The Future of AI in Payments: Insights from Arpit Mittal on Building Trustworthy Financial Infrastructure

What's Inside: AI is redefining payment infrastructure far beyond fraud detection. Explore how enterprises are building intelligent, explainable, and trusted payment systems, where Agentic AI, real-time decisioning, governance, and modern data architectures are shaping the next generation of financial services through insights from Arpit Mittal, a staff software engineer at PayPal.

Over the past few years, payment companies have largely centered their attention on using AI for fraud detection. While that remains one of its most successful enterprise applications, the technology is now reshaping far more fundamental layers of financial infrastructure. Payment platforms are evolving into intelligent systems capable of making real-time decisions, adapting to emerging threats, supporting regulatory compliance, and preparing for a future where AI agents initiate transactions on behalf of people and businesses.

These shifts bring new architectural challenges. Success no longer depends on deploying the most advanced model. It depends on building reliable data pipelines, explainable decision-making, resilient governance, and trust mechanisms that scale across highly regulated financial ecosystems. Organizations that stop treating AI as a standalone capability and instead build it as core infrastructure are the ones moving beyond isolated AI pilots into production-ready systems.

To better understand where enterprise payments are headed, we spoke with Arpit Mittal, Staff Software Engineer at PayPal, researcher, and inventor whose work spans AI-driven fraud detection, anti-money laundering, federated learning, and agentic payment systems. Through this conversation, he explains why the next generation of payment infrastructure will be defined not only by intelligence, but by explainability, interoperability, and verifiable trust built into every transaction.

Introducing Our Expert

Arpit Mittal

Arpit Mittal is a Staff Software Engineer at PayPal specializing in AI-driven payment infrastructure, fraud prevention, compliance, and large-scale financial systems. His IEEE-published research and patented work span explainable AI for anti-money laundering, federated fraud detection, and agentic payment systems, helping shape the future of trustworthy, intelligent financial infrastructure.

In our discussion, we explored topics including:

• How AI is transforming modern payment infrastructure beyond traditional fraud detection
• The architectural foundations required to scale AI across regulated financial systems
• Why trust, explainability, and Agentic AI will shape the future of enterprise payments

In Conversation with Industry Insiders

Below are the key insights from our conversation with Arpit Mittal on the future of AI-powered payment infrastructure.

1. You've spent over 15 years building payment systems across compliance, fraud prevention, and large-scale financial platforms. Looking back, what has been the biggest shift in how payment infrastructure is being designed today?

The most significant transformation in payment infrastructure over the past decade has been the shift from static, rule-based architectures to continuous, adaptive decisioning systems. Early compliance and fraud-detection systems relied on deterministic rule engines — conditional logic reviewed and updated in periodic batches. Today's infrastructure is fundamentally data-first: built on real-time pipelines, model-driven risk scoring, and closed-loop retraining mechanisms that evolve as fraud patterns shift. Federated learning is emerging as a critical next step for fraud detection, enabling institutions to collaboratively strengthen models without centralizing sensitive data.

Equally important is a change in what infrastructure must verify. It is no longer sufficient to confirm that a human authorized a transaction; systems must now establish who authorized an action, under what conditions, and within what limits. That's a requirement made urgent by the rise of agentic and AI-initiated commerce. This demands that interoperability and explainability be architected into systems from inception rather than added retroactively, as regulators and merchants increasingly expect real-time justification of automated decisions.

Together, these shifts point toward a payments ecosystem where fraud prevention, regulatory compliance, and identity verification are no longer separate, sequential checks, but integrated, explainable, real-time capabilities embedded in the infrastructure itself.

2. Everyone is talking about AI in payments, but where do you believe enterprises are seeing the first measurable business value beyond experimentation?

The clearest measurable value today is in fraud loss reduction, reduced regulatory-fine exposure, and false-positive suppression. Together, these build a more resilient payment platform with fewer disputes and chargebacks, and materially higher approval rates.

The second area is checkout optimization. Fastlane is a good example. It speeds up guest checkout by linking a customer's email to their saved payment methods, so returning guests get password-less authentication and an automatically filled-in form instead of re-entering their details. That same trust infrastructure is what agentic checkout will depend on — a user states what they want to buy, an agent finds it and adds it to cart, the user digitally signs off on that specific purchase, and only then does payment execute. Getting that verification step right is what makes agentic payments fast and trustworthy enough for merchants to actually adopt, not just demo.

