One of the laziest assumptions in AI right now is that the foundation model labs will capture most of the value.
It is easy to see why people believe it. The labs move the capability frontier. They attract the attention, the partnerships, and the market narrative. If the models keep getting stronger, it is tempting to assume everything above them collapses into a disposable wrapper.
That conclusion is too blunt.
The better frame came from Bret Taylor's three-layer map at HumanX: foundation models, agent harnesses, and applied AI providers. The map matters because it names something the market still likes to blur. The value stack is not flat. Capability, orchestration, and workflow ownership are different layers, and they do not necessarily collapse into one winner.
This is where the ANSO concept matters.
ANSO is fundamentally an application layer concept. It describes a company where agents are the primary execution layer and where the system produces the primary deliverable rather than merely assisting a human operator. That kind of company does not live at the model layer. It lives close to the work.
That is why the application layer is likely to be larger and more durable than many people expect.
The closer a company gets to the real workflow, the more it inherits things the base model alone does not provide: customer-specific context, domain-specific trust requirements, review loops, regulatory constraints, system integrations, operating data, and the messy details of how a job actually gets done.
Those are not side details.
They are what turn general intelligence into economic value.
Sierra is useful here. Its pitch is not that it owns the frontier model layer. Its pitch is that it can sit on top of enterprise systems, absorb company-specific context, and deploy agents into high-stakes workflows with compliance and operational rigor. That is not a thin wrapper story. It is an applied execution story.
Harvey points to the same pattern from a different direction. Legal is not merely a text market. It is a trust market, a workflow market, and a domain-judgment market. A system that produces usable legal work product under professional review is operating inside a very different business logic than a generic assistant. The more the software produces the actual deliverable, the more it starts to look like an ANSO instead of an AI-augmented tool.
This is why wrapper discourse has become so unhelpful.
A thin UI layer that adds no workflow depth, no trust infrastructure, no domain knowledge, and no operating logic probably is fragile. But an applied AI company that owns meaningful parts of the workflow, absorbs proprietary context, and becomes responsible for real outcomes is not fragile in the same way. It is closer to the customer, closer to the work, and often closer to the risk.
That proximity matters.
It changes the trust requirements. It changes the switching costs. It changes how review happens. It changes how the product improves over time. In many cases it also changes the economics because the company is selling more than access to software. It is selling the output the software produces.
This is why the applied layer should not be treated as a temporary stop on the way to full lab dominance.
Labs can move up the stack. Some will. But not every vertical collapses into a general-purpose product surface. The moment a workflow becomes regulated, high-stakes, operationally embedded, or organization-specific, the value of applied execution rises. The more software has to fit the details of the work, the less plausible it is that a general model interface becomes the whole market.
The application layer also benefits from a dynamic many people still underrate: model progress can strengthen it rather than erase it.
If the company sits on the right side of the ANSO distinction, better models make the execution layer more capable. More of the workflow can be handled. More of the deliverable can be produced. More of the business can move from assistance toward execution. That is not the pattern of a commodity wrapper. It is the pattern of an applied system compounding as the capability frontier moves.
That is what the market is still struggling to name clearly.
The point is not that the labs do not matter.
They matter enormously.
The point is that capability alone is not the whole value chain. The software that turns capability into trusted, domain-specific, reviewable work product has its own category logic.
That is the category logic ANSO is trying to name.
The companies worth watching may not be the ones closest to the model.
They may be the ones closest to the work.
-Larry J. Erwin
Disclosure: I work at OpenAI. Some companies mentioned in this essay use OpenAI's models. The views here are my own.