Atlassian’s Most Controversial Growth Decision | Mike Cannon-Brookes
Summary
- Atlassian’s addressable market extends far beyond developer tooling: more than half its users have no technical job function. Cannon-Brookes’s category definition is deliberate: “We solve people problems. We don’t solve technical problems.” Jira evolved from a 2002 bug tracker into an issue tracker, road-mapping system, and broad workflow engine connecting technical teams with finance, HR, sales, marketing, and other business teams.
- Cannon-Brookes expects AI to compress business processes, not make workflow software disappear. Agents will remove individual boxes and branches from a flowchart, but their files, links, comments, and outputs still re-enter processes for review, approval, or further automation. His counter-consensus call: AI is “an accelerant, not a replacement.”
- The strategic asset beneath Rovo is a permission-aware “teamwork graph” containing more than 100 billion objects and connections, growing north of 50% quarter on quarter. Built since 2019 from links spanning Atlassian products and external SaaS applications, it powers enterprise search, context, ranking, and agent actions. More than 3.5 million people were already using Atlassian AI features monthly: “We don’t want to market AI; we want to ship it.”
- AI is clearly increasing task-to-code speed, but Atlassian refuses to equate faster code generation with customer value or even realized productivity. Coding consumes roughly one-third of a developer’s week, while generated code shifts effort into rebuilding mental models and reviewing output. With token-heavy tools potentially costing thousands of dollars per developer and tens of millions annually across Atlassian, DX combines quantitative delivery metrics with qualitative surveys to find actual ROI.
- Atlassian’s application-layer strategy is to supply context and select models rather than bet the product on one model vendor. Its production AI gateway runs north of 75 models, and Rovo Dev may use two or three on a typical task, including Claude Code or Gemini. A notable internal use case was not greenfield creation but an API change across more than 500 repositories—“mowing the lawn” at scale so engineers can return to “landscape architecture.”
- Cannon-Brookes rejects both mass developer displacement and the idea that only senior engineers will remain employable. He expects Atlassian to employ more developers in five years, with a continuing economic case for pairing one senior with four or five less-experienced engineers whose output AI can amplify. Accountability remains human: whoever commits or approves code owns it, regardless of whether Rovo Dev, Cursor, Claude Code, or GitHub Copilot produced it.
- The operating doctrine is “grow longer, not grow faster,” even as Atlassian reported roughly 21% revenue growth, 26% cloud growth, and 40% RPO growth. Cannon-Brookes wants durable growth north of 20%–30% five years out, not a quarter optimized by spending every available dollar on immediate revenue. His investor-relevant test is whether the company is seeding future demand or merely harvesting yesterday’s work until “the fields are barren.”
Deep dive
1. Atlassian solves people and coordination problems, not programming problems
Cannon-Brookes corrected Biewald’s developer-centric premise: technical teams account for less than half of Atlassian’s active users. Banks, automakers, and drug companies are technology-driven businesses, but Atlassian’s role is connecting their technical and business teams around the processes through which products and services actually get delivered.
Jira’s history explains the misconception. Its first 2002 version was a bug tracker for developers and product managers; feature requests and road maps turned it into an issue tracker, and richer workflows eventually made it a general-purpose engine used well beyond software development.
The common object is a piece of work moving through people and systems. An IT or DevOps queue may look structurally identical to an HR or finance service queue: work arrives, someone manages it, other participants contribute, and the organization needs to preserve state and responsibility.
2. Agents remove workflow steps rather than the workflow itself
Cannon-Brookes’s disagreement with the prevailing narrative is about scope: “Large parts” of business processes will become faster, more accurate, and higher quality, but he doubts that huge numbers of processes disappear completely. People will still coordinate, track things, and apply their brainpower.
His preferred mental model is a flowchart. Agentic technologies from multiple vendors can eliminate “this box or that box and then maybe a whole branch,” while the surrounding workflow remains—an incremental transformation that applies equally to technical and nontechnical teams.
Atlassian has therefore added the ability to assign work to Rovo agents or outside agents from platforms such as Salesforce, Google Agentspace, Cursor, and GitHub Copilot. When an agent finishes, it may return a file, link, comment, or text for another agent or human to approve, authorize, print, send out, or route onward.
Biewald noted that customers are vibe-coding their own dashboards and asking vendors for APIs or MCP servers. Cannon-Brookes agreed, but distinguished creating technology from becoming a software developer: a finance employee generating disposable Python analysis is solving a problem, not necessarily adopting software engineering as a profession.
3. Enterprise vibe coding still requires a governed substrate
Cannon-Brookes sees broader technology creation as “a huge net positive for society”: marketers can build basic sites and applications, while finance teams can perform analysis that once required programmers. That does not mean programmers and developers disappear, especially when generated systems become durable.
