Most durable value from applied AI accrues to companies with proprietary data and repeatable workflows, not to those buying capability off the shelf.
Capability that can be purchased by anyone is priced as a commodity within a short period. That is now the position of general model access, and it will be the position of most tooling built directly on top of it.
The durable position belongs to companies that hold something the model does not: operating history, structured records, regulated workflows, or a distribution relationship that determines where output is consumed.
When we underwrite an AI initiative inside a mature business, the questions are ordinary. Which process is being automated, what did it cost before, what does it cost now including inference, and is the improvement visible in the accounts rather than in a demonstration.
Initiatives that survive those questions are usually narrow, unremarkable to describe, and materially accretive. Initiatives that do not are usually the reverse.