Stripe's Internal AI Platform Hit 83% Adoption. The Math Doesn't
Software / analysis
Stripe's Internal AI Platform Hit 83% Adoption. The Math Doesn't
Kai grew from 296 to 5,000 users in four weeks, Stripe says, but Hacker News readers found its sales-activity and copy claims didn't line up on Sept. 23.

Stripe says an internal AI agent platform called Kai went from 296 users to more than 5,000 in about four weeks and now reaches 83% weekly adoption across the company. Hacker News readers who found the company's account on Sept. 23 spent the day checking whether the numbers around that claim actually held together.
Stripe first described Kai, short for Knowledge AI Platform, in an Aug. 3, 2026 post on its developer blog, which said the tool entered open preview and hit its quarterly adoption target within a week. By the time Stripe wrote the post, weekly active use had reached 83% company-wide, including sales, finance and customer-success teams the company said had felt left behind by coding-focused AI tools. The post drew renewed attention when it hit Hacker News on Sept. 23 as "Stripe built its internal AI platform," gathering 173 points and 108 comments in its first 11 hours.
Who built it, and how fast
Anupam Upadhyay, a Stripe staff software engineer, built the first version of Kai in a single week, according to Stripe's account, before an Agent Foundation team led by engineering manager Sharadh Krishnamurthy took over. The platform runs on LangChain's LangGraph framework using Deep Agents as its underlying harness, a choice Krishnamurthy credited for cutting the scope of what Stripe had to build itself: "The Deep Agents layer solves all the non-Stripey problems, so that we can focus on solving the Stripey agent problems," he said, according to LangChain's own account of the build. Kai connects to more than 1,000 internal tools and skills across over 100 teams, uses filesystem middleware backed by Amazon S3, and integrates with Stripe's data warehouse, Slack and Google Suite.
| Kai metric | Stripe's figure |
|---|---|
| Users at launch | 296 |
| Users after roughly four weeks | More than 5,000 |
| Weekly active adoption | 83% company-wide |
| Total sessions reported | More than 60,000 |
Where the Hacker News thread pushed back
The top comment on the Sept. 23 thread was not about the technology. "Unnecessary AI copy throughout the interfaces like 'Browse, discover, and manage skils for your agents,'" wrote commenter quadrifoliate, adding that the product had an "inconsistent, AI-sloppy look-and-feel with different typefaces spattered across the interface." Commenter chrisvls said they "had to give up reading the document because there were too many paragraphs that contained lots of words but no additional detail."
The more specific objection concerned Stripe's own numbers. Multiple commenters said a claimed doubling in sales activity did not match a separately reported gain in sales opportunities: one said "2x the sales activity" alongside only "17% more opportunities" is a mismatch Stripe's post does not explain, since a metric that truly doubled should move a downstream outcome metric by more than 17 percentage points unless the two figures measure different populations or time windows. Another asked, of the claim that "most of Stripe was using" Kai within two weeks, "by force or by choice?" It is a fair question Stripe's own post does not answer, and it matters for reading the 83% figure: adoption a company mandates and adoption people choose are different signals wearing the same percentage.
The engineering Stripe says it didn't have to build
Krishnamurthy's framing of Deep Agents as the layer that "solves all the non-Stripey problems" points to a specific set of components neither post disputes. Kai's architecture includes sandbox middleware for running generated code, summarization middleware for keeping long multi-turn conversations inside a model's context window, and a federated skills system that loads more than 1,000 internal tools dynamically rather than holding them all in memory at once. That is the part of the build Stripe says took an outside framework rather than a week of one engineer's time, a distinction the company draws clearly even as its adoption math goes unexplained.
What would confirm the number
Stripe's post treats weekly active use as the headline metric, but weekly active use counts a login or a query, not a task finished without a human redoing it. A figure that would actually confirm 83% adoption is meaningful is a completion rate: how many of Kai's more than 60,000 sessions ended with the requester accepting the output as-is, versus escalating to a colleague or ignoring the answer. Stripe's post does not report that number, and neither does LangChain's account of the build, even though LangChain's post goes into detail on the sandbox and summarization middleware that make long sessions possible in the first place. Until one of the two companies publishes a completion or correction rate alongside the adoption figures, the 83% describes how often people opened Kai, not how often it did the job they opened it for.
What Stripe has not disclosed
Stripe's account does not say whether weekly adoption is measured against every employee or against a smaller group with Kai access already granted, which would change what 83% means considerably. Neither the Stripe nor the LangChain post gives an error rate, a task-completion rate, or any measure of how often Kai's output required a human to redo the work, the same category of gap this site found in Anthropic's own enterprise AI push when Claude for Financial Advisors was pulled a week after launch. Stripe has not said whether Kai's adoption figures were reviewed by anyone outside the team that built it, and no independent usage audit of the platform has been published as of Sept. 24. The pattern of a vendor's own account outrunning what its numbers support is one this site has flagged in developer-tool launches before, where a star count and a token-cost complaint told two different stories about the same project; Stripe's gap between adoption and impact is smaller, but it is the same shape.
Sources
More in Software
- 01Rivet, a YC W23 Company, Runs Agent Actors on S3 and Claims 72 KB EachThe Apache 2.0 orchestrator says an actor cold-starts in 4.99 ms at p99. Every figure is Rivet's own, and the Launch HN thread questioned both the premise and the AWS bill.
- 02GitHub Actions Lockfiles: What gh actions-lock Pins and What It Leaves LooseWorkflow lockfiles went generally available on October 28. The extension's own README lists four gaps, including Docker images and reusable workflows in other repositories.
- 03C++26 Reportedly Fails Its ISO Ballot, With Contracts the Sticking PointSecondary reports put the October 23 vote at 12 yes of 20 and nine no votes of 22 bodies. The numbers do not reconcile, and no ISO or WG21 record has been published to settle them.
- 04A Postgres Benchmark Found jsonb Read One Key 2.7 to 10.6 Times Faster, and Larger on Disk in 6 of 11 CasesChris van Eijk of Now-Next loaded 1 million small documents and 5,000 large ones into PostgreSQL 17.11 and 18.6. The speed result is consistent. The storage result depends on whitespace and the 2 kB compression threshold.