The banking sector is currently navigating a paradox of immense opportunity and systemic apprehension. While the market for AI agents in financial services is projected to reach $6.54 billion by 2035, the industry is simultaneously grappling with a staggering 33,125% surge in search interest for “AI bank risks.” As financial institutions (FIs) race to move these systems into live customer workflows, the technical reality of safe deployment has become the industry’s most pressing hurdle.
Neal Lathia, Co-Founder and CTO of Gradient Labs, argues that the path to widespread adoption lies not in avoiding regulation, but in building systems that exceed it. Drawing on a decade of experience in AI and a tenure at Monzo, Lathia sat down with The Fintech Times to discuss the five principles of safe deployment and how to solve the “black box” problem that keeps compliance officers awake at night.
The Transparency Mandate
The primary friction point for any bank executive considering agentic AI is the lack of transparency. Traditional Large Language Models (LLMs) are often viewed as black boxes—systems where data goes in and an answer comes out, but the reasoning remains opaque. Lathia maintains that the highest leverage way Gradient Labs has addressed this is by ensuring the agent is not a closed loop.
“Our agent harness—the code that shapes how the AI agent runs—is built in such a way to keep a strict set of decision traces that can be inspected, understood, and replayed,” Lathia explained. By binding non-deterministic LLMs to specific, narrow tasks, FIs can track not only what the agent did, but the exact logic it followed to reach a conclusion. This level of granularity provides the audit trail that regulators and internal risk committees now demand.
Surpassing the Human Benchmark
One of the most debated topics in AI deployment is how to prove a system is “production-ready.” Lathia suggests that the bar is set by the existing human experience. To exceed this, Gradient Labs leverages internal quality assurance processes to benchmark AI performance against human agents.
“If your goal is to deliver a transformative experience using AI, then the bar is set by the experience you’re currently delivering with human agents,” Lathia commented. To move into production, an agent must demonstrate it can meet or exceed human metrics in accuracy and compliance. This isn’t just about speed; it’s about ensuring the AI can handle the sheer volume of contact reasons inherent in banking. While an e-commerce platform might face 10 distinct customer queries, a bank deals with an order of magnitude more, requiring a significantly higher degree of nuance and reliability.
Navigating Criminal Liability: The Tipping Off Risk
In the UK, “tipping off” a customer about a suspicious activity report (SAR) or an ongoing investigation is a criminal offence. For an AI agent, which pulls from vast amounts of internal data, the risk of inadvertently revealing a sensitive status is a technical nightmare for compliance officers.
Lathia noted that it is almost impossible to prevent an AI from being exposed to information that could lead to a tip-off. “The technical challenge is that even if the AI agent does not have actual access to the state of account investigations, it might still gather enough context to inadvertently tip off a customer,” he said. To mitigate this, Gradient Labs has built an independent, auditable control that runs on all agent output. This secondary guardrail acts as an automated compliance officer, scanning every response before it reaches the customer to ensure no sensitive investigative details are leaked.
Extracting Truth from History
A common fear among Chief Risk Officers is that grounding an AI in historical data will cause it to inherit past human biases or outdated procedural errors. Gradient Labs addresses this through a specialist onboarding agent that extracts “knowledge snippets” or facts from historical conversations.
However, these facts are not simply accepted at face value. “These facts need to be substantiated across multiple conversations and absent from the rest of the AI agent’s knowledge in order to qualify for inclusion,” Lathia added. This process is reinforced by a human-in-the-loop system, where human operators approve and edit facts, ensuring that while the AI learns from the past, it isn’t doomed to repeat its mistakes.
The Control-Plane for the Boardroom
For the C-suite, the success of an AI deployment is measured by its impact on risk appetite. Lathia identifies three “control-plane metrics” that should be reported to the board to prove a system is operating safely: resolution rates, customer-reported satisfaction, and specialised metrics like complaint volumes that capture outcome failures.
These metrics allow a Chief Risk Officer to monitor the system’s health in real-time, aligning AI performance with regulatory expectations. By focusing on these high-level outcomes, banks can transition from viewing AI as a risky experiment to a stable, scalable utility.
Looking Toward 2035
As the industry eyes the multi-billion dollar opportunity of the next decade, the question remains: is the hurdle technological or cultural? Lathia believes the two are inextricably linked.
“I believe that great technology does not subvert regulation; it is supercharged by it,” he concluded. As AI moves further up the value chain and begins to replace traditional human labour in financial decision-making, the role of the regulator will become even more vital in protecting the customer experience. For banks, the winning strategy will not be finding ways around the rules, but building the transparent, auditable, and nuanced systems that make those rules easier to follow.