Compliance automation is close behind. Explainable AML models are cutting investigation time for analysts, a real, measurable operational cost saving. What all of these have in common is that they're narrow, well-instrumented problems with clear before-and-after metrics, not open-ended AI initiatives.

3. From your experience, which parts of the payment lifecycle are best suited for AI today, and where should human judgment continue to play a central role?

AI is strongest wherever the payment lifecycle involves pattern recognition at scale: real-time transaction risk scoring, anomaly detection, identity verification, and routing decisions. These are high-volume, statistically tractable problems where a model can outperform static rules and do it in milliseconds. Human judgment should stay central anywhere a decision carries legal, reputational, or irreversible consequences, like final adjudication on ambiguous fraud cases, regulatory reporting sign-off, dispute resolution, and, increasingly, approving the policies and limits under which an AI agent is allowed to act on a user's behalf.

I'd draw the line this way: AI should own detection and recommendation, humans should own accountability. Even in agentic payment systems, the goal isn't removing humans from the loop, it's making sure the system can always answer who authorized an action, under what conditions, and within what limits, and that the answer traces back to a human decision.

4. Many organizations are adding AI to existing payment systems rather than rebuilding them. What architectural challenges typically emerge when integrating AI into legacy payment infrastructure?

The first challenge is almost always data, not models. Legacy platforms weren't built to expose clean, low-latency, well-labeled data, so the cost of integrating AI into legacy systems includes significant plumbing work just to get a trustworthy feature pipeline. The second challenge is latency: a checkout or authorization flow may have a few hundred milliseconds to decide, and dropping a large model into that path without careful engineering can break the very experience you're trying to improve. Third is explainability and auditability. Legacy compliance infrastructure was built around deterministic, traceable rules, and regulators expect that same traceability from AI-driven decisions, so you need an explainability layer, not just a scoring layer. And because integrating AI into an existing, live, revenue-critical system means doing all of this with zero tolerance for downtime, nearly every architectural decision ends up favoring incremental rollout, shadow testing, and clean rollback paths over a rip-and-replace approach.

5. AI models are only as good as the data they learn from. What role do data quality, real-time data pipelines, and system integration play in determining whether AI initiatives actually succeed?

This is where most AI initiatives quietly fail, long before anyone blames the model. A fraud or compliance model is only as current as the data feeding it, and payment fraud patterns can shift within days, so a pipeline with even modest latency or staleness will systematically underperform no matter how sophisticated the model architecture is. In practice, real-time, well-integrated data pipelines matter more than model choice.

I've seen simpler models on fresh, well-integrated, AI-ready data outperform more sophisticated models running on stale or fragmented data. System integration is the other half of this: when fraud signals, KYC data, and transaction history live in systems that don't reconcile cleanly, you're not just slowing the pipeline down, you're introducing silent inconsistencies that a model will learn as if they were real patterns. Enterprises that succeed with AI in payments tend to treat data infrastructure as the actual product, with the model as a comparatively swappable component on top of it.

6. We've seen growing interest in Agentic AI and multi-agent systems within financial services. Beyond the hype, where do you see these architectures creating practical value for payment platforms?

The practical value is in delegated commerce. An AI agent booking travel, restocking inventory, or completing a purchase on a user's behalf without a human clicking “buy” at that exact moment. That's a real, growing use case of artificial intelligence, but it exposes a gap in how payments have always worked: authentication and authorization models were designed for a human sitting at a keyboard or holding a phone, not for a delegated agent acting on someone's behalf. That's the specific problem agentic and multi-agentic payment systems address, they are giving service providers a reliable, interoperable way to verify not just that a transaction is happening, but who authorized it, under what conditions, and within what limits, even when no human is present at the moment of the transaction. The most immediate practical value today is in agent-to-merchant transactions for well-defined, bounded, recurring tasks (procurement, subscription management, travel) where the constraints can be tightly specified upfront. Fully open-ended agent autonomy over payments is still further out.

7. When enterprises attempt to scale AI across payment operations, what usually breaks first? What patterns have you repeatedly observed?

