Sherrod: Agents are very creative at bringing your infra down. It turns out that agents just like dial up all your failure mode. It just multiplies the amplitude of problems you can get. There were agents that went rogue. There were agents that may have almost taken down core systems. One of the cool things about projects that can be very concrete for people is the idea of tool policies. Let's say you're a person on the HR team who's dealing with a bunch of sensitive information. You really don't want the agent to sort of go rogue and put that sensitive data into some public Google document that all Stripes can access, but you also don't want to tell them, oh, you can't use any tools because your workloads are too sensitive.

Claire: I love this idea of this like three layer triage that a data agent can go through and that's really smart. Find existing reports, then use the analytics layer to find the right query. Then if like you really have to fall down to the data catalog and write your own query. Your data warehouse has to be very resilient to high volume queries because when in doubt an agent will just brute force it.

Sherrod: We could test this out and it's going to kick off what is a human in the loop workflow. Create a calendar invite for me and Wong tomorrow at 11:00 a.m. Pacific. This isn't the fun part. The fun part is what I showed before. Agents are really good. They're very creative. So, we got to put some restrictions on them so they don't go rogue.

Claire: Welcome back to How I AI. I'm Claire, product leader and AI obsessive here on a mission to help you build better with these new tools. Today we have Sherrod, an engineering manager at Stripe and part of the team who built Kai, their internal company brain and company agent. He's going to show us why you might want to build your own custom agent for your company. What are the governance and control mechanisms of Kai that make it super special and how to build not just a single skill, but a whole system.

Sherrod, welcome to the show. Before we get into the live demo, could you give us an overview of what Kai is, why you built it, and how it fits into daily workflows at a company as engineering-forward as Stripe?

Sherrod: Thanks for having me, Claire, super excited to be here. Kai is Stripe's internal AI assistant and platform, essentially what we consider the company's brain and hands. In terms of why we built it in-house, it really came down to three primary pillars:

First is context and permission boundary. Off-the-shelf tools simply don't have deep understanding of Stripe's internal complex landscape—our organizational charts, permission graphs, and data silos.

Second is tool integration and execution. We wanted an agent that goes beyond simple chatting or naive RAG. We needed it to take actions—running queries, spinning up sandbox environments, and driving internal workflows.

Third is governability and extensibility. We wanted every team to easily build their own custom Skills and Projects, while maintaining enterprise-grade guardrails, auditable trails, and cost efficiency.

Today, Kai is natively woven into where employees work daily. From non-technical folks pulling metric dashboards to engineers debugging services, Kai acts as that ubiquitous assistant.

Claire: So just to kind of repeat back what I heard and some of the unique things that you built into Kai is one this sort of bounded context engine. So you know where people sit in the org chart, what they're working on, and their permission scope.

Sherrod: We allow people to select how much they give Kai access to. But out of the box Kai knows like who you are and where do you sit in the org chart, right? And it also knows some helpful global context like what your team does, what services you own, and who you collaborate with frequently. This contextual awareness is progressive, ensuring privacy while eliminating repetitive prompting.

Claire: Great. So, you have this context engine but controlled by the end user to some extent. So, you can decide as an employee how much you want to automatically ingest into the system. You have these deep connections. We'd love to see it in action! Should we jump right into the demo and see how Kai handles a concrete task?

Sherrod: Yeah. Absolutely happy to do it. So I have a few trusty prompts here. What we're going to be doing today is we're actually going to be creating a dashboard because Stripes, we love our data. And so I'm giving Kai a prompt: "Build me an interactive dashboard analyzing Kai's adoption and usage across different orgs." As you can see on the screen, Kai is picking up the prompt and pulling for relevant tools and skills.

Claire: And I see it pulling for tools and skills. So is this like part of the harness which is it's kind of like tuned to go find what it can use to solve this problem?

Sherrod: Absolutely. So there are two things. We all know what tools are and we all know what skills are. Tools basically being the atomic capabilities you can execute, and skills are like essential workflows or specialized recipes.

When Kai gets a request like building a dashboard, it activates that three-layer triage:

1. First Layer: Search existing reports and assets to see if the answer already exists.
2. Second Layer: Use the analytics/semantic layer to find canonical metric definitions and pre-modeled datasets.
3. Third Layer: Fall back to data catalogs and generate custom SQL against raw tables if necessary.

Claire: It looks right. I told it not to make any mistakes.

Sherrod: So, [laughter] that's what these tools and skills are doing. It's safely assembling data and UI components in a controlled environment.

Claire: Yeah. I have a kind of separate question while this is running specifically about making great data agents because I talk to a lot of companies and almost universally the first internal agent they try to build is a data agent, and it often fails. At Stripe's data scale, how do you handle data veracity, messy schema discovery, and hallucination?

Sherrod: I want people that are listening to hear a couple things, and I'm going to make the Stripe team blush. One of the reasons why this works well at Stripe is our deep investment in data infrastructure and data governance pre-dating the AI era.

Agents don't magically fix broken foundations. If your schemas are messy, undocumented, and metrics lack single sources of truth, an agent will just amplify that chaos.

Key things we've invested in:

1. Canonical Metric Layer: Standardized definitions so "active users" or "volume" mean the exact same thing everywhere.
2. Rich Metadata & Catalogs: Clear ownership, descriptions, and verified query templates.
3. Sandbox Execution & Guardrails: Strict timeout, partitioning limits, and compute quotas so rogue queries never take down warehouses.

