Where Copilot's Answers Come From
Copilot answers in two steps: it searches your organization's content, then writes an answer from what it found. The writing is the same for everyone. The content is the variable, and it explains most of the gap between a successful rollout and a disappointing one.
Imagine two companies that buy identical Copilot licenses. Months later one calls it transformative and the other is about to cancel. Same product, same price, same Microsoft. The difference is rarely the AI, the training, or the enthusiasm of the team.
How the answer gets made
When you ask the work-grounded Copilot a question, a search runs before any AI is involved. Copilot looks through the emails, files, chats, and meetings you have permission to see and retrieves the pieces that seem most relevant. Then the AI reads those pieces and composes an answer.
So Copilot does not know your organization. It looks things up quickly and writes about what it found. Put as a formula: answer quality is a constant writer multiplied by your variable content.
Copilot is a brilliant writer working from whatever your organization left lying around.
If what is lying around is current, well organized, and correctly permissioned, the writer looks like a genius. If it is three conflicting versions of the pricing sheet, a policy from 2019 nobody deleted, and a folder named "FINAL_v7_use_this_one," the writer will confidently summarize the mess. A demo runs on a curated environment. Your Tuesday runs on the real one.
Four ways grounding goes wrong
Stale content wins searches. Retrieval finds relevant content, not correct content. The 2022 travel policy is highly relevant to a travel question. It is also wrong, and Copilot has no way to know your organization stopped following it. Old content does not sit harmlessly in a library anymore. It competes with the truth and often wins.
Duplicates split the signal. Five copies of a proposal in five places, each slightly different, means retrieval may grab any of them. People experience this as "Copilot is inconsistent." The content is what changed between answers.
Structure is invisible ink. A wall of text titled "Notes" retrieves worse than a clearly titled document with headings that say what each section covers. Titles, headings, and clear names help people find things, and they help retrieval too. If a colleague would struggle to skim it, Copilot will struggle to use it.
Permissions decide the universe. Copilot searches only what the asking person can access. That is the security model working as designed, and it cuts both ways. Content locked down when it should be shared produces answers with holes. "Copilot doesn't know about our project" often means the project lives in a private site. Content shared more widely than anyone intended gets found, which is a bigger problem we cover in the next post in this series.
Why this is good news
If the divergence came from the AI, you could not fix it. You can fix your content, and doing so upgrades every answer for every user at once. Training helps one person at a time. Prompting patterns help one request at a time. Content cleanup helps every request from now on.
There is a bonus that predates AI. Current content, fewer duplicates, honest permissions, and findable structure were worth having in 2015. Organizations just could not feel the pain concretely enough to act. Copilot turns a vague "our SharePoint is a mess" into specific, visible wrong answers, and specific pain tends to get budget. We made the same argument in SharePoint is already your knowledge base, and it applies to every AI tool that reads your files, as writing so an AI can quote you shows from the outside.
A five-minute self-test
Ask Copilot five questions your team genuinely asks each other:
- What is our current pricing for a product or service?
- What is the process for onboarding a client?
- What did we decide about a specific open question?
- Where is the latest version of a particular document?
- What is our policy on a common situation?
Score each answer as right, wrong but plausible, or "it couldn't find it." Then look at why. Mostly stale answers point to an archiving problem. Wobbly answers point to duplicates. Holes point to permission walls. Wrong but plausible answers everywhere point to all three.
That scorecard is your cleanup priority list. The goal is not a perfect tenant. It is a tenant where the content Copilot retrieves most often deserves to be retrieved.
PROCESS NOTE
This essay represents our original ideas and structural direction, with prose development and editing assisted by AI.
This post is adapted from Chapter 6 of Copilot at Work, our free guide for teams adopting Microsoft 365 Copilot. Download the full guide.