There’s a belief that became conventional wisdom in early AI: whoever owns the proprietary data wins. Train on more data, fine-tune on better data, lock up the data. The moat is the data.

That belief is mostly wrong now. Foundation models have become capable enough that proprietary data provides a diminishing advantage. The model can already do the thing. What matters isn’t what it knows — it’s how deeply it’s embedded in the workflow where the work actually happens.

The 2026 vertical SaaS data is telling: companies that embed in the flow of money, compliance, or operations retain customers at 3x the rate of horizontal tools. The switching cost isn’t the data they’ve accumulated — it’s the workflow they’ve disrupted. It’s the muscle memory, the templates, the prompts, the process steps that have grown up around the tool.

This is why the delivery layer matters more than the feature set for new AI tools.

A standalone web app that does lease abstraction well is a useful tool. You upload the document, get the output, download it, paste it somewhere. It does one thing and does it cleanly. The workflow exists alongside your tool.

A tool embedded in the environment where the analyst already works is different. The abstraction, the rent roll, the operating statement questions, the deal memo draft — all in one thread. No switching. No uploading to an external service. No assembling outputs from three different tabs into one document. The workflow is the tool.

When analysts build their due diligence process around a tool that’s embedded in their primary working environment, the switching cost becomes the entire workflow. Not just “we’d have to migrate our data” — we’d have to rebuild how we work.

The fortress isn’t built from data. It’s built from the daily repetition of the workflow running through it.

Build for the workflow. The moat follows. +++