CertSafari

    Free Microsoft AI Transformation Leader (AB-731) - January 11 onwards Sample Questions

    35 free sample questions from our bank of 350+, covering every exam domain, with answers and detailed explanations. Updated November 2026.

    1: Identify the business value of generative AI solutions

    1.1: Identify the foundational concepts of generative AI

    1.A single pretrained language model can be applied to many tasks, such as summarizing, drafting and question answering, without task-specific training.

    1. A.True
    2. B.False
    Show answer & explanation

    Correct answer: A — True

    • A. Correct. Foundation models are general-purpose, so prompting alone often covers a range of tasks.
    • B. Incorrect. Pretrained models are designed for broad reuse, so the statement is accurate.

    1.2: Identify benefits and capabilities of generative AI solutions

    2.A bank's auditors require that every generated answer can be traced to its source material. Which grounding capability satisfies this requirement?

    1. A.Enable a self-explanation step, so the model states which training data shaped each response it generates.
    2. B.Limit retrieval to one vetted knowledge base, so responses come from a smaller, curated set of documents.
    3. C.Archive each prompt with its response and a timestamp, so reviewers can inspect the interaction history.
    4. D.Attach references to each response, so reviewers can open the retrieved passages and documents behind it.
    Show answer & explanation

    Correct answer: D — Attach references to each response, so reviewers can open the retrieved passages and documents behind it.

    • A. A model cannot reliably introspect on its training data, so a self-explanation is a generated claim, not evidence. It gives auditors no verifiable link from an answer to a source.
    • B. A single vetted knowledge base improves source quality, but an answer still carries no pointer to the passage it used. Auditors cannot verify where a specific answer came from.
    • C. Archived prompts and responses show what was asked and answered, but not which source content supported each answer. The record has no link to source material, so it does not give auditors a trail.
    • D. References that name the retrieved passages and documents tie each answer to the material it was built from. Auditors can open those sources and verify the answer against them.

    1.2: Identify benefits and capabilities of generative AI solutions

    3.Which description best defines retrieval-augmented generation (RAG)?

    1. A.Searching a document store for passages related to the query, adding them to the prompt, and generating the answer from them.
    2. B.Fine-tuning the model on passages from a document store, so its weights absorb the content and it answers from memory.
    3. C.Embedding the query to route it to the best-matched specialist model, which then answers from what it learned in training.
    4. D.Generating an answer first, then searching a document store to check it and remove claims the passages do not support.
    Show answer & explanation

    Correct answer: A — Searching a document store for passages related to the query, adding them to the prompt, and generating the answer from them.

    • A. RAG retrieves passages relevant to the query from an external store, augments the prompt with them, and has the model generate its answer from that supplied context. The model's weights are not changed.
    • B. Fine-tuning on documents updates the model's weights so it answers from learned parameters. RAG leaves the weights unchanged and supplies the retrieved content in the prompt at request time.
    • C. Embedding a query to route it to a specialist model is a routing pattern. Nothing is retrieved or added to the prompt, and the chosen model still answers only from its training data, which is what RAG avoids.
    • D. Checking an answer against a document store after it is written is post-hoc verification. In RAG the retrieval happens before generation, so the passages shape the answer instead of filtering it afterwards.

    1.2: Identify benefits and capabilities of generative AI solutions

    4.An executive asks why a RAG-based assistant costs more per request than a plain chat prompt. Which explanation is correct?

    1. A.The model is fine-tuned on the retrieved documents when a question arrives, adding a training step per call.
    2. B.Answers are held in a manual review queue before release, so reviewer labor is added to each call.
    3. C.Source citations are a separately licensed add-on that is billed per seat, which raises the price of each call.
    4. D.Retrieved passages enlarge the input prompt, and embedding and search work is added to each call.
    Show answer & explanation

    Correct answer: D — Retrieved passages enlarge the input prompt, and embedding and search work is added to each call.

    • A. RAG leaves the model weights unchanged. Retrieved documents are only inserted into the prompt at query time, so no training step runs for each call.
    • B. A RAG pipeline does not include a manual review queue. Answers are generated and returned automatically, so reviewer labor is not what raises the per-call price.
    • C. Citations come from the same retrieved passages that ground the answer. They are part of the response pattern, not a separately licensed per-seat add-on.
    • D. Retrieved passages are added to the prompt as extra input tokens. Each call also has to embed the question and search the index, work a plain chat prompt does not need.

