What is a moat when the model isn't yours?
The labs keep absorbing what used to be startup features. So what is left that a startup can actually defend?
sherlocked, at the speed of a release note
trying to answer →What is a moat in an AI world?Which parts of an AI startup are actually defensible?
Mac developers have a verb for this: to get Sherlocked. In 2002 Apple shipped Sherlock 3, which did most of what a popular third-party app called Watson did, and Watson’s business went away. The platform owner looked at what was selling on top of it and built it in.
Every AI startup built on someone else’s model lives with that risk at a much faster tempo. Document chat, coding help, voice, search, agents: each was a company’s differentiator before it became a line in a lab’s release notes. If the model isn’t yours, the model can’t be your moat, and neither can anything the lab might reasonably ship next.The moats chart scores ten sources of advantage before and after. Features are the one that drops to nearly nothing.
What seems to survive is the stuff a lab doesn’t want to do: the workflow inside a specific customer, the permissions and integrations, the relationships, the data that only arrives because you’re already there, and the grind of execution in one ugly vertical. The moat moves from the technology to everything around it.
Open question I keep circling: is that a moat, or just a head start with better manners? Distribution you rent gets repriced. Workflow you own might not.