Free Microsoft AB-730 Practice Exam with Questions & Answers | Set: 2

Microsoft 365 Copilot operates within enterprise governance and compliance controls. For conversation history, tenant administrators can configure retention and deletion policies. Therefore, users can delete conversation history only if allowed by administrative policy, making Statement 1 correct.

Deleting a Copilot conversation removes the chat interaction but does not delete underlying source files such as documents stored in SharePoint or OneDrive. These files are governed independently by Microsoft 365 storage and retention policies. Hence, Statement 2 is incorrect.

When a Copilot agent is uninstalled, it removes access to the agent interface but does not delete previously generated conversations. Conversations are stored as part of user activity history and remain available unless explicitly deleted (subject to policy). Therefore, Statement 3 is correct.

Thus, the correct responses are Yes, No, Yes .

a specific date range of activity

The Microsoft 365 My Account portal provides users with control over their Copilot activity history in alignment with enterprise privacy and compliance standards. When selecting Delete history , users can remove Copilot activity based on a defined time range rather than deleting only a single conversation or all activity universally.

This functionality reflects Microsoft’s commitment to transparency, user control, and responsible AI governance. Allowing deletion by date range enables organizations and individuals to manage data retention policies efficiently while maintaining compliance with regulatory frameworks such as GDPR and internal data governance policies.

The other options are incorrect because deleting a specific conversation or all conversations with a specific agent is not the primary method offered in the My Account activity deletion setting. Instead, deletion is structured around activity time periods.

This capability reinforces generative AI best practices: secure data management, lifecycle control of AI interactions, and user-directed privacy management within enterprise environments.

The instructions will apply to new chat conversations.

Answer: Yes

The instructions will apply to new conversations with the Researcher agent and the Analyst agent.

Answer: Yes

The instructions will apply to previous responses in existing conversation.

Answer: No

Custom instructions in Microsoft 365 Copilot are a persistent preference mechanism that influences how Copilot responds going forward. Once configured, they are applied when you start new interactions, helping Copilot maintain consistent tone, formatting, and behavioral preferences across future chats. Therefore, they apply to new chat conversations (Statement 1 = Yes).

Custom instructions also carry into new conversations you initiate with Copilot experiences and agents (such as Researcher and Analyst), because the instructions act as a user-level guidance layer that shapes responses regardless of which Copilot “mode” you start—provided those agents are being used in the same Copilot environment and are subject to your settings (Statement 2 = Yes).

However, custom instructions do not retroactively change content that was already generated. Existing conversation history and any prior responses remain unchanged; the instructions affect subsequent turns or new threads, not previously produced outputs (Statement 3 = No).

Microsoft 365 Copilot notebooks use the version of a file that was uploaded or attached at the time it was added to the notebook. When a document such as Process.docx is added from a local folder, Copilot references that uploaded snapshot of the file. If the file is later modified locally, the notebook does not automatically sync or refresh with the updated local version unless the updated file is re-uploaded.

Microsoft guidance on grounding and file references explains that Copilot works with the specific content stored in Microsoft 365 or the attached artifact within the notebook session. Since the updated version remains in the local folder and was not reattached, Copilot continues to use the originally added version.

Therefore, during subsequent chats in the notebook, Copilot references only the original uploaded version of Process.docx.

==================================================

Agents created in the Microsoft 365 Copilot app are primarily designed for use within the Copilot Chat experience and related Copilot surfaces, not directly embedded inside individual applications like Word. Therefore, such agents are not automatically available in Microsoft Word, making Statement 1 incorrect.

However, these agents are designed for collaboration and reuse within an organization. Microsoft guidance confirms that agents can be shared with other users in the same tenant , allowing teams to leverage common tools, prompts, and knowledge sources. Hence, Statement 2 is correct.

For advanced automation scenarios involving complex workflows and third-party connectors, Microsoft provides Copilot Studio rather than the standard Copilot app. The Copilot app focuses on lightweight agent creation using natural language, files, and simple configurations. It does not support complex orchestration with external connectors. Therefore, Statement 3 is incorrect.

Thus, the correct answers are No, Yes, No .

