Information architecture is the work that decides whether Copilot succeeds

Most disappointing Copilot pilots are not caused by the model. They are caused by the content estate underneath it. When an assistant is grounded on a document set containing four versions of the same policy, three of them obsolete, it will answer confidently from whichever version it retrieves. The user sees an AI problem. The organisation has a content problem that predates the AI by several years.

What the assistant actually sees

Microsoft 365 Copilot answers from content the requesting user already has permission to open. That single design decision has two consequences worth understanding before deployment. First, Copilot will not leak content across a correctly configured permission boundary. Second, it will faithfully surface content that a user technically can open but was never meant to browse: the folder shared with everyone in 2019, the site whose membership was never reviewed, the export of an HR spreadsheet saved to a team channel.

Permission hygiene is therefore not a compliance side-task. It is a prerequisite. The practical test is simple: pick five sensitive documents and check who can currently open them. If the answer surprises anyone in the room, the content estate is not ready.

Four preparation activities that matter most

1. Permission review on high-risk locations: broad sharing links, orphaned sites, over-permissive groups and legacy migrated libraries.
2. Duplicate and obsolete content retirement, prioritised by the areas people ask about most, not by storage volume.
3. Ownership assignment: every significant content area has a named owner and a review date. Content without an owner will decay again.
4. Metadata and labelling sufficient to distinguish current from superseded, and internal from restricted.

Grounding is a design decision, not a switch

For custom agents built in Copilot Studio, the grounding set should be explicit and small enough to be reviewed. A curated set of 200 well-maintained documents produces better answers than 20,000 unreviewed ones, and it can be evaluated: you can write test questions with known correct answers and measure whether the agent finds them.

That evaluation set is worth building early. Without it, quality assessment reduces to anecdote, and the first vivid failure in a leadership demonstration becomes the reputation of the programme.

What good preparation looks like at the end

• A named owner for each grounding source, with a review cadence.
• A permission model tested against a representative sample of sensitive content.
• An evaluation set of realistic questions with agreed correct answers.
• Retirement of superseded content in the scenarios in scope.
• A logged route for questions the assistant cannot answer, feeding content improvement.

None of this is glamorous, and it rarely features in the business case. It is, however, the difference between an assistant people trust and one they quietly stop opening.