaningful intelligence layer, you need clean, unified, trustworthy data. Payrails sits at the center of a merchant's payment stack, ingesting transaction data from every PSP, acquirer, and payment method they use. That data arrives in dozens of different formats, with different naming conventions, different error taxonomies, and different levels of completeness.
Our job is to refine all of that into a single, coherent data layer. Think of it as the plumbing that makes everything else possible. The unified analytics dashboard that pulls performance data from Adyen, Stripe, dLocal, and every other provider into one view? It runs on this layer. The ML models we are training? They learn from this layer. The agents we are building? They will reason over this layer.
This is the flywheel at the core of our AI strategy: more unified, cleaner data enables better ML training, which enables smarter co-pilot functionality, which eventually enables autonomous agents. Every improvement to the foundation compounds.
How we build AI: iterate fast, learn from the team
We follow a deliberately iterative approach to applying AI and machine learning. The philosophy is simple: launch quickly, put it in front of real users, gather feedback, and improve based on actual usage patterns.
Our anomaly detection system (which my colleague Jacek Filipczuk recently wrote about) is a good example of this in practice. We built an ML model that automatically learns the distribution of a merchant's transaction segments and surfaces drops or anomalies in real time.
But what matters more than the model itself is what happened after we shipped it. Our Payments Strategy team started using it immediately. They told us the alerts needed more context: "Can you show me which payment method triggered this?" "Can you include the decline reason?" "Can you add the transaction volume so I can tell if this is material?" Every piece of feedback made the system sharper.
Before anomaly detection, our Payments Strategy team spent an hour every morning manually scanning merchant dashboards to spot issues. That is not scalable, and it is error-prone. Now, the model does the scanning and the team focuses on the cases that actually matter. The human expertise did not go away, it got redirected to higher-value work.
Machine learning: the quiet workhorse
These days generative AI gets all the headlines. Indeed, there is a temptation right now to frame everything AI-related as a problem that can be solved by an LLM. But for a lot of what we do, traditional machine learning is actually the better tool, and it deserves more recognition.
Authorization rate prediction is a good example. This is a clearly defined, self-contained problem: given a set of transaction inputs (card type, geography, PSP, time of day, amount), predict whether this transaction will be authorized. A well-trained ML model can do this with a level of speed and accuracy that a general-purpose LLM simply cannot match.
We are currently developing a model that does exactly this. Once mature, it will enable something even more powerful: intelligent PSP selection. Instead of routing a transaction based on static rules, the system can evaluate in real time which provider is most likely to authorize it successfully, at the lowest cost. We have already validated the approach on real merchant data and our payments strategy team, along with a few pilot customers, are rolling it out.
Fee prediction follows the same logic. If the model can accurately forecast the cost of routing a transaction through different providers, it opens the door to systematic cost optimization across the entire payment stack.
These are not speculative capabilities. They are specific, measurable problems where ML outperforms both rule-based logic and LLM-based approaches.
LLMs where they make sense: the analytics co-pilot
LLMs can still be transformative for certain use cases. The one I am most excited about is what we are calling the Analytics Co-Pilot.
Payrails already unifies payment data from every integrated PSP into a single dashboard. The co-pilot adds a conversational layer on top: merchants can ask questions in natural language. "What is my fraud decline rate in Brazil?" "How do my processing costs break down by region?" "Which payment method has the highest failure rate in Germany this month?"
The model translates these questions into queries against our unified data layer and returns answers in plain language. No SQL. No dashboard configuration. Just ask, and get an answer.
We are testing this internally right now, and the early results are promising. What is most interesting about the co-pilot is where it could go. If merchants keep asking the same question repeatedly, the system could automatically generate a persistent graph or alert for that metric. The co-pilot becomes a dashboard builder that responds to how people actually use their data, rather than how a product team imagined they would.
This is what we mean when we say that Payrails acts as an extension of the merchant's team. The co-pilot is not a generic AI layer. It is backed by payments experts who know the right questions to ask, the right patterns to look for, and the right way to interpret what comes back. We are encoding that domain knowledge into the AI so that merchants get the benefit of a dedicated payments intelligence function without having to build one from scratch.
Alongside the co-pilot, we are building toward a daily briefing: a proactive intelligence layer where the system analyzes a merchant's data every morning, identifies anything unusual, and generates a summary of what needs attention. Instead of spitting out a wall of charts, it will produce a concise and prioritized briefing written in natural language: "Your authorization rate in France dropped 3% overnight. Here is the likely cause. Here is what you should check."
