TL;DR
The short version
At enterprise scale, crafts overlap into AI builders but do not collapse. The PM job stays the same; the how changes with tools and org context.
Tamar Yehoshua (Atlassian CPO) shows three cases — Confluence (PM rowing/code), RovoClaw (row then steer), Jira (steer only) — plus an AI Fluency Index and quarterly Builder Weeks to upskill the PMs you already have.
Based on the Lenny's Podcast upload Roles aren't converging—they're expanding | Tamar Yehoshua (28 Sep 2026).
Overlap without erasure
A year of "your craft dies" talk cooled into something more useful: roles overlap, and each can do more. Startups hire builders and skip PM or design seats. Inside Atlassian, Yehoshua focuses on what PMs actually do when the company has thousands of people, compliance, and existing craft roles.
The roles are overlapping, but they are also expanding.
Tamar Yehoshua
PMF, loved products, a business people pay for — unchanged. Note-taking and status chase-downs shrink when Rovo and the Teamwork Graph answer the CEO's "when does this ship?" in minutes. Tip: you cannot track every model; do not believe everything on X; watch what works for customers.
Three cases, three answers
Confluence remix/slides — row. A PM with no terminal experience landed 26 PRs in a month via an eng-built harness, drove evals (~2×), auto-fixed design bugs with Figma MCP, and helped cut cycle time from ~6 months to ~6–8 weeks for enterprise-safe launches.
RovoClaw zero-to-one — row then steer. PM + designer vibe-coded to internal alpha, added eng, then the PM stepped off the keyboard to prioritize and unblock when the team veered — better steerer because he had rowed first.
Jira AI-first — steer only. Too dangerous for PM production commits in a 20-year codebase. Loom→work item→cloud agents, Slack triage, 900+ research clips categorized — cited ~3× throughput and 22 user-facing features in ~10 weeks.
Fluency for the PMs you have
Hire AI-native is a startup slogan. Atlassian built an AI Fluency Index (six capabilities, L1–L5) as a development north star — not a promotion ladder — and quarterly AI Builder Weeks (1,000+ people, 120+ workflows retained). Themes rotate: prototyping, evals, agents, checking in code.
Measurement remains foggy: PRs deployed, features to customer hands, usage, OKRs. Nobody has nailed AI velocity. The operating question stays simple: for this product, in this phase, does the ship go faster if you row or if you steer?
