# Free AB-731 practice questions

> 10 free AB-731 practice questions with the correct answer and an explanation.

Page: https://devopscodelab.com/exams/microsoft-ab-731/free-practice-questions/
Practice exams: https://devopscodelab.com/exams/microsoft-ab-731/

### Question 1 of 10

A retail executive asks how generative AI differs from the predictive model the company already uses to forecast weekly store demand. Which statement BEST describes the difference?

- A. Generative AI always returns the same output for the same input, while predictive forecasting models are designed to vary their answers on each run.
- B. Generative AI only works with structured tabular data such as sales tables, while predictive models are designed to process unstructured documents, images, and audio.
- C. Generative AI requires labeled training data for every output it produces, while predictive models learn entirely without any labeled examples.
- D. Generative AI produces new content such as text, images, or code, while the forecasting model predicts a value or label from patterns in historical data.

**Answer:** D

**Explanation:** Generative AI creates new content (text, images, code, summaries) based on patterns learned during training, whereas traditional predictive models output a number or class such as next week's demand. Generative models are especially strong with unstructured data, not limited to tables. Foundation models are pretrained largely with self-supervised learning rather than labels for every output, and generative output is non-deterministic, so the other statements reverse the real characteristics.

### Question 2 of 10

A legal operations team pilots a chat assistant that drafts summaries of contract clauses. Reviewers find that some summaries cite clauses that do not exist in the source contract. Which generative AI challenge does this describe, and what is the MOST appropriate first mitigation?

- A. Fabrication; ground the assistant on the actual contract text and require clause citations that reviewers can check against the source.
- B. Latency; move the assistant to a provisioned throughput deployment so that summaries are generated faster during the peak contract review hours.
- C. Bias; retrain the underlying foundation model from scratch on a balanced set of contracts from many different jurisdictions.
- D. Data drift; schedule monthly retraining of a classification model so that contract categories stay aligned with the current portfolio.

**Answer:** A

**Explanation:** Content that sounds plausible but is not supported by the source is a fabrication (often called a hallucination). Grounding the model on the real contract and showing citations lets humans verify every claim. Retraining a foundation model to fix bias is costly and does not address invented clauses, provisioned throughput affects speed and capacity rather than accuracy, and drift retraining applies to predictive models whose input data changes over time.

### Question 3 of 10

A customer-support leader wants to reduce the cost of an AI assistant without hurting answer quality. Analysis shows that every request includes a 6,000-word policy manual pasted into the prompt. Which change would MOST reduce token consumption while keeping answers grounded?

- A. Ask agents to rephrase each customer question more politely, because courteous prompts are processed with fewer tokens.
- B. Index the manual and use retrieval so that only the few passages relevant to each question are added to the prompt.
- C. Switch to a larger model with a bigger context window so the full manual fits more comfortably in every request.
- D. Increase the maximum output length setting so the model can answer several customer questions inside one long response.

**Answer:** B

**Explanation:** Retrieval-augmented generation retrieves only the passages relevant to the question, which cuts input tokens dramatically while keeping responses grounded in the manual. A larger context window lets more text fit but does not reduce what you pay for, and larger models usually cost more per token. Politeness does not meaningfully change token counts, and raising the output limit increases rather than reduces consumption.

### Question 4 of 10

A bank wants a model that consistently classifies customer complaints into 40 internal categories with a very specific labeling style. Prompting with examples gives inconsistent labels, and the bank has 20,000 accurately labeled historical complaints. Which approach is MOST appropriate?

- A. Train a brand-new foundation model on the complaints because pretrained models cannot learn custom categories.
- B. Fine-tune a pretrained language model on the labeled historical complaints to teach it the bank's labeling style.
- C. Use retrieval-augmented generation to look up the complaint-handling policy manual for each complaint before the model answers.
- D. Increase the temperature setting so the model explores more category options before settling on a label.

**Answer:** B

**Explanation:** Fine-tuning adapts a pretrained model to a specialized task or output style when there is a substantial set of high-quality labeled examples and prompting alone is inconsistent. RAG adds knowledge but does not teach a consistent labeling behavior. Raising temperature increases randomness and makes labels less consistent, and building a foundation model from scratch is far more expensive than adapting an existing one.

### Question 5 of 10

An IT director explains that Microsoft 365 Copilot can only show users content they already have permission to access. Which component makes this possible by providing access to organizational data with existing permissions?

- A. Microsoft Graph, which connects Copilot to Microsoft 365 data while honoring each user's access rights.
- B. Copilot Studio, which grants each new agent tenant-wide read permissions to SharePoint and Exchange data.
- C. Microsoft Foundry, which hosts a separate copy of the company's mailboxes inside a dedicated model project.
- D. Azure AI Search, which copies all tenant documents into a shared index that every licensed user can query.

**Answer:** A

**Explanation:** Microsoft 365 Copilot retrieves organizational content through Microsoft Graph and the semantic index, and it only returns items the signed-in user is already permitted to see. Azure AI Search is a separate retrieval service for custom solutions and does not underpin Microsoft 365 Copilot's grounding. Foundry is a platform for building custom AI apps, and Copilot Studio agents operate under defined permissions rather than receiving tenant-wide access.