Microsoft guidance for effective prompting emphasizes clarity and context as the two most important factors. Providing context (Option A) ensures that Copilot understands the background, data sources, and constraints relevant to the task. This improves grounding and reduces ambiguity, leading to more accurate and useful responses.

Providing a clear goal (Option D) is equally critical. A well-defined objective—such as summarizing, analyzing, or drafting—helps Copilot determine the correct approach, structure, and level of detail for the response. Without a clear goal, outputs may be generic or misaligned with user expectations.

Using a broad range of instructions can introduce confusion rather than clarity. Acronyms may reduce clarity unless explicitly defined. Keeping prompts generic is discouraged because specificity improves quality and relevance of responses.

Therefore, the correct answers are A (adding context) and D (providing a clear goal) .

This scenario is an example of fabrication, which is commonly referred to in generative AI contexts as a hallucination. Fabrication occurs when an AI system generates information that appears credible but is factually incorrect, invented, or unsupported by verifiable sources.

According to Microsoft AI Business Professional guidance, large language models predict text based on patterns learned during training. They do not “know” facts in the human sense. As a result, when asked to generate reports using web-based information, the model may produce plausible-sounding but fictional details if sufficient grounding or reliable sources are not provided.

Deepfake refers specifically to synthetic media such as manipulated images, audio, or video. Overreliance describes a human behavior risk where users trust AI outputs without verification. Prompt injection is a malicious technique designed to manipulate model behavior. Bias refers to systematic unfairness in outputs.

In this case, the presence of fictional information in the generated report directly aligns with fabrication, making option B the correct answer.

This scenario is an example of overreliance . Microsoft’s AI guidance highlights that Copilot is an assistive tool that can accelerate analysis and drafting, but users remain responsible for validating outputs—especially for high-impact business decisions. Overreliance occurs when someone treats AI output as authoritative and acts on it without applying judgment, cross-checking sources, or validating against current conditions.

Here, the merchandiser uses Copilot to generate stocking recommendations from historical sales data, then places a large order without reviewing the suggestions or incorporating current market trends. Even if the historical data is accurate, demand can shift due to seasonality changes, competitor actions, pricing, supply constraints, macroeconomic factors, or new product launches. Copilot’s recommendations should be treated as a starting point for decision-making, not the final decision.

Option A (verification) is the opposite behavior—checking accuracy before acting. Option C (fabrication) would involve Copilot inventing facts; the prompt doesn’t indicate fictional data, only unvalidated reliance. Option D (prompt injection) involves malicious instructions embedded in content to manipulate the model, which is not described here.

Therefore, the best answer is overreliance .

The correct answer is a page because Microsoft 365 Copilot supports Copilot Pages as a place to capture AI output and then refine it collaboratively. When Researcher generates a report, saving it to a page turns the response into an editable artifact where you can reorganize sections, polish wording, add supporting details, and continue iterating with Copilot while keeping the content structured.

A “conversation” is the chat thread itself and isn’t the primary editable format for producing a finalized deliverable. An “agent” is the experience that generates the content, not the storage/edit surface. A “notebook” is used for broader organization of work, but the direct save-and-edit unit for a generated response is the page , which is designed specifically for drafting, editing, and preparing content for business use (for example, exporting or reusing in Word/Teams/Outlook workflows).

Prompt injection is a generative AI security risk where an attacker inserts instructions (often hidden in text, documents, webpages, or user inputs) to override or manipulate the assistant’s intended behavior. This can lead to unintended actions such as ignoring policy controls, producing unsafe outputs, or attempting to reveal sensitive information. Because generative AI systems follow natural-language instructions, they can be socially engineered to prioritize malicious content unless safeguards are in place. This is why prompt injection can cause data exposure (for example, attempting to extract confidential content from grounded sources) and can also embed harmful instructions that redirect the model’s behavior. In enterprise settings like Microsoft 365 Copilot, mitigations include grounding boundaries, permission trimming, content filtering, and instruction hierarchy (system policies over user instructions). From a business governance perspective, users should treat untrusted inputs (emails, documents, web text) as potentially hostile and apply least-privilege access and validation when using AI outputs in decision-making.