We are also testing something we find genuinely exciting: combining traditional ML with LLMs in a hybrid workflow. When our anomaly detection model surfaces an issue, we pass the context to an LLM and ask it to dig deeper.
This approach combines the best of both worlds. The ML model is fast and precise at spotting that something is off. The LLM is good at reasoning about why and suggesting what to do next.
Where this is heading: AI agents for payments
Everything I have described so far, from anomaly detection to auth rate prediction, the analytics co-pilot and the daily briefing, are all building blocks. They are valuable on their own, and they are already creating real impact. But they are also the foundation for something larger.
At Payrails, we see the future of payment operations moving through three stages:
- Standardization. Centralize all payments and financial data into one modular hub. This is the platform layer, and it is where we started.
- Optimization. Automate the known problems using high-performance routing, reconciliation logic, and ML models. This is where we are now.
- Autonomy. Deploy intelligent agents to handle the unknowns: the exceptions, the edge cases, the complex multi-step problems that currently eat up hours of human effort.
We are actively building specialized AI agents for three domains where the impact is clearest:
Reconciliation. Matching what was sold against what was recorded in the ledger against what actually arrived in the bank account. Our reconciliation agent learns from manual human actions: every time a team member resolves a mismatch, the agent studies the decision and builds internal rules. Critically, it expresses those rules in natural language, so a human can read what the agent has learned and intervene if something looks wrong. Once validated, these rules can be applied to match transactions at scale. Today, all of this is largely a manual cleanup job requiring days of painstaking human effort every single month. The goal: auto-resolve the routine mismatches (which represent the vast majority) and surface only the genuinely anomalous ones so that finance teams can work smarter.
Fee monitoring. Processing fees change. Acquirers adjust their rate schedules. FX markups shift. Scheme fees evolve. Most merchants discover these changes weeks or months later, buried in settlement reports. The Payrails fee monitoring agent will track fees at the transaction level, compare them against contractual terms, and flag discrepancies in real time. If an acquirer introduces an unannounced fee change affecting thousands of transactions, the agent will catch it, quantify the impact, and present the evidence for a dispute.
Chargebacks. Disputing a chargeback requires collecting transaction histories, IP addresses, delivery confirmations, and other evidence, then assembling it into a coherent case within a tight deadline. Our chargeback agent will automate the evidence collection, assess the strength of the case, and draft the dispute response. The human reviewer can therefore focuses on judgment calls, not data gathering.
In each case, the pattern is the same. The agent handles the repetitive, data-intensive work. The human handles the exceptions and the strategic decisions. We are not replacing finance teams. We are redirecting their expertise toward the problems that actually require human judgment.
The question of autonomy
How far should AI operate autonomously in payments? Here is where it gets interesting, and where I think honest conversations matter more than marketing.
This is not a theoretical question for us. Take our auth rate prediction model for example. It could work in two ways: it could generate recommendations that a human reviews and implements, or it could make real-time routing decisions on every transaction without human involvement.
Both are technically feasible. But they are very different propositions in terms of trust, liability, and customer comfort.
There is a historical parallel worth noting. ML-based cancer detection systems have performed at above-human accuracy for around a decade. Symptom diagnosis tools have been provably better than individual doctors for just as long. And yet adoption in the medical field took years. Patients preferred human doctors, even when the machine was statistically more reliable.
Payments are obviously not healthcare, but the dynamic is similar. Merchants will adopt autonomous AI when they trust it, and trust is earned through transparency, not speed.
That is why the principle of “human-in-the-loop” remains critical. The AI operates, the human supervises. Every agent decision is logged with a clear chain of reasoning. Merchants can see why the system chose a specific routing path or flagged a specific mismatch. And critically, every manual correction the human makes feeds back into the system, so the next similar case requires less intervention.
What we are not doing
We are not bolting a chatbot onto a dashboard and calling it AI. We are not promising full autonomy next quarter. We are not hand-waving about "the future of finance" without showing what we are building today.
We are investing deeply in the foundational data layer that makes all intelligence possible.
We are shipping ML models for specific, high-impact problems where machine learning demonstrably outperforms rules.
We are building LLM-powered interfaces that make complex payment data accessible to anyone.
We are developing specialized agents for the operational workflows where autonomous reasoning can eliminate the most manual effort.
And we are just getting started.
Alex Kagoshima is the Head of Data at Payrails, where he leads the Data Intelligence team building ML models, analytics tools, and AI agents for enterprise payment operations.
I am the text that will be copied.
