Photo By: Armand Khoury
For years, “human-in-the-loop” has been one of the easiest assurances for financial institutions adopting artificial intelligence: Let the machines do the work, but keep a person in the process.
The concept sounds straightforward. It is also becoming increasingly difficult to execute at scale.
As banks and capital markets firms experiment with agentic AI capable of taking actions across complex workflows, simply adding a human checkpoint may do little to reduce risk. In some cases, it can create a new problem: overwhelming employees with AI-generated recommendations, alerts and decisions until oversight becomes little more than a rubber stamp.
That tension is becoming more apparent as financial institutions move beyond AI pilots and confront the operational realities of deploying these systems across high-volume environments.
Rahul Naithani, Field CTO for Banking and Financial Services at NewRocket, argues that the industry needs to rethink what human oversight is supposed to accomplish.
“Slapping a human figurehead into a workflow isn’t a control strategy.”
The issue is particularly acute in data-intensive operations. Analysts, underwriters and other financial professionals can quickly become bottlenecks if they are expected to review raw AI outputs, gather supporting information and perform routine validation themselves. Instead of augmenting expertise, poorly designed human-in-the-loop systems can simply shift repetitive work from machines back to people.
Alert management presents a similar challenge. Financial institutions already process enormous volumes of alerts, many of which ultimately prove insignificant. Routing every potential exception to a human may appear conservative, but an endless stream of low-value reviews can produce alert fatigue. Over time, the very control designed to catch unusual activity can become less effective as reviewers struggle to distinguish meaningful exceptions from routine noise.
The alternative is not necessarily less human oversight. It is more selective oversight.
Routine data gathering, baseline analysis and low-risk triage can increasingly be automated, while people concentrate on cases requiring context, judgment or accountability. Automated decisions should also remain subject to defined controls and detailed audit trails, allowing institutions to reconstruct what happened and why.
Customer-facing AI introduces another layer of complexity. Banks deploying systems across chat, contact centers and digital channels need consistent rules governing how those systems interpret customer intent and respond. Without common, pre-approved logic, the same customer could receive different answers depending on the channel, creating operational, regulatory and potentially fair-lending concerns.
That makes governance an architectural issue, rather than something that can be bolted onto an AI application after deployment.
For financial institutions, this also means connecting AI workflows to existing Model Risk Management processes and establishing clear boundaries around what an agent can recommend, execute or escalate. Human reviewers need enough context to understand an AI-generated recommendation, sufficient time to assess it and the authority to reject or override it.
Perhaps most importantly, those interventions need to be recorded. A defensible human-in-the-loop framework should capture automated actions, human decisions and overrides in an auditable trail that can withstand scrutiny from risk committees, internal auditors and regulators.
“The person in the loop actually has to have the context, time, and authority to exercise real judgment,” Naithani said.
That distinction may prove critical as agentic AI moves deeper into banking operations. The question is no longer whether a human appears somewhere in an AI workflow. It is whether that human is positioned to make a meaningful difference.
For banks, genuine human-in-the-loop governance may ultimately have less to do with adding another approval step and more to do with designing AI systems so that human judgment is deployed where it matters most.
NewRocket is the trusted AI partner helping enterprises build AI workflows customers trust. As a launch Select Partner in Anthropic’s Claude Partner Network, we help organizations unlock the value of AI through strategy, governance, adoption, and delivery. Combined with ServiceNow’s workflow platform and more than two decades of enterprise transformation expertise, we design, build, and scale AI solutions that create outcomes that last.
About Rahul Naithani
Rahul Naithani – Field CTO, Banking and Financial Services.
He brings three decades of deep-domain expertise in architecting and delivering high-stakes technology solutions for major global financial institutions.
Throughout his 30-year career, Rahul has been a strategic partner to large banks, helping them achieve capital management efficiency, improve regulatory posture, generate sticky customer experiences, and solve complex operational leverage problems. His deep expertise spans the entire financial ecosystem—including Risk and Compliance, Trade life cycle, Post-trade operations, and Asset/Wealth management across both consumer and institutional banking.
At NewRocket, Rahul bridges the gap between technology architecture and business economics. He works cross-functionally with Sales, Delivery, and Technology teams to engineer and package robust, specification-driven offerings tailored for the modern financial services market. His focus centers on driving immediate client value through data management, legacy modernization, platform engineering, and AI-led transformation.
Rahul holds an MBA from NYU Stern and is based in East Brunswick, New Jersey.
Show Comments