Rowing or Steering: How Atlassian's PMs Became AI Builders

Yehoshua recalls that a year or two ago everyone was pointing at someone else's craft and predicting it would vanish: PMs told engineers, designers told PMs, and everyone was nervous. That talk has faded. What she hears now is that jobs overlap, and that an "AI builder" role is real, with startups hiring builders rather than separate PMs and designers.

The harder question is what this looks like in a company of more than 10,000 people, which is the setting she focuses on, speaking at the Lenny and Friends Summit. She says roles are overlapping but also expanding, so everyone can do far more than before. The PM's mission of finding product-market fit and building a business people pay for is unchanged; AI tools simply change the how.

Asked how to stay on top of AI developments, Yehoshua says you cannot, and advises not believing everything you read on X. Focus on what actually works for your customers. She also points to context as a major shift: Atlassian's teamwork graph gives AI tools the context of the organization, so the PM no longer has to take meeting notes, chase action items or tell go-to-market when things launch.

She shares an anecdote in which Atlassian's CEO asked when a feature would launch and, five minutes later, posted a screenshot of Rovo's answer and asked if it was right. It was, and he apologized for wasting their time. People now have access to this information, she says, but they still need help adopting it, so PMs are in a front-row seat for change management.

Atlassian thinks about acceleration as a ship heading for a destination: intelligence from the models plus context from the organization. Each PM must decide whether they are rowing or steering. After running experiments with different team structures and measuring outputs, the company's answer is that it depends on the product and its phase.

The first case is Confluence's Remix with Rovo and Confluence Slides, a new feature in an existing codebase. A PM who had never written code or used a terminal worked with an engineering partner who built her a harness, and she checked in 26 PRs in a month, more than most engineers on the team. Her reasoning was that there were many UX fixes and too few engineers, and she wanted engineers focused on higher-level work.

PMs also owned evals for Slides, relaying how the team should evaluate the product and getting 2x eval throughput. They used Arize to find prompt issues and send them to engineers faster. The team used the Figma MCP to map designs to code and had a coding agent fix about 14 design bugs in an hour, while test creation went from half a day to 10 minutes.

Remix launched in six weeks and Slides in eight, work that Yehoshua says would have taken about six months before AI, and in enterprise software that must be compliant and secure. The PM was rowing, helped by an engineering manager who provided a clean, isolated repository and a team that deliberately blurred roles and questioned existing processes.

The second example is Rovoclaw, a zero-to-one project distinct from the Rovo product. PM Josh and designer Kevin vibe-coded the whole thing to a working alpha used internally, and only then were engineers added. Soon Josh noticed the engineers veering off track while he was busy coding, and asked whether coding was still his highest-leverage activity.

He stepped back to set direction, prioritize and unblock, and things sped up. Having started by coding, he understood the blockers and challenges, which made him a better steerer. He also built an agent inside Rovoclaw to write the product's own weekly updates, so he no longer writes them by hand: the PM began by rowing and shifted to steering as the phase changed.

Jira is more than 20 years old, spans data center, cloud, isolated cloud and FedRAMP, and serves hundreds of thousands of customers, so Yehoshua says PMs checking in code would be dangerous. Instead, PMs and engineers built AI-first features that let customers assign work to agents. In the period from in progress to shipped, throughput was about 3x, and the team shipped 22 user-facing features in about 10 weeks.

Prototyping started with Loom: record the UI or Figma designs with a talk track, auto-create work items, and moving one to in progress kicks off a managed coding agent in the cloud. The output is compliant, accessible and in the Atlassian design language, which makes handoff to developers faster. Feedback from Slack was triaged by the Jira agent and sent to a coding agent, and more than 900 pieces of customer-study feedback were categorized by a Rovo agent with surprisingly good quality.

Here the PM steered, focusing on unblocking engineers, processing feedback and building reusable prototypes. Overall, PMs now prototype, test, write evals, dig deeper into customer feedback and consume far more information, while no longer doing manual updates, hand-compiled research or slides from scratch. All of it depends on the AI tools understanding organizational context via the teamwork graph.

Where startups say to hire AI natives, Atlassian wants its existing strong PMs to become AI native. Its AI fluency index covers six capabilities, such as using tools, automating data insights, prototyping and technical literacy, scored from 1 (curious) through 3 (capable) to 5 (pioneering). It is a development guide and not a promotion ladder; the company expects L3 everywhere over time and an L5 where the team most needs it.

The training approach that stuck is AI builder week, held once a quarter, with external speakers, advanced internal PMs teaching peers, and a hands-on project. More than a thousand people have attended and over 120 new workflows have come out of it. Each week has a theme: prototyping, then evals, then building agents, then checking in code. Atlassian has published an "AI builder week in a box" for other companies.

On measuring results, Yehoshua is candid that Atlassian has not figured it out and thinks no one has. They track PRs deployed to production rather than written, features delivered from idea to usage, OKRs and overall throughput, looking at both teams and the organization. She closes as an optimist: she thinks this is the best time to be a PM.