a16zAtlassian CEO on the SaaS Apocalypse, AI Agents & What Comes Next
CHAPTERS
- 0:00 – 0:10
AI is ahead of its real-world value: the “chatbox with unlimited power” problem
Mike and Alex open with the gap between what modern AI models can theoretically do and what users actually ask them to do. They frame the core challenge as turning raw capability into practical, workflow-embedded value that normal users can trust and adopt.
- •Models’ capabilities outpace the value most users extract from them
- •Users default to trivial queries when faced with an open-ended interface
- •The main bottleneck is productization (workflow + UX), not model intelligence
- 0:10 – 0:32
From filing cabinets to databases to software that does work
Alex traces software history as digitizing “filing cabinets” into databases—useful, but often adding new overhead (security, IT, provisioning). AI changes the game by making the filing cabinet active: systems can now execute tasks, not just store and retrieve data.
- •1960–2022: software largely digitized records and retrieval
- •Efficiency gains were real but came with new operational burdens (CISO, IT, SSO)
- •AI enables systems (e.g., accounting) to perform actions, not just hold data
- 0:32 – 3:16
Making sense of the “SaaS apocalypse” narrative
Erik tees up market fears about SaaS, and Mike argues the confusion is largely about valuing businesses during disruption. He emphasizes markets are pricing uncertainty and static assumptions (that companies won’t adapt), even as many software businesses continue performing well.
- •Public markets are struggling to value software amid rapid AI disruption
- •Perceived risk has increased even when fundamentals remain strong
- •The dominant fear: “What if AI does this in 2–3 years?”
- •Static thinking ignores real-world adaptation by companies and customers
- 3:16 – 6:39
Risk is up, but software won’t uniformly die—some will adapt and thrive
Mike acknowledges not every SaaS company will make it through the next decade, paralleling past platform shifts. He argues the opportunity is enormous for knowledge-work software—if vendors execute and continuously prove durability to impatient markets.
- •Not all SaaS companies will survive the next platform transition
- •Some software becomes a steady cash-flow business rather than a growth story
- •AI is structurally positive for knowledge-work tooling—execution is the differentiator
- •Markets demand proof over time; patience is scarce
- 6:39 – 10:42
Three types of SaaS businesses—and why markets lump them together
Alex proposes a three-bucket framework for SaaS: (1) seats tightly tied to human work output, (2) seat-based pricing that’s “fairness optics” but not tied to doing work, and (3) middle cases. He argues investors are not distinguishing which revenue streams are most exposed to AI-driven labor substitution.
- •Bucket 1: seat revenue tied to human labor doing the work (more AI risk)
- •Bucket 2: pricing feels fair but seats aren’t tied to outcomes (more resilient)
- •Bucket 3: mixed exposure (e.g., creative tools)
- •Revenue can go to zero—or expand—depending on pricing/model shifts
- 10:42 – 13:40
Why “vibe coding everything” fails: comparative advantage and hidden edge cases
Alex calls the idea that everyone will build their own Workday/Zendesk replacements “preposterous,” citing comparative advantage and the reality of buried edge cases. Software encodes years of learned exceptions (often not documented), making naive replication far harder than it seems.
- •Comparative advantage: even if you can build it, it’s often irrational to do so
- •Real software value often lives in unspoken, learned edge cases
- •Deterministic rule sets emerge from decades of real-world exceptions
- •DIY replacement works only for simple domains with minimal edge-case load
- 13:40 – 20:42
Businesses aren’t “systems of record”—they’re coordinated processes
Mike reframes the “system of record” concept as too static for modern enterprises. He argues businesses are collections of processes constrained by internal goals and external rules (laws, compliance), and AI’s impact depends on which processes are input- vs output-constrained.
- •“System of record” framing can understate process and coordination complexity
- •Businesses run on process coordination more than static data storage
- •Input-constrained work (e.g., support, legal) optimizes for efficiency
- •Output-constrained work (e.g., creative, engineering) reinvests gains into more output
- 20:42 – 22:55
Build vs buy in the AI era: the Goldilocks zone and what’s truly replaceable
Alex discusses when it makes sense to build custom tools versus using existing platforms, noting cost, criticality, and switching friction. They explore “atomic” systems of record and highlight that even infrequently accessed records (like cap tables) can be too important to risk recreating.
- •DIY makes sense only when pain/cost is high enough to justify ownership risk
- •Some systems are low-stakes to swap; others are rare-use but mission-critical
- •Systems of record vary widely in replaceability and business impact
- •Trust and correctness requirements often outweigh frequency of use
- 22:55 – 25:29
Vibe coding’s real leverage: extensibility on top of trusted platforms
Mike distinguishes between replacing enterprise systems and extending them. AI-assisted coding can dramatically lower the cost of building highly specific internal apps that sit on top of stable platforms (e.g., Workday/Gmail), making underlying systems stickier and more valuable.
- •Replacing core systems is terrifying/risky; extending them is powerful
- •AI lowers the cost of building niche internal tools for small teams/use cases
- •Extensions leverage existing data, governance, and business rules
- •Personal apps may remain personal; a few may evolve into companies
- 25:29 – 35:24
Pricing after AI: fairness optics, front-end/back-end decoupling, and “AI credits” confusion
They debate how software pricing evolves when AI changes who “uses” a product and how value is delivered. Mike argues customers dislike consumption/outcome pricing when it’s hard to control, and he critiques AI credit models as opaque “casino chips” that vendors can inflate via new features.
- •Seat pricing persists because it’s predictable and feels fair to buyers
- •Front-end/back-end decoupling raises pressure to reduce paid “seats”
- •Consumption/outcome pricing works best when customers can control usage
- •AI credits are hard to compare across vendors and hard for customers to budget
- •Outcome-based pricing can erode over time as “savings” become the new baseline
- 35:24 – 43:19
How Atlassian is adapting: platform foundations + workflow upgrades + agent insertion
Mike explains Atlassian’s approach as a three-layer problem: understand AI, build durable platform components (AI gateway, graph, compliance), and ship features that improve today’s workflows while enabling new ones. He describes a progression from simple helpful features (like ticket summaries) to agentic steps to reimagined workflows.
- •AI features must work in current workflows while paving paths to future workflows
- •Platform investments: AI gateway, teamwork graph, enterprise controls/compliance
- •Immediate value: workflow accelerators (e.g., service ticket summarization)
- •Agent strategy: insert agents into high-friction steps; support third-party agent platforms
- •Long-term: redesign workflows so legacy artifacts (like tickets) may disappear
- 43:19 – 54:18
Why customer trust is the hardest problem: human-agent loops and new UX patterns
Mike argues trust, iteration, and UX are the limiting factors—not raw model quality. He describes the tension between oversight and annoyance, the difficulty of choosing the right context, and the need for collaborative “human-agent” workflows; he closes with a concrete example of AI-assisted document creation requiring new user habits.
- •Users fear silent autonomous actions; trust requires the right level of confirmation
- •Context selection is confusing when users must choose data sources and scopes
- •Too many agent questions feels like managing interns; too few reduces trust
- •Design patterns for editing/iterating on AI outputs are still immature
- •Example: AI-assisted document creation shifts users from “blank page” to prompt+coauthoring paradigms