The enterprise constraint is everything surrounding the generated interface—permissions, compliance, governance, regulation, auditing, data residency, and secure access to internal information. Making a public website is one problem; safely connecting an improvised internal application to corporate data is another.
Atlassian’s platform investment follows from that constraint. Enterprise search, organized data, and permission-aware connections are necessary before an agent can participate safely in a workflow: “We can’t build an agent without understanding the data layer underneath.”
4. Rovo turns SaaS links into organizational memory
Cannon-Brookes calls links “the shipping containers of the SaaS trade.” A link carries useful structure before its contents are opened: it identifies a Google Doc, Figma mock-up, GitHub pull request, Salesforce record, or monitoring report and reveals how that object relates to work elsewhere.
Atlassian began constructing its teamwork graph around 2019 from links across its applications and outside SaaS products. It now contains north of 100 billion objects and connections and, according to Cannon-Brookes, is growing north of 50% quarter on quarter.
The difficult work is typing relationships and enforcing permissions. A Jira issue can connect to a pull request, repository, project, business initiative, and spreadsheet; understanding those object types lets the system navigate and rank the graph rather than treat every link as undifferentiated text.
LLMs changed that graph from organizational knowledge into “organizational memory.” Rovo can interpret document content, answer questions across hundreds of SaaS applications, or define internal terms such as “fairy dust” and “alchemists” by reconstructing their company-specific history.
5. AI adoption is measured through activity, not licensed seats
Cannon-Brookes characterized Rovo as potentially one of the world’s largest enterprise search engines—and argued it is the best because Atlassian can combine semantic retrieval with relationship, ranking, and permission context. The same graph improves Jira backlogs, Confluence recommendations, search answers, chat, and agent actions.
Atlassian had disclosed more than 3.5 million monthly AI users. Cannon-Brookes prefers this to customer counts or purchased seats because collaboration without activity “doesn’t make any sense”; 100 licensed seats with five users is a “shelfware gap” likely to become 20 renewed seats.
His product conclusion is that the distinction between AI and non-AI products will disappear. Jira becomes more useful when it can reason across Confluence, Figma, Slack, Google Docs, Microsoft Teams, Zoom, and Salesforce—not because it receives an isolated AI tab.
6. Code speed is not engineering-team productivity
Biewald asked whether developers who feel faster with AI are actually producing more. Cannon-Brookes separated code speed—turning an assigned task into committed, tested, or secure code—from team speed: choosing the right work, shipping it reliably, maintaining quality, and delivering value customers want.
Coding itself represents roughly one-third of a developer’s week in the studies and Atlassian data he cited. Tools have accelerated that portion for decades, from assembly to higher-level languages, cloud infrastructure, JetBrains IDEA, VS Code, and now GitHub Copilot, Cursor, Claude Code, and Rovo Dev.
Atlassian tracks “developer joy” because software creation is a creative profession: productive engineers feel in flow, while broken builds and failing tests make them miserable. AI temporarily disrupts that proxy—generated code feels magical, but developers must reconstruct the mental model they previously formed while writing it.
7. DX makes rapidly rising AI-tool bills accountable
The economics have changed from an IDE that might cost roughly $50 per seat to potentially thousands of dollars per developer in token costs. Across Atlassian, Cannon-Brookes said these tools could consume tens of millions of dollars annually: “I’m happy to spend that if I’m getting value.”
DX combines quantitative measures such as pull-request cycle time with carefully designed qualitative surveys. Cannon-Brookes contrasted its “DX Core 4” framework with DORA metrics and emphasized that neither data type is sufficient alone; together they identify actionable constraints.
His simplest productivity intervention was revealing: ask engineers “what sucks about working here?” If enough people cite a build system that takes hours, the organization has a concrete improvement target even when aggregate delivery numbers appear healthy.
Across an R&D organization of more than 10,000 people, Atlassian tracks money, time, adoption, ROI, and context for Rovo Dev, Claude Code, Cursor, and GitHub Copilot, each used by thousands internally. Cannon-Brookes emphasized that results vary by task as models converge.
8. Context makes code maintenance the near-term killer application
Rovo Dev’s objective is to turn Jira issues into approved pull requests; a merged pull request provides a strong signal that the code is good. Its team is also building self-reflection so the agent can improve and avoid repeating mistakes across old, new, and differently structured repositories, much like graduates learn through code review.
The teamwork graph supplies both technical context—prior issues, pull requests, repositories, and prior solutions—and product context from Confluence, Google Drive, or Figma. Atlassian’s production gateway contains north of 75 models, while a typical Rovo Dev task may invoke two or three, sometimes including Claude Code or Gemini.
