Map the vertical value chain
Use seven links rather than the single category “software”:
- acquiring trusted access to a market;
- understanding the workflow and its exceptions;
- translating that knowledge into product behavior;
- integrating with systems and institutions around the workflow;
- earning permission to hold data or perform consequential actions;
- operating, supporting, and improving the service; and
- creating measurable switching value for the customer.
Generated code directly pressures the third link. It may also accelerate experiments elsewhere. It does not automatically produce distribution, permission, institutional integrations, clean data, or accountable operation.
The 2025 DORA report frames AI as an amplifier of an organization's existing capabilities and problems. A METR study found a slowdown for a narrow population of experienced open-source developers using early-2025 tools, despite participants expecting a speedup. Neither source supplies a universal productivity multiplier. Together they warn against turning “code is cheaper” into an unqualified company valuation claim.
Run a substitution test
For each link, ask what a capable customer with an AI coding system can replace within 30, 90, and 365 days.
| Value-chain link | What would make it substitutable? | What might resist substitution? |
|---|---|---|
| Workflow UI | Clear requirements and common components | Unwritten exceptions and frequent change |
| Integrations | Stable public APIs | Certification, contracts, or brittle legacy systems |
| Data | Easy customer export | Network effects, history, normalization, or permissions |
| Operation | Standard infrastructure and low consequence | On-call knowledge, audit duties, or regulated action |
| Distribution | Direct customer reach | Trusted channel, community, or embedded partnership |
The exercise is more useful than calling every vertical product a moat or every generated dashboard a substitute.
Test both failure cases
The “AI destroys vertical SaaS” thesis fails when customers could already build the interface but chose not to own integration, support, risk, and continuous change. In that case cheaper code may expand the vendor's product surface or reduce delivery cost without changing the buying decision.
The opposite thesis—“domain knowledge always wins”—fails when the workflow is simple, requirements are legible, data is portable, switching is cheap, and the vendor mainly charges for undifferentiated CRUD screens. Domain vocabulary alone is not defensibility.
Record what would change the conclusion
A falsifiable moat memo should include:
- the value-chain link believed to be scarce;
- evidence that customers value it;
- the cheapest credible substitute;
- the time and organizational capacity required to switch;
- the portion of vendor work that code generation actually removes; and
- an event that would disconfirm the thesis.
Useful disconfirming events include customers successfully replacing the product with generated tools, falling implementation time without rising support burden, open integrations erasing proprietary access, or incumbents reducing prices without losing retention.
The key question is not whether AI can reproduce a screen. It is whether a customer can reproduce the complete value chain at lower total cost and acceptable risk.