    1.2: Identify benefits and capabilities of generative AI solutions

    5.A research team asks complex, multi-part questions that a single search query handles poorly. Which retrieval pattern is designed for this?

    1. A.Hybrid search, where keyword and vector scores are fused and a semantic ranker reorders the results.
    2. B.Agentic retrieval, where a model splits the request into subqueries and sends them out in parallel.
    3. C.Vector search, where the whole question is embedded once and the nearest chunks are returned.
    4. D.Fine-tuning, where the model is retrained on the document library on a schedule and answers from its weights.
    Show answer & explanation

    Correct answer: B — Agentic retrieval, where a model splits the request into subqueries and sends them out in parallel.

    • A. Hybrid search improves how well one query is matched and ranked, but the multi-part question still goes out as a single query and is never broken into parts.
    • B. Agentic retrieval has a model plan the work by splitting a complex request into subqueries and running them in parallel, so each part of the question gets its own search.
    • C. Embedding the whole question once yields one vector that blends all its parts, so the nearest chunks tend to cover only some of them. It does not decompose the question.
    • D. Fine-tuning changes the model's weights and is not a retrieval pattern. Answers come from what the model absorbed, with no step that searches for evidence for each part of a question.

    1.2: Identify benefits and capabilities of generative AI solutions

    6.A chatbot that summarizes web pages for staff follows hidden text on one page that tells it to reveal the contents of the user's conversation. Which threat is this?

    1. A.Data poisoning: an attacker alters records in the training or fine-tuning set so the model learns bad behavior.
    2. B.Model inversion: an attacker sends crafted queries to the model to reconstruct examples from its training data.
    3. C.Indirect prompt injection: instructions embedded in content the model fetches override how it behaves.
    4. D.Direct prompt injection: a user types instructions into the chat to override the assistant's system prompt.
    Show answer & explanation

    Correct answer: C — Indirect prompt injection: instructions embedded in content the model fetches override how it behaves.

    • A. Data poisoning corrupts the data a model learns from. Here the model is unchanged and the malicious instruction arrives at run time inside a single web page the chatbot is asked to summarize.
    • B. Model inversion tries to recover training examples through repeated crafted queries. Here nobody is querying the model for training data, and the target is the current user's conversation, exposed by an instruction the chatbot obeyed.
    • C. The instruction is planted in external content the chatbot retrieves, and the model treats it as a command, which is indirect prompt injection. Obeying it would expose the user's conversation.
    • D. Direct prompt injection comes from the person typing into the chat. Here the user wrote no malicious input; the instruction sits in a third-party page the chatbot was asked to summarize.

    1.2: Identify benefits and capabilities of generative AI solutions

    7.A bank wants to ground an internal assistant. Which TWO requirements justify choosing a grounding solution over relying on the model alone?(Select 2)

    1. A.Replies must use an energetic, youthful voice when the assistant writes marketing slogans for campaigns.
    2. B.Procedure documents are updated several times a week, and answers must reflect each update without retraining.
    3. C.Staff must be able to translate short travel phrases into several languages at foreign branches.
    4. D.The team needs to brainstorm names for an internal hackathon using open-ended creative prompts.
    5. E.Each answer must point reviewers to the bank's procedure document it came from, so claims can be checked.
    Show answer & explanation

    Correct answers: B, E — Procedure documents are updated several times a week, and answers must reflect each update without retraining.; Each answer must point reviewers to the bank's procedure document it came from, so claims can be checked.

    • A. Writing slogans in an energetic voice is a creative style task. Prompt instructions cover it, and it depends on no private or frequently changing source content, so grounding adds nothing.
    • B. Procedures that change several times a week cannot be built into model weights without constant retraining. A grounding solution supplies the latest documents at request time, so answers follow each revision.
    • C. Translating short, common travel phrases relies on the model's built-in multilingual ability. It needs no bank documents and no current private data, so grounding adds nothing here.
    • D. Brainstorming hackathon names is open-ended creative work with no facts to verify. It needs no source documents, so a grounding solution adds nothing.
    • E. Reviewers must trace each statement to the bank's own procedure document. Only a grounding solution can attach the retrieved source to the answer, which a model alone cannot reliably do.

    1.1: Identify the foundational concepts of generative AI

    8.A retailer's decision-tree chatbot only answers questions that match pre-written intents and fails on unusual phrasing. What main difference would an LLM-based assistant bring?

