Intercom says 90% of PRs AI‑authored

Intercom says its engineering team has reached a point where about 90% of pull requests are authored by Anthropic’s Claude Code. (ideas.fin.ai) The claim appeared in a March 20, 2026 Fin Ideas post by John Moriarty, who wrote that the figure is part of an internal “2x” effort with an explicit goal of doubling productivity with artificial intelligence. (ideas.fin.ai) A pull request is the packet of code and explanation engineers submit for review before changes are merged. Claude Code is Anthropic’s coding agent, which can read a codebase, edit files, run commands, and connect to developer tools. (github.blog ) (code.claude.com) Intercom’s own engineering blog shows the scale of the environment where that system is being used. In a January 26, 2026 post, the company said its Ruby on Rails monolith is its core application, that it averages 180 ships per workday, and that merged code reaches production in about 12 minutes. (intercom.com) That matters because the company is not describing artificial intelligence as a side tool for autocomplete. Intercom has spent the past two years repositioning itself as an “AI-first” company in product strategy, customer support, and internal operations. (intercom.com) Intercom has also tied its customer-facing products closely to Anthropic. In October 2024, it said Fin 2 was powered by Claude, and Anthropic’s customer case study says Intercom uses Claude to help more than 25,000 customers resolve support queries, with resolution rates of up to 86%. (intercom.com) (claude.com) A separate March 19, 2026 Fin Ideas post, summarized by AI Primer, described the internal coding setup in more detail. It said Intercom runs Claude Code with 13 plugins, more than 100 skills, and hooks that turn the model into what the company calls a “full-stack engineering platform.” (ideas.fin.ai) (ai-primer.com) That setup reportedly includes a read-only production console for the Rails application exposed through Model Context Protocol, a standard for letting models call outside tools. AI Primer said the system lets Claude inspect production data for checks such as feature flags, business logic, and cache state while using read replicas, blocked tables, Okta authentication, and audit logs as safeguards. (intercom.com) (ai-primer.com) The same report said Intercom uses hooks to stop Claude from opening a pull request without first generating the business rationale, and background agents to watch continuous integration results after code is submitted. Those details suggest the company is automating not just code generation but parts of review, testing, and release management around it. (ai-primer.com) (intercom.com) Intercom has not published a formal methodology for the 90% figure in the sources reviewed here, and the available detail comes from company-owned channels and a secondary summary of a staff engineer’s post. But taken with Intercom’s published shipping cadence and Claude-heavy product strategy, the number points to a company trying to move artificial intelligence from assistant to author inside a live production workflow. (ideas.fin.ai) (intercom.com)