TL;DR
The short version
Geoff Charles argues winning with AI is not asking people to code faster — it is finding each bottleneck in identify → define → build → review → test → coordinate → improve, and redesigning the factory around it.
Ramp's internal agents (insights, Glass, Inspect, Review Buddy, Testo, Gadget) show the pattern; PMs then split into factory builders, taste-makers, or GMs. Treat Ramp's internal percentages as operator claims.
Based on Geoff Charles's talk on Lenny's Podcast (published 25 Sep 2026; Lenny & Friends Summit).
The system is the limit
Charles opens with racing: the best drivers only perform up to the level of the system. Pit crews cut tire changes from roughly a minute in the 1950s to under two seconds by removing bottlenecks — specialized roles, better tools, practice — not by working 37× harder. F1 even refreshes ~90% of a ~16,000-part car year to year. That is the bar he wants for software factories.
AI simply removes the bottleneck but moves it.
Geoff Charles
Hunt bottlenecks stage by stage
The clock starts at customer pain and stops when a product solves it. Coding got cheaper; define, review, test, and coordination became the drag. Ramp's answer: customer-insight agents over siloed data; Glass for definition wired to Snowflake, research, strategy, and codebase; Inspect for Slack-native coding with deploy previews; Review Buddy for most PR reviews; Testo for browser QA; Gadget so every recurring question becomes an API against a legible org record.
Automate small loops; staff for judgment
Charles warns PMs against living on the dopamine of tiny reactive wins — and points at Claire Vo's same-summit ambition talk. Autonomous loops should route, dedupe, rank, code, and test small UX issues with a thin human gate so people spend judgment on large bets.
Three PM futures: factory PM (remove org drag / ship the harness), taste-maker (hold the bar AI cannot), and GM (own business outcomes across functions). Constraints can force a world-class dimension; bottlenecks should be deleted. Obsess a little less over the next feature and a little more over the factory that builds it.