    1. A.It queries live inventory records while writing each reply, so stock details are current without extra setup
    2. B.It reads varied customer wording and composes its own replies, so answers are not limited to a stored list
    3. C.It holds the current product catalogue in its trained parameters, so no separate knowledge source is needed
    4. D.It applies fixed decision rules that return the same wording on each run, so replies stay consistent between sessions
    Show answer & explanation

    Correct answer: B — It reads varied customer wording and composes its own replies, so answers are not limited to a stored list

    • A. Incorrect. A language model has no built-in link to inventory systems. Current stock data appears in replies only if the team adds retrieval or an integration, so this is not a difference the model brings by itself.
    • B. Correct. An LLM interprets many ways of phrasing the same request and generates new text, so it can respond to wording that a bot limited to scripted intents cannot match.
    • C. Incorrect. Knowledge in a model's parameters is fixed at training time and goes stale. Current catalogue data still has to be supplied from a separate source such as retrieval.
    • D. Incorrect. Fixed rules with identical output describe the scripted decision-tree bot. A generative model is probabilistic, so its wording can differ between runs and sessions.

    1.1: Identify the foundational concepts of generative AI

    9.An HR team's leave policies change monthly, and employees need answers that reflect the latest version. Which approach keeps answers current with the least ongoing effort?

    1. A.Fine-tune the model on the policy documents, then repeat the tuning run and redeploy after each revision
    2. B.Retrieve passages from the policy repository at query time, and replace the stored documents after each revision
    3. C.Tune the model's temperature and top-p sampling settings, and review the answers after each revision
    4. D.Pretrain a custom model on the company's HR documents, and rebuild it from scratch at the end of each quarter
    Show answer & explanation

    Correct answer: B — Retrieve passages from the policy repository at query time, and replace the stored documents after each revision

    • A. Incorrect. Fine-tuning does bake in the policy text, but it needs a new training run and redeployment after each monthly change. That is far more ongoing work than replacing a document that the model reads at query time.
    • B. Correct. Retrieval-based grounding pulls passages from the stored documents when a question is asked, so replacing a document is enough to change the answers. No retraining or redeployment is needed, which keeps the monthly effort minimal.
    • C. Incorrect. Temperature and top-p only control how varied the wording of a response is. They add no knowledge of revised policies, so answers stay outdated and manual review becomes the ongoing burden.
    • D. Incorrect. Pretraining a custom model is the most expensive route. Rebuilding only at the end of each quarter also means a policy change made just after a rebuild stays out of the answers for almost a full quarter, which does not fit monthly updates.

    1.1: Identify the foundational concepts of generative AI

    10.A CIO has been told that fine-tuning will make the model smarter on everything. Which realistic consideration should the CIO weigh before approving it?

    1. A.It needs curated, high-quality examples, and shifting needs mean retraining and re-evaluating, which adds ongoing cost
    2. B.It trains only on company-approved examples, which removes the need to review outputs for fabricated details
    3. C.It bakes company knowledge into the weights, which keeps answers current as new company documents are published
    4. D.It specializes the model on one task, which leaves it handling general language requests poorly afterward
    Show answer & explanation

    Correct answer: A — It needs curated, high-quality examples, and shifting needs mean retraining and re-evaluating, which adds ongoing cost

    • A. Correct. Fine-tuning depends on a good set of labelled examples and has to be retrained and re-evaluated when requirements change, so it costs more than using a pretrained model as is.
    • B. Incorrect. Training on approved examples does not stop a model from inventing details. Fine-tuned models can still fabricate, so their outputs still need grounding and human review.
    • C. Incorrect. Fine-tuned weights reflect only the data present at training time. Documents published later reach the model only through retrieval or another round of training.
    • D. Incorrect. Fine-tuning generally keeps the model's broad language ability while improving the targeted behavior. It does not leave the model handling general requests poorly.

    1.1: Identify the foundational concepts of generative AI

    11.A finance manager asks why usage of a language model is not billed by word count. How do language models measure text for usage and cost?

    1. A.They count the sentences in each prompt and reply, and bill on the total number of sentences counted
    2. B.They split text into tokens, which can be word pieces, short words or punctuation, and bill on the token total
    3. C.They log each request as one call, and bill on the number of calls made, whatever the text length
    4. D.They measure each prompt and reply in kilobytes of data, and bill on the total volume transferred
    Show answer & explanation

    Correct answer: B — They split text into tokens, which can be word pieces, short words or punctuation, and bill on the token total

    • A. Incorrect. Sentences are not a unit models process or meter, and a long sentence takes far more processing than a short one. Usage is metered in tokens, not sentence counts.
    • B. Correct. Models read and generate text as tokens, which can be word fragments, whole short words or punctuation marks. Usage and cost are metered in tokens, which is why word counts do not match the bill.
    • C. Incorrect. Per-call billing would ignore how much text is processed. Usage is metered in tokens, so a long document costs more than a one-line question.
    • D. Incorrect. Data size in kilobytes is not how text usage is measured. The same number of bytes can map to different token counts, and billing follows the tokens read and generated.

