AI-native Software Business Models

Gokul Rajaram's AI-software thesis is that thin AI features are fragile, but AI-native companies can be durable when they own scarce assets, deep workflows, systems of record, data loops, regulated control points, money movement, hardware, networks, or full-stack vertical platforms.

From utility software to work software

Seat pricing still works for access products, such as enterprise AI tools sold as predictable per-user tiers. It breaks when the product is doing work on a user's behalf. Work products should move toward outcome pricing: per contract processed, ticket resolved, workflow completed, transaction, or business result.

Systems of action are not enough

Rajaram argues that many AI companies initially tried to sit on top of legacy systems of record as systems of action. Incumbents can block APIs, charge for data access, or bundle free agents. Durable AI entrants may need to build or replace the system of record, including the migration tools required to move data from Salesforce, Jira, Zendesk, or similar incumbents.

Vertical AI has to own the stack

A single AI function inside a vertical can be a viable business, but Rajaram doubts it becomes a very large company unless it expands to own the full stack. ServiceTitan is his canonical vertical example: many products, deep workflow, and control over a large operational category.

BPO first, hiring replacement later

Rajaram expects AI labour replacement to hit outsourced BPO spend first, because that budget is already explicit and easier to cut. Next comes not replacing departing employees. Direct layoffs are the later and more sensitive step.

What to underwrite

The two questions for pure software are: does the company create a proprietary data asset that improves with every interaction, and is it deeply embedded in a workflow that matters? Without that, model improvements and horizontal agent builders can erode the product quickly.