What are the key takeaways from this The TWIML AI Podcast episode?

Key insights include: **Multi-agent trigger criteria:** Deploy multi-agent architecture only when a problem contains multiple distinct user intents that cannot be resolved by a single deterministic model. Capital One's Chat Concierge required separate agents for intent disambiguation, planning, governance validation, response accuracy checking, and final response formatting — each with a narrowly scoped task.; **Risk-first platform layering:** Separate agent governance into two distinct layers — platform-level enterprise policies covering cyber, compliance, and guardrails that apply automatically at runtime, and domain-specific policies that individual teams layer on top. This split lets developers focus on agent design while the platform enforces mandatory regulatory boundaries without manual configuration per deployment.; **Latency as a product feature:** Treat end-to-end latency as a first-class product requirement, not a non-functional afterthought. In multi-agent systems, latency must be measured across every agent boundary, tool invocation, and model call simultaneously. Capital One uses smaller specialized fine-tuned models via teacher-student distillation to hit latency targets while maintaining personalization quality.

What did Rashmi Shetty discuss on The TWIML AI Podcast?

Rashmi Shetty, Senior Director of Enterprise Generative AI Platform at Capital One, explains how the company built and deployed Chat Concierge, a multi-agent car-buying system, and outlines the platform strategy enabling developers to build governed agentic systems at scale across the enterprise. Key topics include: **Multi-agent trigger criteria:** Deploy multi-agent architecture only when a problem contains multiple distinct user intents that cannot be resolved by a single deterministic model. Capital One's Chat Concierge required separate agents for intent disambiguation, planning, governance validation, response accuracy checking, and final response formatting — each with a narrowly scoped task.; **Risk-first platform layering:** Separate agent governance into two distinct layers — platform-level enterprise policies covering cyber, compliance, and guardrails that apply automatically at runtime, and domain-specific policies that individual teams layer on top. This split lets developers focus on agent design while the platform enforces mandatory regulatory boundaries without manual configuration per deployment..

How long is this episode of The TWIML AI Podcast?

This episode is 54 minutes long. SignalCast provides an AI-generated summary so you can get the key insights in about 3 minutes.

The TWIML AI Podcast

How Capital One Delivers Multi-Agent Systems with Rashmi Shetty - #765

April 16, 2026

54 min episode · 2 min read

Rashmi Shetty

Episode

54 min

Read time

2 min

Topics

Productivity, Leadership, Design & UX

AI-Generated Summary

Published Apr 17, 2026

Key Takeaways

✓Multi-agent trigger criteria: Deploy multi-agent architecture only when a problem contains multiple distinct user intents that cannot be resolved by a single deterministic model. Capital One's Chat Concierge required separate agents for intent disambiguation, planning, governance validation, response accuracy checking, and final response formatting — each with a narrowly scoped task.
✓Risk-first platform layering: Separate agent governance into two distinct layers — platform-level enterprise policies covering cyber, compliance, and guardrails that apply automatically at runtime, and domain-specific policies that individual teams layer on top. This split lets developers focus on agent design while the platform enforces mandatory regulatory boundaries without manual configuration per deployment.
✓Latency as a product feature: Treat end-to-end latency as a first-class product requirement, not a non-functional afterthought. In multi-agent systems, latency must be measured across every agent boundary, tool invocation, and model call simultaneously. Capital One uses smaller specialized fine-tuned models via teacher-student distillation to hit latency targets while maintaining personalization quality.
✓Closed-loop observability design: Instrument agentic systems to capture production failure signals and route them back into the experimentation environment for prompt tuning, model fine-tuning, retrieval adjustment, or context management updates. Design this feedback pipeline before deployment, not after, because production telemetry is where the largest performance gains originate in agentic systems.
✓Beachhead use case selection: Choose the first production agentic deployment from a high-surface-area, low-risk scenario to safely observe real failure modes at scale. Capital One selected an auto dealership customer experience rather than a core banking workflow, generating architectural patterns and observability baselines that informed the broader enterprise platform strategy.

What It Covers

Rashmi Shetty, Senior Director of Enterprise Generative AI Platform at Capital One, explains how the company built and deployed Chat Concierge, a multi-agent car-buying system, and outlines the platform strategy enabling developers to build governed agentic systems at scale across the enterprise.

