a16zAtlassian CEO on the SaaS Apocalypse, AI Agents & What Comes Next
CHAPTERS
AI turns “filing cabinets” into workers: why this shift matters
The conversation opens with a historical framing: traditional software digitized records, but AI can now act on those records. The hosts set up the central tension—models feel far ahead of the real value most users currently extract.
- •Software history as converting paper-based filing cabinets into databases
- •AI’s novelty: the “filing cabinet can do work,” not just store and retrieve
- •Underutilization problem: powerful models, weak real-world value capture
- •Why a simple chat box interface often fails to unlock capability
The SaaS “apocalypse” is a valuation and disruption narrative problem
They unpack why markets are anxious: investors are trying to price software businesses during an unusually disruptive transition. Cannon-Brookes argues risk has increased, but many views assume a static world where companies and customers don’t adapt.
- •Public markets are pricing uncertainty more than fundamentals
- •AI-driven disruption raises perceived risk across SaaS categories
- •Investor behavior is reflexive: betting on what other investors will think
- •Static assumptions (no adaptation) lead to overly bleak forecasts
- •Not all SaaS will thrive, but software won’t broadly “die” either
Three buckets of SaaS companies—and why markets misprice them
Rampell proposes a simple taxonomy: (1) seat-based pricing tied directly to human labor outcomes, (2) seat-based pricing as a fairness heuristic but not tied to usage, and (3) mixed cases. Misunderstanding these buckets leads to overgeneralized sell-offs and confusion about who benefits from AI.
- •Bucket 1: seats tied to human work that AI can replace (riskier)
- •Bucket 2: seats/pricing divorced from actual usage (often safer)
- •Bucket 3: blended cases with partial exposure (e.g., creative tools)
- •Pricing is driven by perceived fairness, not marginal cost
- •AI can either erode seats or expand value via outcome/work automation
Pricing psychology: “fairness” beats rationality in SaaS monetization
Using examples from behavioral economics, Rampell explains why per-seat pricing became dominant: it feels fair even when costs are near zero. This frames how SaaS vendors may need to rethink monetization as AI changes who (or what) does the work.
- •Predictably Irrational: people reward effort/incompetence and resent efficiency
- •Seat pricing feels fair even if provisioning costs are near-zero
- •Perception of fairness shapes churn and willingness to pay
- •Outcome-based pricing must still satisfy fairness heuristics
- •AI forces repricing conversations even when product value increases
Why “vibe coding everything” is overhyped: edge cases and comparative advantage
Rampell argues that the idea everyone will replace enterprise software by “vibe coding” is unrealistic. The hidden moat is the accumulation of edge cases, compliance requirements, and learned business rules embedded over decades.
- •Comparative advantage: even if you can build it, you shouldn’t
- •Enterprise software encodes countless edge cases users don’t anticipate
- •Many rules are implicit/learned, not exposed or easily replicated
- •Simple categories may be rebuildable; complex ones resist replacement
- •Systems that operationalize messy reality gain value with AI automation
Businesses as processes (not just systems of record): input vs output constraints
Cannon-Brookes reframes the enterprise: companies are coordinated processes, shaped by internal goals and external governance. He distinguishes input-constrained work (fixed demand queues) from output-constrained work (creativity/throughput expands with efficiency), which affects how AI changes value creation.
- •“System of record” framing is too static for modern knowledge businesses
- •A company is a collection of interconnected processes
- •Input-constrained examples: support queues, legal requests—optimize efficiency
- •Output-constrained examples: engineering, marketing—efficiency enables more output
- •Compliance/law creates mandatory process constraints (e.g., state-by-state rules)
AI-driven extensibility beats replacement: custom apps on stable “rails”
Rather than rebuilding Workday or Salesforce, AI lowers the cost of tailoring tools via extensions. Vibe coding becomes powerful for niche internal workflows while the core system of record/process logic remains the trusted backbone.
- •Replacing core systems is “terrifying” due to downside and complexity
- •AI makes it cheaper to build bespoke internal tools for small teams/use cases
- •Extensions leverage existing data, governance, and business logic underneath
- •This can increase stickiness of platforms as customers customize more
- •Personal and team-specific apps emerge, while core rails stay constant
The looming pricing battle: seats, backends, and the “front end vs back end” split
They explore how AI might decouple interfaces from underlying databases and process logic, pressuring seat-based pricing in products where many licensed users never touch the UI. The key question becomes what the customer perceives as the value unit worth paying for.
- •Some orgs pay for many seats even when few people log in (Salesforce example)
- •AI could reduce “front-end” seats while backend value remains essential
- •Pricing vulnerability increases when UI usage is divorced from backend necessity
- •Fairness optics still matter: customers want simple, understandable pricing
- •Process logic (business rules) may still require identity/access even via agents
AI credits as “casino chips”: why consumption pricing scares customers
Cannon-Brookes critiques token/credit models: they’re hard to compare across vendors, hard for customers to control, and can spike unpredictably when vendors ship new features. This undermines trust and budgeting, making seats attractive despite their flaws.
- •Credits/tokens lack standardization and portability across vendors
- •Vendors can inadvertently (or deliberately) 10× usage by adding features
- •Customers struggle to control consumption driven by end-user behavior
- •Consumption pricing works best when usage is customer-controlled (logs, storage)
- •Outcome pricing is tricky because “savings” expectations ratchet each year
Why predictable pricing helps vendors too: sales planning and scaling
Rampell explains that predictable pricing isn’t only customer-friendly—it lets vendors forecast revenue per account and scale sales efficiently. Outcome/consumption models can make it harder to know how much a logo is worth, complicating go-to-market execution.
- •Workday-style pricing allows easy outside-in revenue estimation per account
- •Predictability benefits both buyers (budgeting) and sellers (sales focus)
- •Consumption models can obscure which customers are truly high-value
- •Scaling sales is harder when revenue per customer is unknowable upfront
- •Some “small” customers can outperform large logos under usage-based models
How Atlassian is adapting: platform foundations + workflow-native AI
Cannon-Brookes outlines Atlassian’s approach: build core AI platform components (governance, compliance, context layers) while shipping practical features inside existing workflows. They balance near-term usability with longer-term reimagining of products and processes.
- •AI strategy split: platform components vs app-level features
- •Workflow-native improvements (e.g., ticket summarization) deliver immediate value
- •Agent frameworks: Atlassian-provided or bring-your-own (multiple enterprise agent stacks)
- •Transition management: customers must run “today” and “tomorrow” simultaneously
- •Design investment is critical to productizing AI beyond demos
Trust, iteration, and the new UX problem: humans + agents in the loop
The closing section focuses on why customer trust is the hardest part: users need confidence, visibility, and the right level of control without constant interruptions. The challenge shifts from model capability to interaction design—how users guide, review, and iteratively refine AI outputs in real workflows.
- •Trust requires transparency and appropriate confirmation loops (not too many, not too few)
- •Iteration is hard: prompts and outputs must be refined without unwanted changes
- •Context selection is a UX burden today (which sources to use, when)
- •Agentic workflows resemble “managing interns”: productivity vs question overload
- •Example: AI-assisted document creation needs new paradigms and user re-training
