Capability-Value Gap Driven By Productization, Ux, And Trust
Sources: 1 • Confidence: Medium • Updated: 2026-03-08 21:18
Key takeaways
- AI model capabilities are currently far ahead of the real-world value users extract from them.
- Seat-based SaaS pricing persisted in part because it feels fair to buyers even when marginal provisioning cost is near zero.
- The idea that organizations will broadly replace core enterprise systems by end-user AI coding is argued to be unrealistic due to hidden edge cases and operational risk.
- A core software shift is moving from digitizing records to enabling those records to autonomously execute tasks.
- A common bearish SaaS narrative assumes a static world where software companies do not adapt, and that premise is argued to be incorrect.
Sections
Capability-Value Gap Driven By Productization, Ux, And Trust
- AI model capabilities are currently far ahead of the real-world value users extract from them.
- AI's realized value is argued to be limited more by product design and user experience than by model quality because users struggle to operationalize blank chat interfaces.
- Building trust in AI actions requires a design balance where the system previews or explains what it will do without overwhelming the user.
- Prompt input and context selection are major UX problems because users cannot reliably choose what sources and memories to include, leading to agents that ask too many clarifying questions.
- High-quality corporate agentic work often requires iterative human-agent loops, and the number and timing of loops must be tuned to avoid distrust or frustration.
- AI-assisted document creation requires significant user re-learning because it shifts writing from a blank-page paradigm to a prompt-driven doc-plus-chat workflow with commands operating across the document.
Pricing And Packaging Constraints For Ai Features
- Seat-based SaaS pricing persisted in part because it feels fair to buyers even when marginal provisioning cost is near zero.
- AI credit or token schemes are argued to fail customer control and cross-vendor comparability, and vendors can increase customer consumption by shipping new features that burn more credits.
- SaaS businesses can be grouped into three categories: seats tied to AI-automatable work, seats that act as a fairness proxy for headcount, and an intermediate partially exposed category.
- Consumption-based and outcome-based pricing are expected not to 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, making costs predictable and manageable.
- Outcome-based pricing tied to cost savings is argued to become harder to sustain because customers reset expectations to the new lower cost baseline in later years.
Incumbent Resilience Vs Replacement-By-Custom-Coding Narratives
- The idea that organizations will broadly replace core enterprise systems by end-user AI coding is argued to be unrealistic due to hidden edge cases and operational risk.
- AI-assisted coding is reported to be useful for extending and customizing existing platforms rather than replacing them.
- Switching risk for a system of record depends on how directly it touches revenue and business outcomes, making low-stakes records easier to replace than high-stakes ones.
- Building software in-house is argued to be rational mainly in a cost-and-criticality 'Goldilocks zone' where the tool is both expensive relative to costs and business-critical.
- High-consequence accuracy requirements can justify staying with a specialized vendor even when the record is accessed infrequently.
- Systems of record may benefit from AI by adding task-executing capabilities that require being the underlying record system.
Systems Shift From Records To Process Execution And Orchestration
- A core software shift is moving from digitizing records to enabling those records to autonomously execute tasks.
- Businesses are better modeled as coordinated sets of processes (including compliance-driven external rules) rather than static systems of record.
- AI-driven efficiency impacts differ between input-constrained processes with fixed demand queues and output-constrained processes where more output can be produced.
- Even when rules are deterministic and public, a key automation bottleneck can be mapping messy personal data into those rules.
- A differentiated value in tax software is argued to be eliciting the right user information and asking the right questions, not the tax code itself.
- In service management, AI ticket summarization can materially reduce onboarding time for new participants in a multi-person ticket, but it requires careful contextual handling beyond naive LLM summarization.
Market Narrative Uncertainty And Valuation Mechanisms
- A common bearish SaaS narrative assumes a static world where software companies do not adapt, and that premise is argued to be incorrect.
- During the public-market SaaS sell-off, valuations fell broadly without reliably distinguishing business models by AI substitution exposure.
- SaaS risk premia have risen because investors are pricing uncertainty about how other investors will react to AI disruption, not only underlying cash flows.
- Atlassian has had three strong quarters in a row while rapidly changing how it works to adapt to AI.
- Not all SaaS companies will thrive through the next decade; an AI-era shakeout is expected.
Unknowns
- What measurable evidence (attach rates, retention deltas, productivity metrics) shows that AI features are delivering real-world value commensurate with model capability improvements?
- In which SaaS categories are seat counts actually declining due to AI automation, and what pricing/model transitions (if any) are succeeding at offsetting the seat pressure?
- Do enterprises accept any standardized, controllable metering unit for AI usage, or do they force AI value back into seat bundles and fixed commitments?
- How frequently do enterprises attempt to replace core systems with custom-built alternatives, versus building AI-driven extension layers on top of incumbents?
- How well do workflow-embedded AI features (for example, service ticket summarization) generalize across contexts without adding new failure modes or trust breakdowns?