Scuttlebutt

Scuttlebutt

Capital One SVP on fraud prevention: "80, 90% of that is because of the way we use AI"

8 lessons: why a 2016 cloud bet lets Capital One scale AI agents, why Llama can't do its fraud modeling, and how its Macaw framework checks an agent before it acts

Scuttlebutt's avatar
Scuttlebutt
Nov 21, 2026
∙ Paid

Eye On A.I. · Nov 20, 2026 · 39m · 31m saved

With Abhijit Bose, Senior Vice President, AI/ML Platform Engineering (MLX) at Capital One

Capital One SVP on fraud prevention: "80, 90% of that is because of the way we use AI"

AI accounts for 80 to 90% of what Capital One calls one of the best fraud-prevention rates in the industry. Abhijit Bose, Senior Vice President of AI/ML platform engineering at Capital One, gave that figure on Eye On A.I., in an interview about why most companies cannot get AI agents past the pilot stage.

Most enterprises still build AI one project at a time, one business line at a time. Abhijit says Capital One's lead comes from a decision made about a decade ago to move the whole bank onto the cloud and build shared platforms, and that a company starting from zero now faces a hard problem.

"We consider ourselves as a tech company that does banking instead of the other way around." — Abhijit Bose

Abhijit runs MLX, Capital One's AI and machine-learning platform. A four-digit number of employees use it every month, and the applications built on it reach about 100 million customers.

The full interview is covered here so you can skip it. 39 minutes of audio, 8 minutes of reading.

Here are the 8 lessons that matter.


This Scuttlebutt Summary is available in full to premium subscribers. If you're not one already, please consider subscribing to keep reading.


Key Takeaways

  • Capital One says AI accounts for 80–90% of its fraud-prevention rate

    • Some fraud models are updated daily or hourly with no engineer in the deployment loop

  • A company can run five agents without a platform, but not 20,000, in Abhijit's view

  • Capital One's AI scale traces to a cloud migration begun around 2016 and 2017, when cloud economics were worse than today

  • Quarterly loss forecasting now takes about a week, down from weeks of training tens of models

  • All 20,000-plus call-center agents use a generative AI lookup tool, after staff asked to be enrolled

  • Capital One uses Llama for generic tasks but builds its own models for fraud, where open models lack the data

  • Its Macaw framework adds a validation agent that checks each action before it is taken for a customer

1. Platforms, Not Pilots

Abhijit says the technology behind enterprise AI largely works. What separates companies is whether they built governance, monitoring and orchestration for agents before they needed them.

Abhijit says a company can deploy a handful of agents without central infrastructure, but not thousands

"If you think of, oh, I'm going to have five agents in the company in the next year, you might be able to do that without a platform. But if you are envisioning a future where you will have 20,000 agents running in your company... that is a hard problem to solve." — Abhijit Bose

Running agents across customer-facing and internal work means registering, monitoring and coordinating each one centrally. Abhijit gave the same reason for why isolated pilots stall.

"Not having a standard way to orchestrate these applications, not having the platforms, makes it really hard to scale them." — Abhijit Bose

2. A Cloud Bet From 2016

User's avatar

Continue reading this post for free, courtesy of Scuttlebutt.

Or purchase a paid subscription.
© 2026 The Golden Age LLC · Privacy ∙ Terms ∙ Collection notice
Start your SubstackGet the app
Substack is the home for writing, podcasting, and media founders.