Every vertical AI tool is built on the same foundation models. The model is not the moat.

This is the thing that’s easy to miss when you’re building: the AI capability is table stakes. OpenAI, Anthropic, and Google will improve the underlying models continuously, and those improvements will flow to everyone building on the API. Any advantage you have from prompting the model cleverly will erode as the model gets better at the things you were compensating for. Any advantage from model selection will erode as competitors access the same models.

The moat, if there is one, is domain knowledge.

Not abstract domain knowledge — not “we understand commercial real estate.” Specific, operational domain knowledge. What does an analyst actually do with a rent roll? What’s the first thing they check, and why? What does a suspicious variance look like, and how do they flag it? What format does the output need to be in for it to go directly into their model without reformatting?

This knowledge doesn’t come from reading about the domain. It comes from working with practitioners until you know their workflow at the level they know it. The AI tool that has that knowledge embedded in its prompts, its output structure, its validation logic — that tool is doing something a general AI cannot replicate by getting smarter. The model getting better at understanding documents doesn’t help if the tool doesn’t know what to do with those documents in the context of this profession.

The founders who are winning in vertical AI right now are domain experts first and AI builders second. Or they’ve found a domain expert co-founder who makes them that combination. The AI capability they can acquire. The domain knowledge is the thing that’s hard to replicate.

Build the domain knowledge in from the start. It’s not a feature you add later. +++