Arthur Azizov, founder of B2BROKER Group and B2BINPAY, analyses AI deployments in financial services and their impact on legacy fintech infrastructures.
These days, 81% of financial industry companies are deploying AI in one form or another, which is, on one hand, a moment of truth for the technology. It shows that AI has moved from its experimental stage to a real finance business case. On the other hand, a recent analysis of 15 million customer interactions shows that AI in finance has actually fragmented operations. Even though a small group of fintechs reached 70% automation on complex workflows, most are stuck below the 25% threshold.
Chat support, internal tooling, and document processing are procedures where AI performs best, because the technical environment is relatively forgiving. If you connect the same systems, say, to payments, know your customer (KYC) workflows, or compliance engines, they’re most likely to fall apart. The infrastructure around them wasn't built for this, and to this day, it isn't ready either.
So, in this piece, let’s break down where exactly the infrastructure fails, and what the next phase of AI adoption in finance requires.
Legacy payments infrastructure: built for the past, expected to support present behaviour
What's currently happening with payments infrastructure is caused by the systems banks and fintech platforms run on, which weren't built for real-time operation. They were primarily designed to close the books at the end of the day, process checks overnight, and reconcile everything by morning. Although this has worked fine for decades, with AI's rapid breakthrough, it is suddenly no longer the case.
Now, a model that screens a transaction for fraud or, for instance, prices a trade in real time, needs current-to-the-second information, as hours-old data might just as well be a week old. As a result, AI can look flawless in isolation, yet make mistakes when it's connected to a live environment, where every weakness is exposed.
Let’s look at the KYC procedure to clarify what this means. The fraud detection systems that many companies use were, in fact, devised for older types of fraud: forged documents, stolen identities, and other ‘classic’ schemes. Now, they’re forced to confront AI-generated attacks, and the main problem is that many of these systems haven’t fully caught up even with traditional fraud schemes, let alone AI’s ability to generate multiple synthetic identities within seconds.
Well, this doesn’t mean that there's something wrong with AI. On the contrary, it tends to perform exactly as designed, as the underlying infrastructure wasn’t prepared to absorb a technology that evolves this quickly.
Rebuilding the payment rails
To some extent, the industry has already begun to recognise this issue and act on it. According to KPMG's Pulse of Fintech report, global fintech investment rebounded sharply in 2025, reaching USD 116 billion, up from USD 95.5 billion in 2024. Although the numbers are solid, it’s not fully clear how much of it is actually going toward rebuilding infrastructure.
The point is that if investment goes mainly into AI models, pilots, anything like this, when everything around them remains underfunded (from data pipelines, compliance logs, audit trails, monitoring, and reconciliation to how humans interfere for cases the system cannot resolve by itself), scaling will only exacerbate the problem.
That’s why running AI in production isn't the same as stress-testing its potential. This is especially true for mid-sized fintechs, for which AI testing is cheap but deploying it at scale is not. The companies that do manage to scale already have an established infrastructure layer, such as standardised data feeds, stable APIs, and controls built to support automated decisions for live volumes. The next phase of AI adoption in finance starts with infrastructure; flashy interfaces don’t matter if the technology isn’t supported as needed.
Final thoughts
AI has already proven itself, sometimes working even better than most expected. Still, everything it must run through is, so far, left exposed: ledgers built for future reconciliation and review queues sized for backlog caseloads. None of it works in real time.
The USD 116 billion invested in fintech last year will either go toward fixing that, or toward more tools that pile on top of it, idling away. It’s all up to the builders.
About the author
Arthur Azizov is a serial fintech entrepreneur and the founder of B2BROKER Group and B2BINPAY. With over 15 years of experience in fintech and financial markets, he focuses on building technology-driven solutions for broker-dealers, financial institutions, and trading firms worldwide.
About B2BROKER Group
B2BROKER Group is a global fintech powerhouse providing liquidity, trading technology, and brokerage infrastructure to financial institutions. Founded in 2014 and headquartered in Dubai, the company serves brokers, exchanges, and hedge funds across Europe, the Middle East, and Asia. For more information: https://b2broker.com/
