The model is probably the most replaceable part of an enterprise AI stack.

The harder asset is the system that knows what exists, who can access it, how it connects, and which source each task should query.

That is the layer @Glean has been building .

They think enterprise AI needs a unified index across applications, with permissions, metadata, authorship, activity signals, and entity relationships preserved before inference begins.

It built one cross-application layer that preserves entity relationships, permissions, authorship, activity signals, and metadata.

Memory carries learning across sessions, connectors pick the retrieval path per source, and tools let the assistant act on what it finds.