When AI Goes Wrong in Banking: Trust and a Kill Switch, with Gagan Singh

In the AI: Trust but Verify podcast, our host, Alec Crawford (@alec06830), Founder and CEO of Verapath (https://www.verapath.com), interviews guests about how they are using AI in business, where you can trust AI, and where you need to put up guard rails.

About Our Guest

This episode's guest is Gagan Singh, Head of Product and Strategy for Enterprise AI at one of the five largest banks in the U.S., who joined banking after years building and scaling products at big tech companies, including Google. Gagan has worked on AI since 2016, from early statistical and rule-based models to voice interfaces in the early days of Alexa, and today leads AI architecture at enterprise scale in a regulated industry. He and Alec discuss where banks are getting real ROI from agentic AI, how to make decades-old systems AI-ready, and why kill switches, traceability and humans in the loop are what will earn AI the trust it needs to scale. They also cover open-source versus frontier models, AI ethics after deployment, and Gagan's advice for tech professionals moving into banking.

Top 5 Themes

• AI takes on back-office work and first-pass fraud review. Banks are getting the most value by automating middle- and back-office operations that already run on well-defined handbooks, such as tier-one review of potential fraud. AI can follow those processes faster and often more accurately, while people handle escalations and decisions.
• Making decades-old bank systems AI-ready. AI is only as good as the data and tools it can reach. Gagan breaks large legacy systems into smaller pieces, modernizes them, and builds a central data layer that AI can rely on, with different views for each line of business.
• The kill switch: governance in a regulated bank. AI use cases pass through cyber risk, model risk, first- and second-line risk and a risk committee that sets how much autonomy each one gets. Every pipeline needs points where a person can stop an outcome that can't be traced back to how the decision was made.
• Three layers of trustworthy AI. Traceability that satisfies risk and cyber stakeholders, clear rules and data definitions that keep models in bounds, and a human in the loop. Tool gates across the pipeline let teams stop a result the moment it drifts, rather than after it is finished.
• Open source vs. frontier models: competition is good for banks. The gap between open-source and frontier models has narrowed sharply. Simpler tasks can run on a cheaper open model, like Llama, inside the bank's own data center with extra guardrails, saving frontier models and their higher token costs for harder work.

Resources

Here are a few of the companies, products and resources mentioned during the episode:

• Verapath: Secure, private AI for banks and wealth managers. https://www.verapath.com
• Claude Cowork: The Anthropic tool Gagan cites for automating enterprise work such as redlining legal documents. https://claude.com/product/cowork
• Gemini Live: Google's voice conversation mode for Gemini, which Gagan uses daily. https://gemini.google/overview/gemini-live/
• ChatGPT: The OpenAI chatbot that kicked off the current wave of AI adoption. https://chatgpt.com
• Llama: Meta's open-source model family, Gagan's example of a model a bank can run in its own data center. https://www.llama.com
• OpenClaw: The open-source, self-hosted AI agent discussed as a consumer example of autonomous agents. https://github.com/openclaw/openclaw
• All-In Podcast: One of Gagan's go-to sources for AI and tech commentary. https://allin.com

Copyright (c) 2026 Artificial Intelligence Risk, Inc. All Rights Reserved