Summary

The “SaaS apocalypse” is a repricing of uncertainty before it is proof of universal impairment. Mike Cannon-Brookes concedes that software has become riskier and “not every SaaS company is going to thrive through the next decade,” but argues markets are extrapolating two- or three-year AI scenarios while assuming incumbents remain static. Atlassian has delivered three great quarters, and for his knowledge-work business, “this is the best thing that’s happened to our business” — subject to execution through the transition.
Alex Rampell’s three-bucket test separates impaired seat models from systems whose AI upside is being ignored. Zendesk-like seats directly fund work that agents may eliminate, so without repricing “that revenue stream is 100% going to zero”; with outcome pricing, revenue might instead triple or quadruple. Workday’s employee-based seats are not tied to outcomes, while Adobe sits between those poles — distinctions Rampell says public investors are failing to price.
The durable moat is accumulated process knowledge, including edge cases that cannot be recovered from a prompt. Rampell invokes David Ricardo’s comparative advantage and the Indiana employee-on-maternity-leave problem: companies could theoretically vibe-code core software, just as they could grow their own food, but recreating decades of hidden rules while engineers have other work is terrifying and economically unattractive. Cannon-Brookes’s sharper framing is that businesses are collections of processes, not databases.
AI will affect input-constrained and output-constrained work differently. Customer support and legal teams face fixed incoming queues, so faster processing can improve efficiency and reduce cost; marketing, creative work and software development can absorb efficiency gains into more output. That split matters because the same AI capability can compress seats in one workflow while expanding activity and software value in another.
Vibe coding is more credible as an extensibility engine than as a replacement for core systems. Cannon-Brookes calls the idea of running a self-built Workday “terrifying,” but sees enormous value in cheaply generating a 20-person Miami application on top of Workday’s data and rules. Erik Torenberg characterizes that as making the underlying platform “stickier in the enterprise and more valuable,” even as bespoke interfaces proliferate.
Software pricing remains governed by perceived fairness, control and predictability — not technical purity. Customers tolerate consumption pricing when they choose the unit, as with Splunk logs or S3 storage, but AI credits feel like opaque “casino chips” that vendors can consume by adding features. Outcome pricing has another flaw: after software cuts support spending from $20 to $10, the customer resets $10 as the baseline and asks to reach $5.
The near-term AI bottleneck is product design and trust, not model capability. “Give people a chat box that can do unlimited power and they’re like, ‘Tell me a dad joke,’” Cannon-Brookes says; the models are far ahead of realized value because users need contextual workflows, understandable agent behavior and well-timed human checkpoints. Atlassian is addressing that through its AI gateway, Teamwork Graph, workflow summaries, agent integrations and Rovo’s hybrid document-and-chat interface.

Deep dive

1. AI turns the filing cabinet into a worker

Cannon-Brookes describes software’s history from 1960 through 2022 as taking a filing cabinet and turning it into a database. Generative AI changes the pattern because “the filing cabinet can do work.”
Rampell later applies that idea to systems such as Intuit: QuickBooks need not merely retrieve accounting information for a human; software could collect outstanding receivables or perform another task itself, turning a passive record into a potential labor layer.

2. The market is pricing one apocalypse across three SaaS buckets

Cannon-Brookes sees a rational rise in perceived risk becoming a broad valuation response that may fail to distinguish among software businesses. Investors are not only estimating discounted cash flows; they are betting on “what other people think that other people think” investors will do amid competing visions of AI’s future.
His objection is that many bearish scenarios are static: AI changes radically over two or three years while incumbents, customers and workflows supposedly do not. Atlassian’s three great quarters sit uneasily with that response; he says such results can seem disconnected from the reality on the ground, though the company must keep proving its adaptation.
Rampell’s most exposed bucket contains seats tied directly to work. If agents answer the support requests that once required Zendesk operators, required seats could fall to zero; Zendesk must change its product and pricing, although successful outcome pricing could also make the opportunity “triple or quadruple.”
At the opposite pole, Workday prices by total employees even though those employees are not each producing an outcome inside Workday. Rampell says “nobody’s going to get rid of QuickBooks,” mentions a 45% decline around February 26 or 27 while discussing Workday and Intuit, without clearly identifying the company in the excerpt, and places Adobe in the less-clear middle.

3. Process knowledge, not stored data, is the durable moat

Cannon-Brookes dislikes “system of record” because it makes a business sound like a static database — the conceptual equivalent of a floppy-disk save icon. Atlassian’s more than 10,000 people arrive carrying their brains and leave with them each evening; the business exists by coordinating their processes.
He divides those processes into input-constrained and output-constrained work. A support team cannot generate ten times more customer questions merely by working ten times faster, while marketing or software teams can use the same productivity gain to create more output and customer value.
Processes also encode laws, governance and compliance: the Indiana employment edge case is not optional just because a generated application overlooks it. Rampell’s point is that much software consists of deterministic rules accumulated “in the wild,” with the valuable rules embedded rather than publicly exposed.
Intuit illustrates why even published rules do not eliminate process value. The tax codes may be downloadable, but Intuit’s special ability is knowing how to ask users the right questions about messy life data; AI makes companies consider which of their presumed 50 “secret sauce” processes are genuinely unique.

