Mike Cannon-Brookes on Atlassian's AI Strategy

The worst interface in modern software may be the one every company is now building its future on: a blank white box, inviting a user to type anything and promising it can do everything. Mike Cannon-Brookes, the Australian co-founder and CEO of Atlassian, the twenty-four-year-old company behind Jira and Confluence, thinks this is a problem the industry has not solved. He said as much on The Twenty Minute VC with host Harry Stebbings, in a conversation that keeps circling back to one idea: as AI makes it easier to build software, the thing that makes software work gets harder.

The Brainless Bet

Cannon-Brookes opens with a flat rejection of a popular idea in enterprise software circles, that products should go "headless," stripping out the interface and letting AI agents interact directly with the underlying data. "Headless is brainless by definition," he says (05:22). If a company gives up on its interface, he argues, it has given up on solving the customer's problem at all.

"If you're giving up on your user interface, if you're giving up on evolving your experience for a customer to solve the problems that you set out to solve for them, if you are not a significant part of that experience, yes, I think that's a brainless bet." - Mike Cannon-Brookes [05:22]

This is not a rejection of agents. Atlassian has, by Cannon-Brookes's account, one of the top three to five MCP servers in the world (05:48). MCP, short for Model Context Protocol, is the open standard that lets AI agents read data from a company's software and write actions back into it. Atlassian leans into it. But Cannon-Brookes separates the plumbing from the experience: agents can fetch information, yet most enterprise work, he says, involves "taste and interaction and collaboration" that a pipe cannot replace (03:58).

The design argument gets sharper when he turns to the tools people use today. Chat interfaces, he says, hand users "unlimited power" through "a blank one-line text box," and most people respond by treating a large language model like a search bar (11:24). That is a design failure, not a user failure. He thinks the last time the industry had to solve a problem this foundational was the arrival of mobile phones, when touchscreens forced every app to be rethought from scratch (10:06). Questions he says remain open: how should an agent announce itself inside a shared workspace, how should people tell AI-touched content apart from human-written content, and does that distinction even matter (10:27).

Underneath this is a competitive claim. Cannon-Brookes estimates that a brand new app today can beat ninety percent of existing software on design with modest effort. In five years, the same app will only beat twenty percent, because user expectations keep rising (06:52). If building a product is getting cheap, he reasons, the thing that remains expensive, and valuable, is making it feel good to use.

From Information to Imagination

Two more threads run through the conversation, and both are about trust and cost rather than taste. The first is geopolitical. Cannon-Brookes says Atlassian hears real pushback from large customers about using models associated with China, even though those same customers show no objection to open weight models in general (17:29). Atlassian runs Nemotron, a model family built with Nvidia, and uses some Mistral models, but keeps Chinese built models out of production entirely (17:49). He separates the geopolitical question, who controls the technology, from the technical one, whether an openly published model can be trusted because its weights can be inspected. As a comparison, he points out that roughly forty percent of contributions to Linux, the open source operating system, now come from China, and almost nobody treats that as disqualifying (18:28).

The same skepticism shows up when he talks about "specialized models," the current venture pitch that a narrow, purpose built model beats a general one. "The jury is really out on the benefits of building specialized models at the moment," he says (20:52).

"The jury is really out on the benefits of building specialized models at the moment." - Mike Cannon-Brookes [20:52]

His objection is partly semantic and partly empirical. Semantically, he argues that most companies claiming to have "built their own model" have actually taken an open weight model, often Chinese, and fine-tuned it for a narrow task, a much cheaper undertaking than training from raw data (22:19). Empirically, he says general foundation models have improved fast enough to erase the price-performance edge specialized models had at launch (23:59). The exceptions he grants are real: domains with large proprietary datasets, like robotics, where a specialized approach still makes sense (21:14).

Cost discipline, in his telling, is becoming a core skill for any serious technology company, in the same way financial discipline is. He calls it FinOps, short for financial operations applied to cloud and AI spending, and says managing it well will be "a core strength and attribute of a great technology company" (26:35). Atlassian routes every AI feature internally, blending multiple models to balance price, quality, and speed for each request, rather than relying on third-party routing tools (28:04). Internally, he tells product teams not to worry about inference cost while a feature is small. "If five people are using a feature, I don't care how much it costs. If five million people use a feature, I'm going to care a lot about cost" (29:37).

That distinction between production and consumption becomes the episode's clearest thesis. "Right now, the production of technology is faster than the consumption of technology. We've reached a tipping point," Cannon-Brookes says (31:29). Companies can now build features faster than customers can absorb them into daily work. The constraint on AI's usefulness is no longer supply. It is attention and imagination. He credits Didier Elzinga, the founder of the employee feedback company Culture Amp, with a phrase he has adopted: the economy is moving "from a world of information asymmetry to imagination asymmetry" (32:57).

"We're moving from a world of information asymmetry to imagination asymmetry." - Mike Cannon-Brookes [32:57]

Knowing something other people do not used to be the advantage. Now, he argues, the advantage belongs to whoever can think of a genuinely new way to apply a technology everyone already has access to.

That idea explains Atlassian's bet on what it calls the Teamwork Graph, a system meant to link a company's documents, code, chat messages, and the people behind them into a single layer that AI agents can query (36:08). Cannon-Brookes makes a small, telling point about where organizational knowledge actually lives: "How many times are PowerPoint slides the place where people put actual facts and figures? A lot" (36:38). Formal databases, in his account, often miss the decisions that matter most, because those decisions get made in a slide deck and never migrate anywhere else.

The graph's trickiest property, he says, is that it expires. "The social graph of a company is valid today, and only today," because people join and leave constantly, changing who works with whom (38:04). The system has to behave like a search index, inferring likely answers in advance rather than looking things up live, while constantly refreshing itself to stay current (37:09).

It is worth noting what this system is built on. Confluence, Atlassian's documentation product, generates more than $1.5 billion in annual recurring revenue and was still growing faster than twenty-five percent a year as of the company's last disclosed figures, a scale that would make it a strong public company on its own (42:13). Cannon-Brookes is candid that writing text is not, by itself, a defensible product; there are, he notes, roughly twenty ways to write a document on any given computer already. What sustains Confluence's growth, he argues, is the same thing he opened the conversation with: continuous, unglamorous investment in design and integration, not the existence of one clever feature.

The conversation closes on a smaller, more personal note that still fits the theme. Cannon-Brookes, twenty-four years into running Atlassian with ten thousand employees (45:13), says his current struggle is not motivation but time, carving out space for "deep, long thinking" in a culture that rewards instant, fragmented action (34:00). Stebbings worries aloud that venture capital has made founding a company too easy, pointing out that raising a fifteen million dollar pre-seed round is no longer unusual (48:43), which lets people take on the founder role without building the discipline that makes companies last. His advice, half to Stebbings and half to himself, matches everything else he said: slow down enough to notice what is actually worth building, because the parts of the job that are genuinely hard, taste, trust, and patience, are the parts AI has not made any easier.