AB-730 sample exam questions

Sharing a prompt by link only works inside the same organization. A colleague at another company gives you a link outside your tenant, so Copilot reports the prompt as not found. Outdated, scheduled, or labeled-file prompts do not produce this message.

Prompts, responses and Graph data are not used to train foundation models. Copilot never grants or changes permissions; it only respects the access a user already has, using the same controls as other Microsoft 365 services, so oversharing must be fixed separately.

Researcher responses include citations that link to the sources used, so you can open them and verify validity. Instructions, memory, and capabilities describe how an agent is configured or personalized, not the origin of facts in a report.

Information that sounds plausible but is invented is fabrication, also called a hallucination. A deepfake is synthetic media imitating a person, overreliance is trusting output without checking, prompt injection is malicious input, and bias is unfair skew in output.

A notebook gathers pages, files, and other references in one place and lets you hold many different Copilot conversations grounded on that shared material. Adding pages to a single conversation or downloading them does not give that reusable multi-chat workspace, and stories are not a Copilot feature here.

Grounding is improved by pointing Copilot to specific files or sources, using the / file reference, so answers draw on real content. Shorter prompts, avoiding jargon, and specifying output format affect clarity or style but do not tie the answer to your data.

Researcher is a prebuilt, Microsoft-managed agent, so users cannot edit its instructions, suggested prompts, or other settings. To get customized behavior, you would build your own agent in Copilot Studio or the agent builder instead.

Instructions tell an agent how to behave, including its tone, persona, and rules, so a warm and friendly style belongs there. The description only summarizes the agent's purpose, capabilities add features, and suggested prompts offer users starting questions.

This canned refusal appears when Copilot's responsible AI safety filters flag the prompt as violating content guidelines. A vague prompt gets a clarification or generic answer, and there is no five-request limit. Exceeding the context window gives a different length-related message.

Custom instructions let you set standing preferences, such as tone, style, and format, that Copilot applies to every conversation. Actions on your behalf come from agents and plugins, security rules come from admin policies, and personal details like name and role are only part of the optional about-you section.

A page is collaborative, so you can mention the colleague in it with an at-mention or share the page link. Opening in Word does not itself involve a colleague, and adding the page to a notebook organizes reference material rather than sharing it.

Copilot can still fabricate or misattribute, so the dependable way to ensure accuracy is to manually open and check the cited sources against the original document. Copilot, agents, or notebooks verifying their own output cannot replace human review.

Copilot respects your permissions through Microsoft Graph, so it only reaches meetings you can access and cannot answer about meetings you were not invited to. Within that access, it can summarize chats, channel posts, meetings, and upcoming calendar events.

Agent knowledge can include Teams chats alongside sources such as SharePoint, email, and documents. Copilot conversations are not a selectable source, and knowledge files are text-based formats such as Word, PowerPoint, Excel, and PDF. Supported sources change, so check current documentation.

Copilot in Teams can summarize the meeting so far while it is in progress, which is the fastest way to catch up. The transcript is slow to read, the Microsoft 365 Copilot app lacks live meeting context, and the intelligent recap is available only after the meeting ends.