The SaaSpocalypse was supposed to go like this. AI agents read your data directly, build their own middleware, and your expensive software subscriptions become a database with a login screen.
Atlassian just grew around 30 percent in a quarter on a $7 billion run rate. That is the fastest it has grown in about two years.
Mike Cannon-Brookes, Atlassian's CEO, sat down with Nilay Patel on The Verge's Decoder this week, and the interview is the most useful counterweight to the doom narrative I have heard this year. Partly because he is obviously biased. Mostly because he brought numbers.
Agents made customers spend more
The claim that stopped me: customers using Atlassian's MCP server and CLI grow their annual spend at twice the rate of customers who do not. They create more Jira issues and add seats faster.
And more than 98 percent of people using the MCP server still use the regular user interface. Agents did not replace the product. They increased how much the product gets used.
This matches what I see in our own operations. When we connected agents to our project management and reporting stack, usage of the underlying tools went up, not down. The agent creates tickets, updates statuses and pulls data. Humans still open the interface to decide, review and argue.
The SaaSpocalypse thesis assumed interfaces exist only to move data around. In practice, interfaces are where teams coordinate. Take that away and you do not get efficiency. You get fifteen agents producing output nobody has agreed on.
Business is not simple, and that is the moat
Cannon-Brookes makes a point that deserves more attention. If a frontier model could run your entire business, you would have a relatively simple business. Most are not simple.
They are layers of rules, compliance, local law, staff in different countries and plenty of what he calls human inconsistency. He uses a sales exception process as the example. AI can approve the customer who wants 45-day payment terms instead of 30. The customer who wants 90 days, split payments and a custom clause still goes to a human.
That 80 percent automation figure is the honest version of what AI does in operations. It covers the width of a process, the routine cases, not the full distance. The remaining 20 percent is where the judgement, the risk and usually the margin live.
For software vendors, the lesson is clear. The value was never the screen. It was the encoded rules, permissions and workflows that make a messy organisation behave consistently. Agents need those rules more than humans do, because agents do not ask a colleague when something looks wrong.
The org chart is where AI actually lands
The part of the interview I would send to every CEO is about structure. Atlassian cut about 10 percent of staff in March, and its HR leader is now titled chief people and AI enablement officer.
Cannon-Brookes put system changes and talent changes for AI in the same function on purpose. His argument is that most "is this a talent problem or a system problem" debates are a mix, 60 and 40 in one direction or the other, and you cannot fix one without the other.
Product managers, designers, marketers and finance people now write and ship code at Atlassian. That only works because the company built internal MCPs and CLIs that understand its design system, so vibe-coded features still match what customers expect. The designers did not disappear. Half their job became building the system everyone else builds with.
I wrote about the org chart becoming a loop a few weeks ago. This is a very large company running that experiment in public, with the growth rate to show it is not falling apart.
What this means for your software stack
Three implications for anyone deciding what to buy, cut or build in the next budget cycle.
First, do not cancel your system of record because an agent can read it. The agent needs the rules the system enforces. Cancel the tool nobody coordinates in, not the one everyone argues in.
Second, ask every vendor for their agent story in numbers. Do they have an MCP server? Do customers who use it grow or shrink? A vendor that cannot answer is either early or worried.
Third, expect pricing to shift. When agents create issues and read documents, seat-based pricing starts to look strange. I covered why SaaS pricing doesn't work for AI earlier this year. Atlassian's numbers suggest the transition may be more gradual than predicted, because humans are not leaving the seats.
The browser is the next battleground
One more detail deserves attention. Atlassian builds its own browser, Dia, and says 96 to 97 percent of the company uses it daily.
His reasoning is that knowledge workers spend most of the day in a browser, but no longer reading web pages. They are running applications: email, calendar, chat, project tools, video. The browser has quietly become the operating system for work.
That matters because agents need somewhere to act. Whoever owns the surface where applications, documents and agent sessions meet gets to decide how work flows between them. Expect every large software vendor to want a piece of that surface in the next two years.
Cannon-Brookes also said he tells large customers not to centralise everything on his platform. He wants Atlassian to be the connected station, not the control tower. Self-serving, maybe. Still the right architecture advice for 2026.
The SaaSpocalypse assumed software was a screen. The companies still growing are the ones that treated it as a set of rules.