Displacement Vs Extension: When Ai Reduces Seats And When Platforms Get Stickier
Sources: 1 • Confidence: Medium • Updated: 2026-04-11 17:47
Key takeaways
- Broadly replacing core enterprise systems by 'vibe coding' equivalents is unrealistic because of hidden edge cases and operational risk.
- Seat-based SaaS pricing persists partly because it feels fair to buyers even when marginal provisioning cost is near zero.
- Businesses are better modeled as coordinated sets of processes (including compliance-driven external rules) rather than only as 'systems of record.'
- Surviving the AI transition requires redesigning human-software collaboration loops and trust boundaries rather than merely adding an AI feature.
- AI model capabilities are currently far ahead of the real-world value users extract from them.
Sections
Displacement Vs Extension: When Ai Reduces Seats And When Platforms Get Stickier
- Broadly replacing core enterprise systems by 'vibe coding' equivalents is unrealistic because of hidden edge cases and operational risk.
- SaaS businesses can be grouped into three buckets: seats tied to work that AI can perform, seats that are primarily a fairness proxy for headcount, and an in-between set with partial exposure.
- Switching risk for a system of record depends on how directly it touches revenue and business outcomes.
- Mike Cannon-Brookes reports internal gains from using 'vibe coding' to extend and customize existing platforms rather than replacing them.
- Building software in-house is only rational in a cost-and-criticality 'Goldilocks zone' where the tool is expensive relative to costs and business-critical.
- High-consequence accuracy requirements can justify sticking with a specialized vendor even if a record is accessed infrequently.
Pricing Constraints: Fairness, Predictability, And Meter Design
- Seat-based SaaS pricing persists partly because it feels fair to buyers even when marginal provisioning cost is near zero.
- AI token/credit pricing is problematic because credits are not standardized across vendors and vendors can inadvertently increase customer consumption by shipping new features that burn more credits.
- Usage-based pricing is most acceptable when customers can directly control the metered input and compare it across vendors, whereas AI credit/token schemes fail this control and comparability test.
- Consumption-based and outcome-based pricing will not become the dominant SaaS pricing model because customers dislike unpredictable bills and pricing units they cannot control.
- Customers accept usage-based pricing when the metered unit is controllable and transferable, which makes cost predictable and manageable.
- Outcome-based pricing tied to cost savings becomes harder to sustain over time because customers reset expectations to the new lower cost baseline in later years.
Business Reality Is Process- And Data-Mess-Driven, Not Rule- Or Record-Driven
- Businesses are better modeled as coordinated sets of processes (including compliance-driven external rules) rather than only as 'systems of record.'
- AI-driven efficiency impacts differ for input-constrained processes versus output-constrained processes.
- Even when rules are deterministic and public, a hard part of automation can be mapping messy personal data into those rules.
- Intuit's differentiated value may be its process for eliciting the right user information and asking the right questions to apply the tax code rather than knowing the tax code itself.
- Knowledge-economy firms can retain durable value because proprietary internal playbooks encode hard-to-replicate operational know-how.
- Many non-software products effectively sell accumulated cultural and operational knowledge built over long periods rather than just an explicit recipe.
Trust, Context, And Human-Agent Loop Design As The Binding Constraint For Agents
- Surviving the AI transition requires redesigning human-software collaboration loops and trust boundaries rather than merely adding an AI feature.
- In service management, AI ticket summarization can materially reduce onboarding time for new participants in a multi-person ticket but requires careful contextual handling beyond naive summarization.
- Building trust in AI actions requires previewing or explaining what the system will do without overwhelming the user, because too much transparency becomes noise and too little feels unsafe.
- Prompt input and context selection are major UX problems because users cannot reliably choose what sources and memories to include, leading to many agents generating too many clarifying questions.
- High-quality corporate agentic work often requires human-agent iterative loops, and the number and timing of loops must be tuned to avoid distrust or frustration.
- Atlassian has shipped Jira agents that can be assigned work, including the ability to chat with an agent while it is performing a task.
Capability-Value Gap And Near-Term Adoption Path
- AI model capabilities are currently far ahead of the real-world value users extract from them.
- Design and experience, more than model quality, are limiting AI's realized value because users struggle to operationalize blank chat-box interfaces.
- AI-assisted document creation requires significant user relearning because it shifts writing from a blank-page paradigm to a prompt-driven, doc-plus-chat workflow where commands can operate across the entire document.
- Near-term AI adoption is driven by improving existing workflows with practical in-flow features rather than fully agentic redesigns.
Unknowns
- What measurable productivity lifts (cycle time, throughput, quality) are enterprises actually realizing from in-flow AI features versus agentic redesigns?
- What are attach rates, retention impacts, and willingness-to-pay for vendor AI add-ons in major SaaS categories?
- In labor-tied seat categories, are seat counts declining due to AI substitution, and are vendors successfully migrating to new pricing or offerings?
- How frequently do enterprises attempt to replace core systems via custom builds, and what are success/failure rates compared to extending existing platforms?
- Do AI customization layers measurably increase platform stickiness (net retention, switching costs, expansion) or do they introduce maintenance/security overhead that offsets benefits?