An employee asks Microsoft 365 Copilot a straightforward question about a client relationship and gets a generic, slightly wrong answer. The instinct is to blame the model. In most of these cases, the model is not the bottleneck. The actual answer exists somewhere inside the organization — in a Slack thread from eighteen months ago, a slide deck on someone's local drive, a SharePoint folder with the wrong permissions set — and Copilot was never given access to any of it.
This is a much less glamorous problem than model capability, which is probably why it gets less attention despite being the more common cause of failure. A frontier model reasoning brilliantly over an incomplete context window will still produce a wrong or hollow answer, because the limiting factor was never its reasoning; it was what it could see. This is precisely the problem enterprise search tools like Glean were built to solve, by indexing across Slack, Google Drive, and SharePoint with permissions respected — and even Glean's own customers still run into gaps where a document exists but was never connected.
Organizational knowledge is scattered by design, not by accident. Permissions exist to protect sensitive information, and different departments use Slack, SharePoint, and Notion with different retention and search behavior for defensible operational reasons. That fragmentation was manageable when humans did the searching, because a colleague could remember an exception to the rule that a search index cannot represent. Microsoft 365 Copilot inherits none of that tacit knowledge unless someone has deliberately engineered it in.
The harder version of the problem is staleness. A SharePoint document that was accurate two years ago and has not been updated does not announce its own obsolescence to a retrieval system. A copilot that surfaces it with full confidence is not hallucinating in the sense usually meant by that word — it is faithfully reporting something a human organization forgot to correct or delete.
Solving this requires investment that looks nothing like buying a better model: permission-aware retrieval systems like Glean, systematic content freshness tracking, and often an uncomfortable audit of how much internal documentation is simply wrong. Most enterprises have not done this work, because it is unglamorous, cross-departmental, and does not show up as a line item anyone gets credit for funding.
This creates a mismatch in expectations. Buyers evaluate Microsoft 365 Copilot or Glean in demos using clean, well-organized sample data, then deploy against the actual mess of a real organization's document sprawl and are surprised when performance drops sharply. The gap is not the vendor overselling the model. It is the buyer underestimating how much value depends on organizational context no vendor can fix from outside.
The companies getting the most value from enterprise AI right now are not necessarily the ones with the best model access. They are the ones that had already done the unglamorous work of organizing and permissioning their internal knowledge before Copilot or Glean ever arrived to make the absence of that work visible.