Foundation Models Commoditizing Vertical AI
Raycaster
The first wave of vertical AI was easy for model labs to catch because its main job was reading the same public documents everyone else could access. Once OpenAI and Anthropic shipped research agents that browse the web, read PDFs, and synthesize sources, products built around earnings calls, patents, and web filings lost much of their standalone edge. The durable layer moved from public data retrieval to internal workflows, permissions, review rules, and audit trails.
-
Raycaster’s own product direction reflects this shift. It moved away from public document research toward tech transfer packs, specs, methods, batch records, change control, and Module 3 regulatory work, where the system has to know a company’s templates, approvers, and source of truth systems, not just answer questions from the open web.
-
Harvey and Hebbia point to the same pattern in adjacent fields. The winning products are not generic chat boxes. They package repeat workflows, control who can review outputs, and fit into how lawyers and finance teams already work. That is a much harder surface for horizontal model companies to replace directly.
-
The labs are still moving into this territory fast. OpenAI says deep research can find, analyze, and synthesize hundreds of online sources, and Anthropic offers both web search and Research for agentic, multi step investigation. That makes open internet research a feature of the model layer, not a moat for an application company.
Going forward, vertical AI companies will win by owning the messy middle inside enterprises. That means ingesting private documents, tracking edits and approvals, encoding compliance rules, and collecting feedback traces that improve future work. The market is heading toward workflow software with AI inside it, not AI wrappers around public information.