The default Copilot experience inside Word, Excel, and PowerPoint became Agent Mode in early 2026. Most teams still treat Copilot as a chat panel that lives to the side of the document. The actual shift is in the document itself — Copilot now edits directly, asks clarifying questions, and works alongside the author rather than at arm's length.
This is the walkthrough of using Agent Mode to rewrite an annual review document against a brand voice guide. Eight minutes of video showing the full flow including the clarifying question moment and the two-pass review pattern.
The setup
The user has two documents open in the tenant.
The first is an annual review prep brief that needs to go out to a client. It is roughly half written but does not yet pass the team's brand voice rules.
The second is a brand voice and guidelines document. It defines the team's house style: punctuation conventions, banned words, tone register for internal versus external communications, sentence-length guidance, structure rules.
The annual review brief is open in Word. The Copilot pane is open on the right, defaulted to editing mode. The brand voice guide is referenced as input to the prompt.
The first prompt: rewrite by rule
The user's prompt establishes the job:
"Can you please check this annual review against our brand voice and guidelines and identify anywhere where it's breaking the rules? Where you find that it is breaking the rules, I want you to rewrite that section for me one by one so that I can see it on the screen and then also present what you have changed, why you changed it, and what rule that it does break."
Two things are being asked. First, flag every rule violation. Second, rewrite each section with a justification.
Copilot reads the brand voice guide, scans the annual review document, identifies the violations, and starts rewriting. Each rewrite appears live in the document (highlighted in blue, as tracked changes). For each rewrite, the chat pane explains what changed and which rule was broken.
This is the first pattern that breaks the chat-only mental model. The output is not a single reply in the chat pane. It is a sequence of in-document edits with paired explanations.
The clarifying question moment
Mid-rewrite, Copilot stops. A question appears in the chat pane:
"Do you want these rewrites to stay as an internal tone or shift to external?"
The annual review is going to a client, so the user replies "external". Copilot continues the rewrite with the external tone applied to every subsequent section.
This is the moment that makes Agent Mode genuinely different from a generative one-shot. Copilot does not guess. When the disambiguation matters, it pauses and asks. The answer shapes the rest of the work, not just the next edit.
For workflows where tone is contextual (the same source content could go to different audiences), this is the single feature that prevents Copilot from getting halfway through a document in the wrong register and forcing the author to redo the work.
The two-pass review pattern
After the first rewrite, the user accepts all the edits and runs a second pass:
"Can you please review that document one more time against the brand voice document and let's see if we are all good now."
Copilot re-reads the now-rewritten document, compares against the guide, and reports back differently this time. The report is structured into two categories:
Hard rules: pass. No em-dashes, title reframed, formatting adjusted. The unambiguous rule violations from the first pass are now resolved.
Recommended fixes. Long sentences with multiple clauses in sections one and four. Some sentences still dense. These are softer style suggestions that need author judgement, not absolute rule breaks.
The user picks one ("let's go ahead with number four") and Copilot rewrites that specific section. The other softer suggestions are left for the author to decide on.
This two-pass pattern — first pass for hard rules, second pass for softer suggestions you can accept or reject — is the workflow that most matches how a human editor would work over a document.
Why this changes how teams use Word with Copilot
Three patterns I see in client tenants once teams adopt Agent Mode beyond the chat panel.
Style consistency at scale. Teams with a brand voice guide finally get to enforce it without manual reading of every document. The cost of rewriting against a style guide drops to single-digit minutes per document.
New starters land faster. A new team member writes a draft, runs it through Agent Mode against the team's style guide, and gets a rewritten version that matches the team's voice on day one. The institutional style knowledge transfers through the rewrite, not through years of feedback.
External communications stop sounding internal. Documents that were drafted as internal notes but need to go to clients can be tone-shifted in a single pass. The clarifying-question pattern catches the ambiguity before Copilot guesses wrong.
Where this fits with Copilot in SharePoint
Agent Mode is a Microsoft 365 Copilot capability, not a Copilot in SharePoint capability specifically. But the two complement each other.
When Copilot in SharePoint generates a Word document (via the What to Produce capability), the output lands in your SharePoint library. Opening that document in Word, you can then use Agent Mode to refine, rewrite, or polish against your team's style guide.
Same when a Skill produces a Word document as its output. The Skill creates the structure; Agent Mode refines the voice.
The combination is more useful than either alone. SharePoint Skills handle the structured workflow; Agent Mode handles the per-document polish.
What to do this week
If your team writes documents in Word that need to match a consistent voice, three actions.
-
Open one document that should have followed your brand voice but does not. Reference your voice guide. Ask Copilot in Agent Mode to flag every rule violation with a rewrite. See how close it gets on the first pass.
-
Run the second pass after accepting the obvious fixes. Compare which softer suggestions you agree with versus skip. This calibrates your expectation for how aggressive the rewrites should be.
-
Document any rule misses. If Copilot misses a rule that is in your voice guide, the guide probably needs to be more explicit. The clearer the guide is written, the better Agent Mode performs against it.
For sites where the voice guide is sitting unused, a SharePoint.md context file makes the rules available to Copilot in SharePoint automatically, not just to Word Agent Mode on demand. The combination — the rules in SHAREPOINT.md for site-wide application, plus Agent Mode in Word for per-document polish — is the production pattern most useful teams settle on.
If your organisation wants to roll the Agent Mode + SharePoint.md + Skills combination out across multiple teams, the Copilot in SharePoint Immersive covers the full pattern in one day.
