Generative AI Fundamentals

Ask Microsoft 365 Copilot to "summarize the Contoso Q3 launch decisions" and it does not answer from the language model's memory. It first runs a step Microsoft calls grounding: it gathers the relevant context it is allowed to reach (the file you have open, items in Microsoft Graph, or web results), adds that to your prompt, and only then sends the enriched prompt to the large language model (LLM) to write the answer. That single idea organizes everything on this domain. Because the answer is built from whatever context Copilot could ground against, the same prompt gives different results in different apps, the answer never escapes your existing permissions, the way you scope Copilot (open Chat, a purpose-built agent, or the Copilot inside Word or Teams) just changes what it grounds on, and an answer with no grounding behind it is the fabrication you have to catch. The classic exam trap is treating a Copilot answer as a finished fact: it is a grounded draft, and its quality only ever tracks the quality of the context it could reach.

Read this page as a map, then follow the four subtopics in order. How Microsoft 365 Copilot Works explains the grounding mechanic itself and the split between the free web-grounded Copilot Chat and the licensed work-grounded Microsoft 365 Copilot. Data Privacy and Protection in Copilot covers the rules on that grounding: your data stays inside the Microsoft 365 service boundary and is not used to train the models, Copilot only surfaces what your account could already open, and sensitivity labels carry through. Chat, Agents, and App Experiences is the where: open-ended Chat for one-off questions, a scoped agent when a team repeats a prompt, and the in-app Copilot whose capabilities differ by what each app is for. Responsible AI: Risks and Verification closes the loop: name the three risks (fabrication, prompt injection, over-reliance) and verify the output, with the rigor scaled to the stakes. Each subtopic carries its own mechanisms, tables, and traps; this page only shows how they fit together.

Across this domain the instinct the exam rewards is the same: let Copilot do the grounded drafting, but keep the person responsible for access, judgment, and the final call. Reach for the licensed work-grounded Copilot when the task depends on your own files, mail, or meetings, and the free web-grounded Chat when it does not. Trust permissions and sensitivity labels to bound what Copilot surfaces rather than expecting the model to police itself. Pick the scoped tool (an agent, or a specific app's Copilot) over a generic blank prompt when the work repeats or belongs to one app. And treat every output as something you still own: check the citations and apply human oversight before you rely on it, more so the more a wrong answer would cost. The wrong answer on AB-730 is almost always the one that hands the decision to Copilot instead of keeping the person in the loop.