Copilot comes with Microsoft 365. Why doesn't it know your facts?

Here is a scene I have watched in more than one form. A Monday meeting at a consultancy of fourteen people. The founder asked Copilot a simple question in Teams: what do we charge for the quarterly retainer once the volume discount kicks in? Copilot answered in seconds, fluently, with a citation. The citation pointed to a pricing deck last edited in October 2023. The retainer had been repriced twice since. Nobody in the room caught it until the number was already halfway into a customer email.

That scene is why Microsoft Copilot gives wrong answers about your company even though it can read everything you have. Copilot grounds its answers in whatever documents rank highest in SharePoint and OneDrive, so an outdated file produces a confident, cited, wrong answer. Company Brain fixes the source. It holds a small set of company facts your team has reviewed and approved, and delivers them to Copilot through MCP, so answers start from what is actually current.

The rest of this post explains where Copilot's answers come from, why the failure keeps happening to teams with perfectly good document hygiene, and when the built-in grounding is honestly all you need.

Where Copilot's answers actually come from

Microsoft 365 Copilot is grounded in your Microsoft Graph content. When you ask a question, it retrieves the files, mails and messages that rank highest for your query across SharePoint, OneDrive, Outlook and Teams, respects your permissions while doing it, and writes an answer from what it found.

That is real power, and it deserves to be called that. Ten years of accumulated files become searchable in plain language. Meeting recaps, thread summaries, "find the proposal we sent that logistics client", all of it works, and for questions about what a document says, retrieval is exactly the right tool. Company Brain does not replace any of that.

The trouble starts with a different kind of question: what is true at our company right now?

Documents are where facts go to age

Retrieval ranks documents by relevance to your question. Relevance has no concept of superseded. That October 2023 pricing deck ranks beautifully: it has the right keywords, a clean table, and a history of being opened often. The decision that replaced it happened in a meeting nobody minuted, so as far as the document estate is concerned, it never happened at all.

Your files hold every version of the truth at once. The draft nobody deleted sits next to the final. The policy from the old owner sits next to the one you actually follow. The "FINAL v4" naming convention was abandoned around v7. Copilot can tell which document matches your question. It cannot tell which document your company still stands behind, because that knowledge lives in people's heads and chat threads, and often was never written down anywhere.

The deck in that Monday meeting answered the question asked, precisely and with a source. It was simply no longer true. Confidence plus a citation reads as truth, which is exactly what makes this failure expensive.

Why better search will not fix it

The instinctive response is hygiene: archive the old decks, tidy the SharePoint, agree on naming rules. Teams try this, and it helps for a quarter. Then the estate grows faster than anyone curates it, because curating documents is a full-time job that a fourteen person company has given to nobody.

Even perfect hygiene leaves the deeper gap. Many of your most important current facts never live in a document at all. The discount you never give in writing. The customer name that must always be spelled a certain way. The approval threshold that changed last month. Ranking a document pile is the wrong instrument for facts like these, no matter how good the ranking gets.

One thing this argument is careful about: Copilot is extensible. Through Copilot Studio, an admin can build agents that reach knowledge outside the Graph, including external actions. The limits described here are about what Copilot grounds on by default, out of the box, for the whole team.

What a reviewed fact layer changes

Company Brain is a reviewed set of company facts that your AI tools read through MCP, and once an admin connects it, Copilot reads them too. Instead of asking retrieval to guess which of nine pricing files is current, your team writes the fact down once: "The quarterly retainer is priced per the 2026 rate card; volume discount applies from the third seat." What your team writes passes a curator's review before it becomes shared context, so the deck that should have retired in 2023 never gets a vote.

The connection is deliberately simple. Your company gets one shared MCP address. An admin adds Company Brain to a Copilot agent in Copilot Studio and publishes it once, and from then on the whole team's Copilot answers start from the reviewed facts, with the document estate still available for everything documents are good at. The step by step setup lives on our page about how Company Brain connects to Microsoft Copilot.

The same facts serve Claude, ChatGPT and Cursor through the same connection, so "what we charge" stops depending on which app someone opened. Company Brain's free plan includes 50 company facts, one curator and the whole team connected, and it stays free.

When Copilot's built-in grounding is enough

Sometimes it genuinely is, and you should know before adding anything.

If your team is small enough that one person knows every current number and reviews anything customer-facing, human review is your fact layer, and it works. If your documents are under real document control, as in some regulated firms where superseded files are withdrawn on a schedule, the estate itself stays current and retrieval over it is trustworthy. And if company-specific questions are rare in your Copilot usage, the occasional wrong deck is an annoyance, an annoyance you catch, and that can be fine.

The gap shows when the same questions about pricing, policy, product names and terminology come back every week, from different people, in drafts that go straight to customers. At that point the cost of one uncaught October 2023 answer exceeds the cost of writing twenty facts down properly.

The lesson from that Monday meeting

What stayed with me was how good the wrong answer looked. Copilot had done its job exactly as designed: found the most relevant document and reported it faithfully. The failure was upstream, in a pile that contained three eras of pricing and no marker for which era is now.

Give Copilot its pile, it is genuinely great with the pile. Then give it a small, reviewed set of facts for the questions where "current" matters more than "relevant". If your team has been compensating by pasting context into every chat, our guide on how to stop repeating your company context in every AI chat walks the whole escalation ladder, and where it ends.