Developing Governed Agentic AI for Financial Operations

Financial institutions are discovering that deploying agentic AI across core workflows is less a model problem than an operational engineering problem. Emerj’s Yolandi de Weerdt explored this shift with Yoav Naveh, Co-Founder and Co-CEO of Reindeer, and Ajay Swamy, Senior Executive Product Director for GenAI Products, AI/ML Platform Management and Governance at JPMorganChase, in a three-episode series examining how regulated institutions move from isolated experiments to supervised autonomous systems.

Why BFSI Workflows Resist Automation

The operational bottleneck is structural. According to the CFA Institute, 90% of enterprise data is unstructured, and BFSI environments pull disparate document formats from core databases, risk engines, and document repositories. Heterogeneity without governance inflates error rates in document-heavy processes such as Anti-Money Laundering (AML) and Know Your Customer (KYC).

Compliance pressure compounds the problem. Research highlighted by the ProSight Financial Association puts the average annual cost of non-compliance at $14.82 million, 2.71 times the cost of maintaining compliance infrastructure. Meanwhile, analysis published by the IEEE Computer Society found that while 83% of technology leaders start AI projects, only 9% successfully operationalize them. Research published by MR Online reports that current AI task execution achieves an average success rate of 30% on complex end-to-end workplace processes.

Redesigning Workflows Before Adding Autonomy

Naveh frames automation as a structural challenge. A single step is rarely difficult; ten simple steps spanning five systems, three teams, and inconsistent document formats become something else entirely, with exceptions multiplying at every seam.

He and Swamy outline a sequence: mirror the workflow exactly as humans run it today, digitize it end-to-end to expose hidden dependencies, map every system, data store, and judgment checkpoint, then stabilize exception handling as a first-order requirement. Only after that baseline holds should institutions layer in new capabilities such as public-record enrichment or dynamic document validation. Swamy adds that workflows fail at the seams, where fragmented data, inconsistent entitlements, and shifting policy interpretation collide.

Governance as a Precondition, Not a Safeguard

Swamy argues that governance determines whether agentic AI can be used at all. "A number or a decision that you cannot trace is a number or a decision that you cannot defend," he said, adding that the technology is the easy part because demos look great; what institutions actually evaluate is whether they can supervise it and explain every step end-to-end.

Naveh notes that institutions often treat accuracy as the primary risk when the real exposure emerges during uncertainty or when work inevitably changes. Agents must escalate ambiguity rather than mask it, since silent failure is more dangerous than an explicit handoff. Together they set four supervisory conditions: surface full decision lineage, make human judgment checkpoints explicit, detect divergence in real time, and govern agent evolution through controlled versioning, regression testing, and approval cycles.

What to Watch

The National Institute of Standards and Technology has emphasized that reliable deployment requires continuous human oversight and explicit fallback mechanisms when model confidence declines. For BFSI leaders, the open question is no longer whether agents can execute tasks, but whether institutions can engineer the systems map, judgment map, and exception architecture that make autonomous execution defensible.