🟦 Microsoft 365 Copilot
• AI grounded in your work data (Microsoft Graph), across Word, Excel, PowerPoint, Outlook & Teams
• Prompt with Goal + Context + Source + Expectations; name sources with /
• Honors your permissions — but fix oversharing before you roll it out
Where Copilot lives
Two surfaces: Copilot Chat (the cross-app assistant grounded in your work data and the web, in Teams/Outlook/the web) and in-app Copilot embedded in Word, Excel, PowerPoint, Outlook, and Teams. Chat reasons across everything you can access; in-app helps with the document or task in front of you. There is a free Copilot Chat tier and the paid Microsoft 365 Copilot add-on that unlocks deep work-data grounding and the in-app experiences.
Prompting: Goal, Context, Source, Expectations
Microsoft's framework is Goal + Context + Source + Expectations (add persona/tone if useful). The biggest quality move is naming the Source explicitly with the / picker so the answer is grounded and citable. Specify an output contract (audience + format + length). Iterate with follow-ups rather than expecting a perfect one-shot, and use Copilot Lab for vetted starter prompts.
In the apps
Word — draft from a source, rewrite/adjust tone, summarize, and ask the document questions. Excel — insights, formula columns, conditional formatting, charts; requires structured tables. PowerPoint — build decks from a Word doc, redesign, summarize. Outlook — thread summaries, drafted replies, coaching. Teams — intelligent recap, action items with owners, real-time catch-up. The recurring pattern: ground in a real source, give a clear audience/format, review the draft.
Agents, security & rollout
Agents are scoped Copilots grounded on specific knowledge and instructions, built as declarative agents or in Copilot Studio (low-code) and shared via the agent store. Governance: permissions are honored but oversharing is inherited, so remediate loose SharePoint/OneDrive access first; sensitivity labels are respected and admins have a management dashboard. Rollout is the real bottleneck — train on prompting, seed a shared prompt library, name champions, measure outcomes.