Microsoft exam

AB-731 Sample questions

Microsoft Certified: AI Transformation Leader

Independent: not affiliated with Microsoft. Free to try; the study bundle adds a study guide and two practice exams.

5 original questions, none of them in the practice exams.

Pick an answer. Show answer explains every option and links its Microsoft source.

5 questions / Explanations link Microsoft sources / Last updated Oct 26, 2026

Question 1

Multiple choice

Domain 1 / Task 1.1 Foundational concepts of generative AI

A car-rental company's customer-care team uses a large model in Microsoft Foundry to draft replies to booking-change requests. Reviewers rate the drafts highly, but at 60,000 requests a day the cost per reply and the time to produce each draft are above target. The team wants similar quality at lower cost and latency. What should the AI lead propose?

A Add more example replies to every prompt sent to the large model, so that each draft needs less work from the model. Your answer B Fine-tune the large model on the highest-rated replies; fine-tuning always improves results, so the current setup can be retired without a comparison. Your answer C Fine-tune a smaller model and switch off the assistant's content filters, because a fine-tuned model applies its own safety controls. Your answer D Fine-tune a smaller model on replies generated by the large model, then compare it with the current setup on held-out requests before switching. Correct

Show answerHide answer

Answer: D

Why

Microsoft's fine-tuning overview lists 'Specialize a smaller model' among the reasons to fine-tune: 'Teacher-generated examples can help a smaller model meet task requirements at lower cost and latency.' It also says 'Benefits aren't guaranteed; held-out evaluation helps you compare the fine-tuned model with your baseline.' Training a smaller model on the large model's replies and testing it against the current setup follows both statements.

Why not the others

AExamples in the prompt are sent with every request, so adding more raises token cost and latency. Microsoft describes fine-tuning as adjusting the model's weights 'rather than adding examples to each prompt' and lists 'Reduce prompt overhead' among its uses: 'Fewer prompt examples can lower token costs and latency.'
BMicrosoft says 'Benefits aren't guaranteed; held-out evaluation helps you compare the fine-tuned model with your baseline.' Retiring the current setup without that comparison skips the check, and fine-tuning 'adds training and hosting costs' of its own.
CMicrosoft says fine-tuning 'doesn't replace retrieval for current information or application-level safety controls', so the content filters stay in place.

Source: Fine-tuning in Microsoft Foundry (Microsoft Learn).

Question 2

Multiple choice

Domain 1 / Task 1.2 Benefits and capabilities of generative AI solutions

A bakery chain wants a model that flags which of its 300 industrial ovens are likely to fail in the next 30 days, so that engineers can service those ovens first. It has two years of sensor readings for every oven (temperature, hours run, door openings), but breakdowns were only ever recorded as a yearly repair cost per region, never against the oven that failed. What should the AI lead recommend first?

A Start recording each breakdown against the oven that failed, so that readings can be paired with known outcomes for a supervised model to learn from. Correct B Treat each region's yearly repair cost as the outcome for its ovens and train a regression model on their readings to rank the ovens by predicted cost. Your answer C Run clustering on the sensor readings of all 300 ovens, which reveals the ovens that are likely to fail without needing any record of past failures. Your answer D Keep collecting readings from every oven for another year, so that a classification model has enough sensor data to learn which ovens fail. Your answer

Show answerHide answer

Answer: A

Why

Microsoft defines supervised machine learning as algorithms 'in which the training data includes both feature values and known label values', trained 'by determining a relationship between the features and labels in past observations, so that unknown labels can be predicted for features in future cases'. The sensor readings are the features, but without a record of which oven failed and when, there is no label to learn from; a yearly repair cost per region cannot be tied to individual ovens. Recording each breakdown against the oven that failed gives the chain labels to pair with its readings.

Why not the others

BRegression predicts a numeric label for each observation from that observation's features, as in the page's examples (ice creams sold on a given day, the selling price of a property). A cost recorded for a whole region over a year is not an outcome recorded for any one oven, so each oven's readings still have no known outcome, and a cost is not the same as failing within 30 days.
CClustering is unsupervised: it trains on data with 'feature values without any known labels' and 'groups the data observations based purely on similarity of features'. Ovens with similar readings land in one group, but nothing in that grouping records which ovens failed, so it does not reveal which ones will.
DSupervised learning needs training data that 'includes both feature values and known label values'. More readings add more feature values but no record of which oven failed, so a classification model would still have no known outcomes to learn from.

Source: Types of machine learning model (Microsoft Learn).

Question 3

Multiple choice

Domain 2 / Task 2.1 Microsoft 365 Copilot and Microsoft Copilot

A city council has eligible Microsoft 365 licenses, and its 2,500 staff use Copilot Chat (Basic). No one has the Microsoft 365 Copilot add-on license, and no pay-as-you-go billing has been set up. The digital team proposes four agents for all staff. Which one can staff use in Copilot Chat at no additional cost?

A An agent that answers questions from the HR policy library stored on the council's own SharePoint site. Your answer B An agent built from instructions and the national tax agency's public website, for tax questions. Correct C An agent that drafts replies to residents from each staff member's own mailbox and meeting notes. Your answer D An autonomous agent built in Copilot Studio that logs every planning-permit email when it arrives. Your answer

Show answerHide answer

Answer: B

Why

Microsoft's Copilot overview says: 'Declarative agents that are grounded in instructions and public websites are included with Copilot Chat. Access to custom or other agents is pay-as-you-go only.' Its license comparison lists 'Access to agents that use web data' and 'Pay-as-you-go work data agents' for Copilot Chat (Basic). An agent built from instructions and a public website uses web data only, so staff can use it without the add-on license or a billing policy.