4. Vibe coding extends core systems more readily than it replaces them

Rampell grounds his skepticism in Ricardo’s 1817 theory of comparative advantage: being able to grow food or weld aluminum does not make self-production economical. “I could theoretically vibe-code myself some Workday,” but recreating its edge cases while engineers have other work is a poor trade.
Replacement economics also have a Goldilocks zone. A rarely used Miami conference-room system may be acceptable to change, while a Carta cap table is accessed infrequently yet cannot be wrong; its high consequence and relatively low cost favor buying the proven system.
Cannon-Brookes nevertheless sees a major extensibility gain. A company may now afford a bespoke Miami front-desk app for 20 people, incorporating local HR policy while running on Workday’s data and global rules — customization that previously could not justify an internal technology team.

5. Pricing survives when customers can see and control the value

Rampell uses Dan Ariely’s locksmith story to explain why pricing follows fairness rather than marginal cost: customers resent $500 for a 90-second rescue yet tip the incompetent locksmith who struggles for nine hours. Humans are “capable and willing to pay for incompetence,” and per-seat SaaS feels fair despite near-zero provisioning cost.
Salesforce exposes the tension between front end and back end. Rampell says his firm may have roughly 600 Salesforce licenses although he never logs in; he sometimes consumes the relationship-data output, illustrating why the underlying record can remain essential even if direct interfaces and seats shrink.
Cannon-Brookes rejects consumption pricing as the universal answer. Splunk logs and S3 gigabytes are understandable and controllable units; AI credits are opaque “casino chips” whose consumption can jump because a vendor adds automatic summaries, leaving customers unable to know whether providers’ units are comparable or control their bill.
Outcome pricing can destroy its own reference point: cutting support cost from $20 to $10 works in year one, then the buyer demands $5 because $10 is the new reality. Rampell praises Workday’s two-way predictability; at his payments companies, Walmart underperformed expectations while Casper unexpectedly made the economics work.

6. Atlassian is straddling today’s workflows and agent-native ones

Atlassian begins with a favorable substrate: collaboration problems across service, HR, finance, business and software teams involve enormous amounts of text. But customers must move large organizations gradually, so Cannon-Brookes says Atlassian must take them “one year, two years, and five years into the future simultaneously.”
The company separates platform foundations from individual features: its AI gateway, Teamwork Graph, enterprise compliance and controls should support whatever future models and interfaces emerge. Customer-facing capabilities can then evolve without rebuilding the organizational context and governance beneath each application.
A deliberately unmagical Jira example is ticket summarization. When four, five or six people have touched a complex enterprise issue, the next expert might spend 30 minutes absorbing conversations and attachments; a context-aware summary removes that boot-up time without changing the workflow “one iota.”
Atlassian then inserts agents into expensive workflow steps while also exploring workflows where the service ticket disappears entirely. Cannon-Brookes expects most businesses to operate three to five large agent platforms, so Jira must accommodate Atlassian’s framework alongside systems such as Agentforce or Gemini.

7. Trust and context are now the binding constraints

Cannon-Brookes calls the gap between model capability and delivered value almost trite, yet enormous. An unconstrained chat box produces paralysis — “unlimited power” becomes a request for a dad joke — because ordinary users want outcomes without learning models, prompts or technical plumbing.
Trust depends on revealing enough agency without creating approval fatigue. An inbox agent that silently sends 15 emails is frightening, but repeatedly asking “Are you sure?” is maddening; the interface must earn autonomy through a history of correct work and proportionate checkpoints.
More context is not automatically better. The Teamwork Graph may remember that Cannon-Brookes wrote code in 2002, but that fact should influence an explanation only when relevant; forcing users to toggle web search and organizational search pushes poorly understood context decisions onto them.
His management analogy captures the coordination tax: 50 interns produce plenty of work but may ask “50 questions a minute.” Jira agents therefore let users ask what they are doing mid-task, yet Cannon-Brookes says users reject both extremes — an opaque box and “a thousand steps” of telemetry.

8. The winning AI interface still has to be taught

Rampell’s Nano Banana 2 example isolates the unresolved editing problem: it can one-shot an impressive Japan etiquette infographic, but should a user revise its text, manipulate graphics, prompt a new version or combine all three? Torenberg’s conclusion is that users need to “get your head into the model” for trust and iteration.
Atlassian’s Create with Rovo offers one answer: 75% of the screen is an editable document and 25% is chat. Users can research a section, ask how a board member might interpret the draft or command “make every heading blue,” while retaining direct control of the prose.
Power users rapidly alternate between the two panes and call the system amazing; ordinary business users ask whether they should simply type on the left. Cannon-Brookes expects the paradigm to become learned infrastructure, like Excel or mobile gestures, but the speakers agree that neither the industry nor Atlassian has fully solved it yet.

Why Leonardo was a saboteur, Gutenberg went broke, and Florence was weird – Ada Palmer OpenAI's Fog of War + Betting on Iran + Hard Fork Review of Slop