Governance breaks first, almost every time, not the model. An enterprise gets a pilot working beautifully on one team's data with one team's assumptions about what “good” looks like, then tries to scale it across business units with different risk appetites, different regulatory obligations, and different data quality, and the model's assumptions quietly stop holding.

The second thing that breaks is monitoring. Teams instrument a model well enough to launch it but don't build the ongoing drift-detection and retraining infrastructure to keep it healthy as fraud patterns and customer behavior evolve, so performance degrades slowly and invisibly until losses start ticking up.

The third pattern is organizational rather than technical: AI initiatives that live entirely inside a data science team, without deep integration into compliance, risk, and engineering from the start, tend to produce models that are accurate but not deployable at the standard a regulated payment platform requires. The common thread is that scaling AI in payments is mostly a systems and governance problem wearing a machine learning costume.

8. Looking ahead, how do you see payment platforms evolving over the next three to five years as AI becomes more deeply embedded into financial infrastructure?

I expect payment platforms to shift from systems that process transactions humans initiate to systems designed natively for a mix of human- and agent-initiated activity, with identity, consent, and trust infrastructure built to handle both from the ground up rather than as an afterthought.

Compliance and fraud detection will keep moving from periodic, batch-based review toward continuous, real-time decisioning that's explainable by default, because regulators are already signaling they expect that level of transparency from automated systems.

I also expect significant standardization and interoperability work, the kind of protocols and trust frameworks that let one company's agent transact safely with another company's payment platform, because right now nearly every organization is solving agent authorization independently, and that doesn't scale. The platforms that win in that world will be the ones that treat trust and verifiable authorization as core infrastructure, on par with uptime and latency, rather than as a compliance checkbox added at the end.

9. If you were advising a financial institution planning its AI roadmap today, what capabilities would you prioritize first to create long-term business value instead of isolated AI pilots?

I'd start with data infrastructure engineering before any model work. A real-time, well-governed pipeline with clear lineage and quality controls, because every AI capability built later inherits the quality of that foundation.

Second, I'd invest in an explainability layer as a shared service rather than something each team rebuilds, that's what lets you deploy AI in fraud, compliance, and credit decisions without a separate slow-moving review process for each one.

Third, I'd build authorization and identity infrastructure that already anticipates agentic use cases (delegated, bounded, verifiable authority) even if agent-initiated transactions are a small fraction of volume today, because retrofitting that later is far more expensive than designing for it now.

Only after those three are in place would I prioritize expanding model capabilities themselves. I

10. Finally, what's one misconception enterprise leaders still have about AI in payments that you believe needs to change?

The biggest misconception is that AI adoption in payments is primarily a model-selection problem. Pick the right algorithm, choose the right vendor, and the value follows. In my experience, the organizations that actually see returns spend most of their effort on data quality, integration, explainability, and governance, and comparatively little on the model itself, which at this point is often close to a commodity.

The second misconception, and one I think will age quickly, is treating “AI in payments” and “agentic AI in payments” as the same conversation. They're not. One is about better predictions on transactions humans initiate. The other is about verifying trust for transactions a system initiates on a human's behalf, which needs fundamentally different non-fraudulent infrastructure, not just a better model.

Leaders who conflate the two will keep solving yesterday's problem well while the more consequential one, trustworthy delegation, goes unaddressed.

Radixweb's Take on the Future of AI in Payment

As Mr. Mittal explains, the future of AI in payments is not about deploying increasingly sophisticated models. It is about building trustworthy artificial intelligence solutions across financial infrastructure. Organizations need real-time data pipelines, explainable AI, resilient governance, and identity frameworks that can support both human- and agent-initiated transactions. As autonomous commerce grows, trust, authorization, and compliance will become foundational capabilities rather than features added later.Across our work with financial institutions and payment providers, we've found the same pattern. The AI initiatives that deliver lasting business value are built on strong integration, high-quality data, and architectures designed for continuous monitoring and regulatory compliance. Production-ready AI succeeds when it becomes part of the payment ecosystem, not another disconnected layer. If you're evaluating how AI fits into your payment platform or modernization roadmap, schedule a no-obligation chat with our AI specialists who can help you plan and build a payments solution that's secure, scalable, and enterprise-ready.