Sherrod: It turns out that agents just dial up all your failure modes; it multiplies the amplitude of problems you can get. If a human engineer writes a bad query, it fails once. An agent might aggressively retry or brute-force variants dozens of times in parallel, causing massive compute spikes. Resilient infra and circuit breakers are non-negotiable.

Claire: I love it. Okay, so we've yapped while Kai ran. Let's show what Kai actually generated using these skills and tools in Sandbox.

Sherrod: Yeah, of course. So here's what you see. You see that Kai adoption is looking good and pretty much everyone at Stripe uses Kai. We can see high engagement across engineering, marketing, and support, with full interactive slicing and dicing.

Claire: [laughter] I love the fact that our marketing team is like a 100% all-in, right?

Sherrod: They need it, right? I don't know a single marketing person that doesn't either want some sort of app built or some sort of dashboard. So, I think you have product market fit.

Claire: To power this for 10,000 Stripes across the company, that must require a huge dedicated team, right? Do you have like 20 engineers on this?

Sherrod: Definitely not 20 engineers. Stripe runs fairly lean, very nimble, and very fast. Our core team is actually fewer than 10 people. We built the V1 in just a couple of weeks and scaled it across the company. We could do that because we leveraged Stripe's existing robust infra and focused our efforts on the agent harness, context engine, and governance.

Claire: I love it. This episode is brought to you by Hyper Agent, the platform for deploying always on agents that actually run your business. With Hyper Agent, you build agents in the cloud and deploy them securely with integrations into your workflows and data sources.

Sherrod: I'm going to do two things: I'm going to show skills and I'm going to show projects. When people see the dashboard, everyone loves it, but you don't want to re-explain complex prompt instructions every time.

So we built Skills. It should be familiar to folks who've used Cursor or modern AI tooling, but tailored for enterprise collaboration.

And even more crucially, Projects. A project is not just a folder for grouping chats; it is an isolated boundary for permissions, context, and security policies.

Sherrod: For example, if you're on the HR team handling confidential compensation or performance data, you want the agent to use specialized tools within that project, but you must prevent sensitive data from leaking into public Google Docs.

Within Kai Projects, we built Tool Policies:

• Explicitly configure allowed vs. denied tools per project.
• Bound agent execution strictly to that project's sandbox and context.
• Enable high-sensitivity teams to leverage AI safely without security anxiety.

Claire: Yeah, I love this because I think a lot of the existing tools let you maybe configure some of this at the individual level, but then it applies to every session. It's not contexted to what you're working on or the risk profile.

Claire: This is what you see when you do it live. You have kind of a skill creator, a specific spec, and a skill editor that you can test, edit, optimize, and manage.

Sherrod: While engineers love fussing around in Markdown and Python files, having a little IDE-like UI allows everyone across the company to get that quality of life improvement.

Claire: And with so many people creating skills, how do you prevent skill sprawl? For example, if a skill isn't used for 30 days, do you notify the owner and deprecate it?

Sherrod: 100%. You cannot separate quality and quantity when it comes to these systems because context is everything. The more unrelated context you throw into the AI, the worse it performs.

We built automated lifecycle management for skills:

• Monitor invocation frequency and quality metrics.
• Alert owners when skills go unused.
• Unmaintained skills transition to deprecated status and are removed from runtime discovery, keeping the agent's prompt context lean and focused.

Claire: Amazing! I just want to recap for folks the high-level takeaways about Kai:

1. Personalized Bounded Context: Org-chart awareness with progressive disclosure.
2. Three-Layer Data Triage & Sandbox: Existing reports -> semantic layer -> raw catalog, executed safely.
3. Project-Level Tool Policies: Strict security boundaries for high-risk domains.
4. Lean Team Leverage: Serving 10,000 Stripes with fewer than 10 engineers by building on solid infra foundations.

That is an incredible achievement.

Sherrod: That was a fantastic summary. And of course, credit goes to all the underlying platform infrastructure teams across Stripe whose work made this possible.

Claire: Before we get out of here, let's do two lightning round questions. First: putting Kai aside for a second, what is your favorite personal AI workflow outside of core work?

Sherrod: Honestly, I have this whole workflow around using AI to make sure I don't miss important threads or tasks, surfacing key items so nothing slips through the cracks.

Claire: Exactly. We think a lot about how to put agents to work, but I often want agents to put me to work: "Claire, please fill out this form; Claire, please review this."

Sherrod: Yes! For a few more years at least, until they figure out everything we're doing.

Claire: Last question: when AI isn't listening or doing what you want, what do you do? Are you a yeller in prompts?

Sherrod: I might repeat the prompt or occasionally use all-caps, but I don't like shouting at AI. It feels wrong, almost like I'm shouting at the engineers who trained it.

Claire: Like many of our guests, you gentle-parent the AI: "I know you can do better, I believe in you." [laughter]

Sherrod: "I'm not mad. I'm just disappointed, Kai. You should have done better."

Claire: I love this! This has been super insightful and gave me so many ideas for enterprise AI adoption. Where can folks follow you and your work?

Sherrod: Thank you so much for having me, this was a lot of fun. Huge thanks to the audience, and it's been an amazing time!

Claire: Thank you! Thanks so much for watching. If you enjoyed this show, please like and subscribe here on YouTube, or leave us a comment. You can also find this podcast on Apple Podcasts, Spotify, or your favorite podcast app. Visit howiaipod.com for more episodes. See you next time!