- Most of the fintech tooling on display analyzes, scores, alerts and recommends, but very few tools decide while the transaction is still in flight; a fraud model that flags a payment 400 milliseconds after it settles has documented a loss rather than prevented one.
- Test one is whether the decision happens inside the transaction path at transaction speed; if a vendor cannot show you that, you are looking at a reporting tool wearing a decisioning costume.
- Test two is whether the system acts on live transactional state or stale aggregated data, because a generic model demoed on synthetic data reflects the vendor’s average customer rather than your institution’s actual risk.
- Test three is whether it survives real-time rails volume, since FedNow, SEPA Instant and UPI push latency and throughput expectations that batch-era clearing infrastructure was never designed to meet.
- A 90/10 model works best: a deterministic system executes the majority of transactional decisions and escalates to agents for recommendations, with human approvals recorded deterministically and promoted into known scenarios once the evidence crosses a threshold.
TL;DR
FinovateFall wrapped in New York last week: three days, dozens of live demos, and an agenda dominated by AI in nearly every form. If you are a bank or payments executive who attended, or watched the coverage, you came away with a long list of things that looked impressive on stage.
There are a few questions that have to be asked:
- How many of these innovations are chatbots disguised as a tool for a human operator?
- Which innovations require data to leave the banks’ infrastructure i.e. compromise sovereignty and open up a new surface area for data loss?
- Which of these technologies sit in the prevention path vs post-mortem path?
- Which processes require scale vs correctness or low latency? Are there any that require any combination or all three?
- What happens to these systems or processes when an adverse event like a hardware or network failure happens? What is the plan for business continuity?
After two decades of watching financial infrastructure decisions succeed and fail, I use three tests to separate real value from demo-stage hype.
Test one: do the decisions happen in the transaction path, or after it?
Many of the tools on show this year analyze, score, alert, and recommend. Very few actually decide while the transaction is still in flight. That distinction is important. A fraud model that flags a payment 400 milliseconds after it settles has not prevented anything; it has documented a loss. A late decision is a wrong decision. If a vendor cannot show you the decision happening inside the payment path, at transaction speed, you are looking at a reporting tool wearing a decisioning costume.
Test two: does it act on your live transactional state, or stale aggregated data?
The clearest signal in this year's analyst coverage of fraud and financial crime is that AI adoption is no longer the differentiator. It is the baseline. The institutions pulling ahead are the ones whose models act on their own live data, because fraud patterns at a regional bank look nothing like those at a global processor. A generic model demoed on synthetic data tells you what the vendor's average customer looks like. Your risk lives in the difference between that average and your reality. Ask where the decisioning layer gets its state, and how fresh that state is at the moment of decision.
Test three: can it survive real-time rails volume?
FedNow, SEPA Instant, and UPI are pushing transaction volumes and latency expectations that batch-era clearing house infrastructure was never designed for. A capability that performs beautifully at demo scale often degrades exactly when it matters most, at peak volume, when fraud attempts spike and settlement windows disappear. The question is not whether it works. It is whether it works at the millisecond and the million-transaction mark simultaneously.
The pattern behind the tests
All three tests come down to the same architectural point. Intelligence that arrives after the moment of transaction is commentary. What banks need is an operational intelligence layer that holds live transactional state and acts on it inside the payment path without deteriorating the consumer’s experience. This becomes even more important as agentic AI enters the conversation: I advocate a 90/10 mental model where a deterministic system executes the majority of the transactional decisions and escalates to agents for recommendation. These recommendations then go through a process of systemizing them eventually:
- Human “approver” decides whether the recommendation is accepted/rejected or even an alternate decision is taken
- This choice by the human approver is then recorded deterministically and fed into learning systems and when the empirical evidence crosses a threshold, it then gets promoted to a known scenario from being an edge case scenario.
- This would create an adaptive learning environment that actively and autonomously manages false positives/negatives.
- This process adjustment will not work without a decisioning layer fast enough to sit in the path.
The innovation on display at FinovateFall is real. The gap between a compelling demo and a production decisioning capability is also real. Apply the three tests, and the shiny objects sort themselves out quickly.
Explore how Volt Active Data approaches real-time decisioning for financial services at voltactivedata.com/bfsi.
About Author
Featured Resources
