Vertical AI's real moat isn't tech—it's domain experts you hire

Building applied vertical AI is surprisingly similar across hedge funds and pharma: the engineering stack is largely commoditized. The real challenge isn't coding agents—it's knowing whether they actually work in a specialized domain.

You cannot judge a pharma or finance AI output without a domain expert. Engineers lack the mental model to evaluate if a trade thesis or drug candidate is valuable. LLM-as-judge fails because it patterns matches jargon without understanding causality.

Proprietary data is the only real moat. Public data is available to everyone, but failed experiments and trade theses are hidden behind NDAs. Neither OpenAI nor Anthropic has this data. The only path is hiring the user—embedding domain experts directly into the engineering loop.

The iterative loop never ends: domain experts refine prompts, curate data, and judge outputs. Ship when the tool delivers alpha over general models—but only if it justifies ROI immediately. Finance and pharma won't wait years.