### Question 6 of 10

A company must decide whether to buy, extend, or build for three needs: drafting documents in Word, an HR agent grounded on internal policies used inside Microsoft 365 Copilot, and a customer-facing agent that uses its own fine-tuned model and custom orchestration. Which mapping is MOST appropriate?

- A. Buy Microsoft 365 Copilot for drafting, extend it with a declarative agent for HR, and build the customer agent in Microsoft Foundry.
- B. Build all three solutions in Microsoft Foundry so that every one of them uses the same custom orchestration and the same fine-tuned model.
- C. Buy Microsoft 365 Copilot for all three needs, because in-app Copilot can host fine-tuned models for external customers.
- D. Extend Microsoft 365 Copilot for drafting, buy a prebuilt HR app from a partner, and build the customer agent in Agent Builder.

**Answer:** A

**Explanation:** Out-of-the-box drafting is a buy decision through Microsoft 365 Copilot, an HR agent that reuses Copilot's orchestration and grounding is an extend scenario with a declarative agent, and a customer-facing agent with its own model and orchestration is a build scenario suited to Microsoft Foundry. Building everything ignores existing capabilities. Agent Builder cannot host a custom fine-tuned model with custom orchestration, and in-app Copilot is not a platform for external customer agents.

### Question 7 of 10

A retailer wants to extract item names, quantities, and totals from thousands of scanned supplier invoices each month and send the results into its ERP system. Which Microsoft service is MOST appropriate?

- A. Azure Document Intelligence in Foundry Tools, using its prebuilt invoice model to extract fields and line items.
- B. Azure AI Search, which indexes the scanned invoices so staff can run keyword queries over them.
- C. Microsoft 365 Copilot in Excel, which summarizes each scanned invoice that an employee opens or pastes into a workbook.
- D. Azure Translator in Foundry Tools, which converts each invoice into the default language of the ERP system.

**Answer:** A

**Explanation:** Azure Document Intelligence provides prebuilt models such as the invoice model that extract structured fields like vendor, line items, and totals from scanned documents at scale for downstream systems. Azure AI Search makes content searchable but does not output structured invoice fields. Translator changes language rather than extracting data, and Copilot in Excel is an interactive per-user experience that does not process thousands of scans automatically.

### Question 8 of 10

A company has several thousand employees with Microsoft 365 E3 licenses. Leadership wants every employee to have secure, web-grounded AI chat with enterprise data protection at no additional per-user license cost. Which option meets this requirement?

- A. Microsoft 365 Copilot Chat, which is available to Microsoft 365 users signed in with their Entra ID accounts.
- B. Microsoft 365 Copilot add-on licenses assigned to every employee so that they receive in-app Copilot features.
- C. Microsoft Copilot with personal Microsoft accounts, which employees can create at no cost for business use.
- D. Copilot Studio capacity packs assigned to each employee so that they can build their own chat agents.

**Answer:** A

**Explanation:** Microsoft 365 Copilot Chat is included for users with Microsoft 365 business or enterprise subscriptions and offers web-grounded chat with enterprise data protection when they sign in with their work account. Assigning Microsoft 365 Copilot licenses adds a per-user cost. Personal consumer accounts lack organizational protection and governance, and Copilot Studio capacity is purchased per tenant for agent usage rather than to give everyone a chat experience.

### Question 9 of 10

A company wants all employees to use AI chat with enterprise data protection while only some roles receive Microsoft 365 Copilot. Which licensing statement is accurate?

- A. Microsoft 365 Copilot is included with every Microsoft 365 subscription, while Copilot Chat requires a separate per-user license.
- B. Copilot Chat and Microsoft 365 Copilot are both free for users who sign in with personal Microsoft accounts on work devices.
- C. Both Microsoft 365 Copilot and Copilot Chat require a Copilot Studio capacity pack for each user before they can be used.
- D. Copilot Chat is included with eligible Microsoft 365 subscriptions, while Microsoft 365 Copilot is a separate per-user license.

**Answer:** D

**Explanation:** Copilot Chat is available at no extra cost to users with eligible Microsoft 365 subscriptions signed in with work accounts, and Microsoft 365 Copilot is purchased as a separate per-user license that adds work-data grounding and in-app features. The reversed statement is wrong, Copilot Studio capacity is for agent usage rather than per-user enablement, and personal accounts do not receive organizational Copilot entitlements.

### Question 10 of 10

Which set lists Microsoft's six responsible AI principles?

- A. Fairness, availability, confidentiality, integrity, transparency, and accountability.
- B. Fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability.
- C. Accuracy, reliability and safety, scalability, inclusiveness, explainability, and ownership.
- D. Fairness, profitability, privacy and security, inclusiveness, transparency, and automation.

**Answer:** B

**Explanation:** Microsoft's responsible AI standard is built on fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability. Profitability and automation are business goals rather than principles. Scalability and ownership are not on the list, and availability, confidentiality, and integrity are the classic information-security triad rather than responsible AI principles.

Ready for a timed run? Take the free 15-question AB-731 mini exam: https://devopscodelab.com/app/exams/microsoft-ab-731/mini/

DevOpsCode Lab is not affiliated with Microsoft. AB-731 and related names are trademarks of their owners. We write every question ourselves; these are not real exam questions.
