Your Model Vendor Is Becoming Your Software Vendor Too

In March 2026, Harvey, the legal AI company built on top of the big foundation models, raised two hundred million dollars at an eleven billion dollar valuation. Seven weeks later, Anthropic launched Claude for Legal. It integrates Harvey, and it also competes with it. Around the same time, OpenAI hired the co-founder of Ironclad to build its own legal product. The companies that make the engine are now building the car.

For the last couple of years, the working assumption in enterprise AI was that the foundation labs would sell raw intelligence, and a layer of specialized application companies would sit on top of it, do the domain-specific work, and own the relationship with the customer. Harvey for law, others for finance, healthcare, and the rest. The labs would be the utility; the apps would be the brands. That assumption is now being tested in public, because the labs have started climbing into the application layer themselves.

Part of the why is visible in the labs' own numbers. Anthropic reports that its engineers now merge eight times as much code per day as they did in 2024, with more than 80% of its production code written by Claude itself. When software gets that much cheaper to build for the people who own the models, the application layer stops looking like someone else's territory.

Where the value is moving

This is a value-migration story, and it is one finance and legal ops teams have seen before in other categories. When the supplier of a critical component decides to sell the finished product too, everyone who built a business in between has to ask where they still add value, and every buyer has to ask who they are really betting on.

The awkward position belongs to the companies in the middle. A vertical application built on someone else's model is, to some degree, exposed to that model's owner deciding to offer the same thing. Some of those applications have real moats, in workflow, integrations, trust, and domain depth that a general model vendor cannot easily replicate. Others have less than their valuations imply. Telling the two apart is now a live question rather than an academic one.

What this means if you're buying

For a buyer, the lesson is not to pick the winner of this fight. It is to notice that the fight makes the layers of your stack behave differently, and to decide on purpose which ones you want to be able to change.

The model will keep changing hands and changing price. That is fine, and you want to keep it swappable so you benefit from the competition. The application layer is now contested, which means the vendor you standardize on today might be undercut by its own supplier tomorrow, or might be the supplier doing the undercutting. You want flexibility there too. The one layer you actually control, and the one that holds its value no matter who wins above it, is your own data and how it is organized.

That reframes build versus buy. The question is less about which product to buy and more about which layer to own. Buy the model. Buy or rent the application. But the structured record of what your own documents say, the obligations in your contracts, the figures in your filings, the terms you have agreed to across thousands of agreements, is the asset that stays valuable regardless of how the contest above it resolves.

Where we sit

That last layer is the one we build, and we build it to belong to the customer rather than to us. But the point does not depend on us. When the company that sells you the engine starts selling the car, the sensible move is to hold tightly to the one thing neither the engine maker nor the carmaker can take with them: the organized knowledge of your own business. Keep that yours, and you can let the layers above it compete for your money for years.