Key Questions Answered

•Multi-agent trigger criteria: Deploy multi-agent architecture only when a problem contains multiple distinct user intents that cannot be resolved by a single deterministic model. Capital One's Chat Concierge required separate agents for intent disambiguation, planning, governance validation, response accuracy checking, and final response formatting — each with a narrowly scoped task.
•Risk-first platform layering: Separate agent governance into two distinct layers — platform-level enterprise policies covering cyber, compliance, and guardrails that apply automatically at runtime, and domain-specific policies that individual teams layer on top. This split lets developers focus on agent design while the platform enforces mandatory regulatory boundaries without manual configuration per deployment.
•Latency as a product feature: Treat end-to-end latency as a first-class product requirement, not a non-functional afterthought. In multi-agent systems, latency must be measured across every agent boundary, tool invocation, and model call simultaneously. Capital One uses smaller specialized fine-tuned models via teacher-student distillation to hit latency targets while maintaining personalization quality.
•Closed-loop observability design: Instrument agentic systems to capture production failure signals and route them back into the experimentation environment for prompt tuning, model fine-tuning, retrieval adjustment, or context management updates. Design this feedback pipeline before deployment, not after, because production telemetry is where the largest performance gains originate in agentic systems.
•Beachhead use case selection: Choose the first production agentic deployment from a high-surface-area, low-risk scenario to safely observe real failure modes at scale. Capital One selected an auto dealership customer experience rather than a core banking workflow, generating architectural patterns and observability baselines that informed the broader enterprise platform strategy.

Notable Moment

Shetty reframes Capital One's competitive AI advantage not as model sophistication but as data infrastructure built over a decade. The argument is that specialized fine-tuned models only outperform general ones when enterprise-grade data pipelines already exist — making prior data investment the actual prerequisite for agentic success.

Know someone who'd find this useful?

Episode Transcript

Capital One's tech team isn't just talking about multiagentic AI. They already deployed one. It's called Chat Concierge, and it's simplifying car shopping. Using self reflection and layered reasoning with live API checks, it doesn't just help buyers find a car they love. It helps schedule a test drive, get approved for financing, and estimate trade in value. Advanced, intuitive, and deployed. That's how they stack. That's technology at Capital One. To learn more about AI at Capital One, visit capital1.com/tech/ai. We moved from a classic ML world to a world where we have LLMs, generating responses. And now we want to move on to a world where actions need to be taken. And when the problem that we are working on is a complex one, that's where multi agentic comes into play. Alright, everyone. Welcome to another episode of the Twimble AI podcast. I am your host, Sam Charrington. Today, I'm joined by Rashmi Shetty. Rashmi is senior director of enterprise generative AI platform at Capital One. Rashmi, welcome to the podcast. Thank you, Sam. Thanks for having me here. Thanks so much for joining us. We're gonna be digging into the why and how of Multi Agentic AI at Capital One. To get us started, I'd love to have you tell us a little bit about your personal journey. How'd you get to where you are? Yeah. So, I think I think my personal journey is something many many folks might might relate to. It's it's it's it's, funny how you ask me about my personal journey in a day and age where agentic AI is in its explosive form. So we have stopped pausing to think about it. How did we get here? So yeah. I mean, yeah. It it's it's it's been more or less organic and and natural evolution in my specific case, but but if you ask me how can I pin it to a specific phase of my life, I would go back to my my thesis, my thesis work? The the topic of my thesis was around pervasive computing. Very interestingly, it was around, perceiving your environment, getting, signals from your environment, taking context aware decisions in real time and having actuators trigger actions to change that environment based on your your specific goals. It was a physical world in which this was manifesting. But in essence, if I think about it, the principles were very much the same. It's it's about distributed intelligence. It's about intelligence being embedded in the system that you're living and operating in. And, if you if you think about the governing principles where where we were working on a design at that point in time where which my thesis topic was around was specifically around, decisioning systems, real time intelligence. And that's where where we are today. I think moving from academia to to the industry, just kind of taught me how to scale that in an enterprise in an enterprise world. So the next phase …

Get the full transcript (8,206 words) + summary by email — free

One-time email with the complete transcript and AI summary of this episode. No account needed.

One email, no spam. We’ll also show you what SignalCast does.

Browse all The TWIML AI Podcast transcripts →