Cannon-Brookes’s best example was an internal service whose URL and request shape changed. More than 500 repositories required configuration or code edits; the team used Rovo Dev to write and commit a few examples, then agents used those examples and the graph to find analogous changes.
Blindly generating 500 pull requests merely transfers the bottleneck to human review. A small initial error can compound through a thousand automated loops into an “AI worm” problem, so Cannon-Brookes insists on repeated human involvement rather than AI-written code reviewed only by another AI.
9. Humans supply the knowledge that gives agents direction
Cannon-Brookes phrased the limitation cautiously: AI “generating net new knowledge is a pretty tricky thing.” It can transform accumulated organizational and personal memory extremely well, but product ideas, customer insights, and genuinely new direction still require information and judgment held by people.
His alternative to the “magic button” is an agent behaving like an intern: it completes a pass, returns for feedback, and tries again. The productive unit is therefore the human-AI collaboration loop, not an agent receiving a task and disappearing until the work is supposedly finished.
In his whiteboard example, an agent searches permissioned material across Teams, Google Docs, Jira, and Confluence to create an initial brainstorm. Five or six people then add sticky notes, vote, reject suggestions, and discuss the results—the human signals from which another agent can draft a memo, flowchart, or Confluence page.
A person can record a Loom over that draft, adding previously unwritten knowledge; AI can then summarize the video for time-constrained colleagues. Repeating the loop converts tribal knowledge into organizational memory while preserving the human contribution that Cannon-Brookes believes will remain indispensable.
10. More AI could mean more developers—including juniors
Biewald supplied the counterpoint: junior developer hiring appears harder, their traditional tasks look increasingly automatable, and demand seems to be rotating toward senior engineers. Cannon-Brookes acknowledged uncertainty—“I might be completely wrong”—but rejected both that conclusion and mass developer displacement.
He expects Atlassian to employ more developers in five years, echoing a similar prediction he heard from AWS. Senior skills will evolve as they did from on-premises to cloud systems, and he sees a continuing economic case for pairing one experienced engineer with four or five less-experienced people who can become substantially more capable with AI.
The non-negotiable is accountability: committing generated code makes the developer responsible for it, and approving a pull request makes the reviewer responsible. “Rovo Dev wrote that” or “Cursor wrote that” cannot excuse defects or security failures.
Cannon-Brookes also argued that juniors may introduce new working methods before seniors do. Loom spread partly through younger, video-native employees questioning synchronous meetings; he offered “20% useful, 80% unnecessary” as an illustrative ratio, while noting every business could choose its own number.
11. Atlassian optimizes for duration, not the fastest revenue headline
Cannon-Brookes celebrated how cloud infrastructure, payments, SaaS, and app stores have accelerated distribution, but warned of survivorship bias: technology circles celebrate companies reaching $100 million ARR rapidly and ignore the many going “from zero to zero.”
Around 2010, Atlassian was eight or nine years old, producing roughly $50 million in revenue and $20 million in profit while growing about 50%; approximately $100 million in revenue was then considered the threshold for going public. Founders missing today’s compressed milestones should first ask whether customers value the product and distribution works.
Atlassian instead aims to “grow longer, not grow faster.” Its latest quarter grew revenue roughly 21%, cloud revenue 26%, and RPO—remaining performance obligations, or contracted future revenue not yet included in 12-month revenue—40%, with more than a billion-dollar-and-change of quarterly revenue.
The “$10 game” captures the allocation problem: spending all $10 on next quarter’s revenue leaves nothing for architecture, surviving AI disruption, or other long-term investments. Cannon-Brookes wants decisions capable of sustaining growth north of 20%–30% five years out, citing Amazon’s long investment in infrastructure and logistics as the canonical pattern.
12. Sustainable growth requires measuring seeds as carefully as harvests
Monthly active users are only the first layer. Atlassian also examines “MCU,” which Cannon-Brookes called “monthly collaborative users” but said he did not know what the acronym stood for, to distinguish Confluence readers, creators, commenters, reactors, and linkers.
Product-led funnels track trials, seven-day usage, D30, 2D30, and return behavior, while enterprise growth also depends on IT gatekeepers evaluating security, compliance, governance, controls, analytics, and rollout. Each subgroup has a different definition of value and a different reason to block adoption.
Sales pipeline presents the same time-horizon choice: this quarter’s results were seeded in a multiyear pipeline perhaps two years earlier. Cannon-Brookes’s retired co-founder framed management as balancing seeding and harvesting—cutting or harvesting faster than resources regenerate eventually leaves a barren field.
The deeper test is whether a business lives off the “interest” of its accumulated resources or unknowingly consumes the principal. AI may generate more reports, but humans still interpret them, decide what needs attention, and collaborate; Cannon-Brookes closed by calling curiosity and creation fundamental to remaining in technology.