Why not the others

AThe HR library is organizational content on SharePoint. For Copilot Chat (Basic), agents that use work data are pay-as-you-go, so this one needs a billing policy or Microsoft 365 Copilot licenses.
CMailboxes are organizational data that Copilot Chat (Basic) can't use through Microsoft Graph, and Microsoft lists email as agent knowledge only for users with Microsoft 365 Copilot licenses, not for Copilot Chat users even with pay-as-you-go billing.
DCopilot Studio is licensed independently of Copilot Chat and its agents' usage is billed in Copilot Credits; only declarative agents grounded in instructions and public websites are included with Copilot Chat.

Source: What is Microsoft Copilot? (Microsoft Learn).

Two practice exams with 100 more like these

Question 4

Multiple choice

Domain 2 / Task 2.2 Foundry Tools

A national court service wants to use key phrase extraction from the Language service in Foundry Tools to pull the main points out of incoming case filings, so that clerks can route each filing faster. Its data rules say filings may not be sent to the cloud for processing, and the program director concludes that Foundry Tools are ruled out. What should the AI lead tell the director?

A Every Foundry Tool can run on premises, so the court can use any feature without checking whether it is offered as a container. Your answer B Containers on premises are capped at a low number of transactions per second, so the court should expect slower processing than in the cloud. Your answer C The Language key phrase extraction container runs the same API on the court's own servers, so the filings can be processed where they are stored. Correct D Foundry Tools are offered only as cloud APIs, so the court must either change its data rules or build its own model. Your answer

Show answerHide answer

Answer: C

Why

Microsoft says Foundry Tools containers 'let you use the same APIs that are available in Azure, on-premises', bringing the tools 'closer to your data for compliance, security or other operational reasons', and lists 'Control over data: Choose where your data gets processed by Foundry Tools. This can be essential if you can't send data to the cloud but need access to Foundry Tools APIs.' Key Phrase Extraction is one of the Language containers it lists as generally available. The containers still send metering information to Azure for billing.

Why not the others

AMicrosoft says 'Container support is currently available for a subset of Foundry Tools', so the feature the court needs has to be on the container list.
BMicrosoft says containers 'don't cap transactions per second (TPS)' and can scale up and out to handle demand if the hardware is provided.
DMicrosoft provides Docker containers for a subset of Foundry Tools so that they can run on premises; cloud APIs are not the only option.

Source: What are Foundry Tools containers? (Microsoft Learn).

Question 5

Multiple response

Domain 3 / Task 3.1 Align with Microsoft responsible AI policies

A city housing department is about to launch an AI system that ranks applications for emergency housing so that caseworkers know which to review first. To meet a launch date set by the mayor's office, the project sponsor proposes skipping the responsible AI review. Which two points should the AI lead make? Choose two.

A Because the ranking affects people's lives, applicants and caseworkers need to be able to understand how the system arrives at it. Correct B The ranking must not be the final authority on who is housed; caseworkers keep meaningful control, and the department stays accountable. Correct C Strong accuracy across all applicants on average shows that the ranking treats each group of applicants fairly, including the smallest groups. Your answer D Once the ranking has behaved as designed on last year's applications, reliability and safety are covered, since unusual cases are rare. Your answer E Because an outside vendor built and trained the ranking model, the vendor rather than the department answers for how it operates. Your answer

Show answerHide answer

Answers: A and B

Why

Microsoft's transparency principle says that 'When AI systems inform decisions that impact people's lives, it's critical that people understand how those decisions are made.' Its accountability principle says that 'People who design and deploy AI systems must be accountable for how those systems operate' and that accountability norms 'help ensure that AI systems are not the final authority on decisions that affect people's lives and that humans maintain meaningful control over highly autonomous systems'. A review before launch is how the department shows both.

Why not the others

CMicrosoft says AI systems 'should treat everyone fairly and avoid affecting similar groups differently', and the Responsible AI dashboard's error analysis helps identify 'cohorts (subsets) of data with a higher error rate than the overall benchmark'. A strong average can hide a small group that is ranked less accurately, so overall accuracy does not show the ranking is fair.
DReliability and safety asks that systems 'function as designed' and also 'respond safely to unexpected conditions'. The page says a system can underperform 'for rarely observed input conditions in the training data', so the fact that unusual cases are rare does not make them safe.
EMicrosoft says 'People who design and deploy AI systems must be accountable for how those systems operate'. The vendor designed the software, but the department deploys it, so accountability does not pass to the vendor alone.

Source: What is Responsible AI, Azure Machine Learning (Microsoft Learn).

Want 100 more?

The study bundle has two practice exams with 100 original questions, an answer key and an explanation for every answer, plus the 52-page study guide. Practice exam explanations are shorter than the ones on this page: a short paragraph per question, with no source links.

Study bundle

$19

100 practice questions / Study guide / Answer keys

Every bundle includes 6 months of access: download the PDFs, and any new editions, until 6 months after your purchase date. Your receipt has a license key. Enter it at overqual.com/download for the three PDFs, stamped with your email and order number. They are yours to keep.

Buy now - $19 AB-731 study bundle

Secure checkout, in a new tab.

AB-731 study bundle Buy now - $19