Korva is a scroll of the internet’s best ideas.
Model labs are moving into apps
Labs take customers
The pattern is playing out in real businesses. A company running a multi-billion dollar operation can lose its biggest customer overnight, because the customer wants to work with OpenAI, which looks more AI forward than a traditional SaaS business. Harvey's CEO ranks the model labs as his biggest competitive worry.
Who is exposed
How exposed a company is depends on how it reaches customers. Designers can pick up a design tool and use it right away. A legal AI company has to build relationships with law firm partners, then deploy to junior lawyers who don't want to use it because they think it will take their jobs. That deployment is the heavy lifting.
Who adopts instead
Companies built on network effects or operations are likelier to adopt the big labs than to compete with them. A systems integrator like Infosys is a long way from a lab's core business, while a scalable software product like a legal chatbot or insurance claims automation is something a lab could build. Legal is probably priority number 12 for Anthropic, and the only thing a legal AI company lives and breathes.
Why do labs move up into applications?
Clayton Christensen described the law of conservation of attractive profits: when one layer of a product becomes interchangeable, profits move to an adjacent layer. If several models perform about the same, customers can swap one for another, and the model alone earns less. The layer that owns the customer relationship is harder to swap out. Christensen laid out the idea in his 2003 book The Innovator's Solution.
Why does legal AI depend on relationships?
Lawyers owe clients a duty of confidentiality, set out in American Bar Association Model Rule 1.6, which requires reasonable efforts to prevent unauthorized disclosure of client information. Software that touches client material therefore goes through partner-level review before it reaches junior staff. That review is a sales process, and it runs on trust built over months. Firms also run security questionnaires on any vendor that handles privileged documents.
How do platforms end up competing with their customers?
Developers have a name for it, Sherlocked, taken from Apple's 2002 Sherlock 3 search tool, which added features that a third-party app called Watson had been selling. The pattern repeats across the industry: a platform owner watches which products succeed on it, then builds a version into the platform. Amazon's private-label products and Microsoft's bundling of its browser are other well-known cases. Each time, the outside vendor faces a rival that controls distribution.
More from 20VC with Harry Stebbings
