Why Vertical AI Agents Win in Regulated Enterprise B2B

Abridge listens to a patient visit, drafts the clinical note inside the physician's actual workflow, and pushes it into the hospital's electronic health record. What makes it hard to compete with was never the quality of its underlying language model. It is the exam room itself. The product does not exist as a product until it is positioned inside a clinical encounter, integrated into infrastructure a hospital already depends on, and trained on the accumulated pattern of real doctor patient interactions a new entrant has no way to access on day one.

That distinction, position in the workflow rather than model quality, is quietly becoming the dividing line in enterprise AI, and it explains why vertical agents in regulated categories are pulling away from the general purpose tools that dominated the conversation two years ago.

The generic wrapper economics have already broken

Companies built as a thin interface layered on top of a foundation model API are now seeing 40 to 70 percent of revenue consumed by API costs alone, with essentially no switching costs protecting them once a platform provider ships a native feature that replicates the wrapper for free. That is not a temporary pricing problem. It is a structural one, because the wrapper never controlled anything a competitor, including the model provider itself, could not immediately copy.

What the winners actually have that a wrapper does not

The companies pulling ahead in regulated verticals share a specific pattern, not just impressive fundraising numbers. Harvey trains on the actual contract archives of the law firms it serves, data no generic model can access, while Abridge's product cannot be sold as a better version of itself because what differentiates it is its position in the clinical workflow and the accumulated doctor patient interaction data behind it, not the language model underneath. Each new customer interaction becomes a training signal a competitor starting from zero cannot replicate, which is a very different kind of advantage than a well designed chat interface.

The scale numbers reflect how real this advantage has become. Harvey reached an 11 billion dollar valuation on roughly 190 million dollars in annual recurring revenue in March 2026, with the majority of the AmLaw 100 as customers, a scale of enterprise trust a horizontal legal assistant has never approached in a category where every output gets reviewed for liability.

Regulation is not the obstacle here. It is the moat.

The categories where this pattern is sharpest, healthcare, legal, and increasingly industrial B2B, are exactly the categories a generalist investor treats as harder to underwrite because of regulatory complexity. That complexity is precisely why the winners are so hard to dislodge once they are positioned correctly. HIPAA requirements, FDA clearance pathways, and Epic's dominance of hospital IT infrastructure are not friction a horizontal tool can shortcut around. They are exactly what keeps a well funded generic competitor from simply copying the product once it works.

This is the pattern our applied deep tech thesis already looks for

MedTech, IoT, and enterprise B2B were never chosen as focus areas because they were easy. They were chosen because the same regulatory and workflow depth that scares off generalist capital is what makes a properly positioned company difficult to replace. Working Excellence's digital engineering strategy is what actually earns a portfolio company that position inside a customer's workflow rather than adjacent to it, and BetterWorld Technology's cybersecurity services build the governance layer a regulated buyer requires before any of that integration work can go into production.

Ask where the product sits, not how good the model sounds

A demo tells you what a model can do in isolation. It tells you nothing about whether the product occupies a position a competitor cannot simply plug into. Our investment thesis starts from that distinction, because in regulated enterprise B2B, the moat has stopped being the model and started being everywhere the model is not.