TL;DR
The short version
Stripe put an agent it calls Kai in front of 10,000+ employees, most of them non-engineers. What made it land wasn't the model — it was two unglamorous things: governance boundaries that set the rules once per workflow, and years of data-platform investment built for humans that held up for agents.
Engineering manager Sherrod's framing on How I AI: the hard problem was never providing AI, it was providing the governance structures so everyone can use it and trust it will do the right thing.
Built on How I AI with Claire Vo and Stripe's Sherrod (7 Sep 2026). Stripe's own Kai metrics are self-reported and vary across sources — treat them as directional.
The failure mode is amplification, not intelligence
Agents don't fail by being dumb. They fail by being effective in the wrong direction. Stripe saw agents brute-force queries, and some went rogue or nearly took down core systems before the team hardened things. Sherrod's line for it: an agent "just multiplies the amplitude of problems you can get."
So the first design job isn't a smarter prompt. It's containment. Stripe's data warehouse had to survive agents hammering it — infrastructure hardening plus early work on "agentic identity," declaring what an agent is and what job it is doing so the system can prioritize and shed load. None of it is glamorous. All of it is why the demo doesn't end in an incident.
Governance is a shared boundary, not a per-user setting
The clever primitive is what Stripe calls a project: a boundary drawn around a workflow that sets the rules for everyone inside it. A project picks a default model (often a cheaper one, to cap token spend), scopes which tools are available, and marks which actions need a human to confirm before the agent acts — writing to a shared doc, say, or sending an invite.
The reason this matters is who makes the decision. A few people who understand the cost, latency, and safety trade-offs configure the project once. Everyone else inherits it and just works. Most tools let you set permissions per person, but then they apply to every session and can't be shared across a team. Stripe scoped the controls to the job instead.
Sherrod's argument against the usual approach is really an argument about human behavior: "If you put too much friction in front of people, they're just going to do unsafe things, because that's how humans are." Prompt someone to approve every tool in every session and eventually they click the wrong button, so approvals are reserved for genuinely sensitive workflows, not applied globally.
The moat was built before AI arrived
The most transferable lesson is one leaders can act on today. Stripe's edge with agents came from platform work done years earlier for human efficiency: a resilient query layer on Trino, a tiered "blessed" analytics layer, and a searchable data catalog. The tiering — existing reports first, the analytics layer next, the raw catalog only as a last resort — is what guards correctness, because an agent can answer the question but has no idea whether it was the right query or the right table.
These investments were made for humans but have held up really well for agents.
Sherrod, Stripe
Host Claire Vo drew the operator's conclusion: when people ask how to ship more with AI, she doesn't talk about product — she says double the size of your developer-experience team and your data team. Agent leverage is downstream of platform quality. If the foundation is weak, agents amplify the weakness.
What to copy
Stripe built its first version with roughly 1.5 engineers in about two weeks, and a company-wide demo triggered the adoption ramp. (Stripe's own published account says one engineer in one week — treat the exact figures as directional; they're self-reported and don't fully match across sources.) The point stands: a thin first version, shown to people, answered questions no memo could.
Three moves worth stealing:
- Contain before you enable — assume the agent will amplify your worst failure mode, and harden that path first.
- Set rules at the workflow level — one knowledgeable owner configures model, tools, and approvals for a job; everyone inside inherits it.
- Fund the platform — the data catalog, query resilience, and dev tooling are the real prerequisites. Agents make good foundations look magical and bad ones look dangerous.
Sherrod was careful to say Stripe hasn't figured this out — it is in the earliest parts of the journey. The playbook isn't finished. But the order of operations is clear: govern and invest first, then let people loose.
