Every team that wants an AI assistant ends up needing the same thing underneath it: a way to retrieve the right internal documents and ground a model's answer in them. Most organizations let each team build that retrieval layer themselves, once per project, which means paying for the same mistakes — chunking, embedding choice, freshness — five separate times.
I treat retrieval as a shared platform capability the CoE owns, the same way infrastructure teams own a database or a message queue. One well-maintained index, one evaluation set for retrieval quality, one team responsible for keeping it current — consumed by every downstream team instead of rebuilt by each of them.
The payoff isn't just efficiency. A shared retrieval layer is also the single place you can enforce access control and data governance consistently, instead of auditing five bespoke implementations with five different blind spots.