What We Learned at FinNext 2026

FinNext 2026 brought together leaders from banking, fintech, payments, lending, and technology to talk about where the financial services industry is heading. AI, data infrastructure, cloud, and tokenization all got their time on stage, but a few themes kept coming up regardless of who was speaking or what the session was officially about.

Here’s what we took away.

AI Is Moving into Production

The mood has shifted. A year or two ago, most conversations about AI in financial services were about pilots and proofs of concept. At FinNext, people were discussing large-scale deployments, particularly in customer experience, risk management, fraud prevention, and operational efficiency. The excitement is real, but so is the anxiety about what happens when these systems are making consequential decisions in real-world environments.

As expected, agentic AI attracted considerable attention. The idea that systems can reason through complex situations and act without constant human direction is genuinely compelling to people running high-volume operations. But the question that kept surfacing wasn’t whether it works in a demo. It was about whether you could trust it when it mattered.

Our own Dheeraj took the stage with a presentation built around what he called the 90/10 Rule for Decision-Making in the Agentic AI Era, and the booth was busy for the rest of the day. The argument is straightforward: 90% of decisions in financial services need to run on deterministic systems. Transactions, fraud controls, compliance checks, operational workflows — these need to be fast, consistent, and auditable every time. The remaining 10% is where AI earns its keep, handling the ambiguous situations and novel patterns that benefit from contextual reasoning and a human in the loop.

It’s not a framework that undersells AI. It puts it where it actually works, and the practitioners we spoke with recognized it immediately because they’ve seen what happens when that line gets blurred.

Fraud Came Up in Almost Every Conversation

It didn’t matter whether we were talking to someone from a large bank, a fintech startup, or a payments platform. Fraud kept coming up. It cuts across product, engineering, risk, and compliance, and the volume and sophistication of attacks isn’t going down.

What’s striking is that detection isn’t usually the problem. Most organizations have invested heavily in identifying suspicious activity. The gap is between spotting something and acting on it before the damage is done. Fraud that gets caught after the fact still costs money, and it still damages trust with customers who feel like they should have been protected. The question people were wrestling with is how to close that gap at the scale and speed at which modern financial systems operate.

Real-Time Data Infrastructure Has Moved Up the Agenda

We had good conversations with technical leaders in risk, data platforms, and engineering who are actively exploring ways to modernize their real-time processing. These weren’t just passing conversations. These were people who had reached the limits of what their current setups could do and were trying to figure out what comes next.

The questions were specific: 

• What guarantees can you make on stream processing?
• What does latency look like under load?
• How do you maintain consistency when the system is under pressure? 

There’s a clearer understanding now that moving data quickly and making decisions reliably are separate problems, and that a lot of the industry has solved the former without fully addressing the latter. 

Agents Are Only as Good as the Data They’re Working From

One thing that came through consistently was that teams building with Agentic AI are hitting a practical wall. Most architectures give agents access to batch data — historical snapshots from data warehouses that might be hours or days old. When an agent is trying to reason about a live transaction or a risk scenario in progress, it’s working off of information that no longer reflects reality.

The consequences are predictable. Recommendations become unreliable not because the reasoning is bad but because the inputs are incongruent. There’s growing interest in approaches that give agents access to the current operational state as in input for them to work through a decision, rather than making them rely on whatever the last batch job managed to capture.

BNPL Is Creating Risk Problems That Don’t Have Easy Answers

Buy Now Pay Later came up in several conversations as an area where the industry is still figuring things out. Extending credit to people without conventional credit histories is the point of the product, but it also means the usual signals aren’t available. Different data, different models, and tighter decision windows — because the gap between someone applying and expecting to use credit is often measured in seconds, not days.

Most Banks Are Still in Experiment Mode, by Design

Many of the institutions we spoke with are establishing centers of excellence to address questions as discussed above before committing to changes in production systems. It’s a reasonable way to approach it. Financial services don’t have much tolerance for getting infrastructure decisions wrong, so the caution is understandable.

What these teams are often looking for is a way to think clearly about where new approaches fit within what they already have. The 90/10 framing seemed to help with that because it doesn’t ask organizations to throw out what works. It asks them to be clear about which decisions need certainty and which ones can benefit from something more adaptive.

In Conclusion

Walking away from FinNext, it was immediately evident that the industry knows the cost of a bad decision at the wrong moment, or of not making the right decision at the right moment. The pressure to move faster is real, but so is the understanding that speed without reliability is actually detrimental in a business where trust is foundational. Those two things need to be solved together, and that’s the conversation that begs delving deeper in earnest.