    1.1: Identify the foundational concepts of generative AI

    12.A team expects sporadic automated summarization by an internal app, with thousands of calls on some days and none on others. Which cost alignment is best?

    1. A.On-demand pay-per-token pricing, billed on the amount of text processed in each summarization request
    2. B.Per-seat subscription licensing for the whole staff, billed monthly at a fixed rate for each named user
    3. C.Provisioned throughput on a one-year commitment, billed hourly at a fixed rate for reserved model capacity
    4. D.Metered storage pricing, billed monthly on the gigabytes of source documents kept in the internal library
    Show answer & explanation

    Correct answer: A — On-demand pay-per-token pricing, billed on the amount of text processed in each summarization request

    • A. Correct. On-demand per-token billing charges only for text actually processed, so spend rises on busy days with thousands of requests and is zero on days with none.
    • B. Incorrect. Seat licences charge a constant amount per named user, which has no link to how many automated requests the app makes. The team would pay the same on quiet days as on peak days, and the licence prices the wrong unit.
    • C. Incorrect. Provisioned throughput is billed for reserved capacity around the clock under a fixed commitment, so the team would keep paying on the many days with no requests. It suits steady, predictable load, not bursts.
    • D. Incorrect. Storage metering follows the size of the document library, which stays roughly constant. It does not follow the amount of summarization performed, so spend would not track the bursts and gaps in usage.

    2: Identify benefits, capabilities, and opportunities for Microsoft’s AI apps and services

    2.1: Identify benefits and capabilities of Microsoft Copilot

    13.A manufacturer wants Copilot to return records from an external quality system that is not part of Microsoft 365. The standard way to bring that content into Copilot responses is a Microsoft Graph ______.

    1. A.sensitivity tier that upgrades the quality system into a native Microsoft 365 app
    2. B.connector that indexes external content so permitted users get it in answers
    3. C.retention label that moves the quality records into every user's mailbox
    Show answer & explanation

    Correct answer: B — connector that indexes external content so permitted users get it in answers

    • A. Incorrect. There is no sensitivity tier that turns an external system into a Microsoft 365 app.
    • B. Correct. Graph connectors let Copilot reference third-party content, returned when the user has access.
    • C. Incorrect. Retention labels manage lifecycle and don't import external data.

    2.2: Identify benefits and capabilities of Foundry Tools

    14.A telecom company stores 20,000 hours of recorded customer calls and wants written transcripts so analysts can search them. Which Foundry Tool should the business map to this need?

    1. A.Azure Language in Foundry Tools, using key phrase extraction to index what is said across the recorded calls
    2. B.Azure Translator, using document translation to convert each stored call file into a written version
    3. C.Azure Document Intelligence, using its read model to extract text from the audio files of the calls
    4. D.Azure Speech in Foundry Tools, using batch transcription to turn the call audio files into text records
    Show answer & explanation

    Correct answer: D — Azure Speech in Foundry Tools, using batch transcription to turn the call audio files into text records

    • A. Key phrase extraction analyzes text that already exists. The calls are only audio, so nothing can be indexed until speech has been converted to text, and this tool does not do that conversion.
    • B. Document translation converts the language of existing text documents. It cannot listen to a call file and produce a written version of what was said, so the transcripts would still be missing.
    • C. The read model extracts printed and handwritten text from documents and images. It does not accept audio, so it cannot return the spoken words from a call recording.
    • D. Azure Speech batch transcription converts large collections of stored audio files into text. That suits 20,000 hours of calls, and analysts can then search the resulting text.

    2.2: Identify benefits and capabilities of Foundry Tools

    15.Which description best captures the layered security model that Foundry Tools provide?

    1. A.Encryption of traffic in transit and data at rest, plus Azure region restrictions that decide which callers are trusted
    2. B.Authentication with one shared administrator login for the subscription, plus Azure virtual network restrictions
    3. C.Anonymous access to the endpoint by default, plus Azure Policy rules and application checks that secure each call
    4. D.Authentication with Microsoft Entra credentials or a valid resource key, plus Azure virtual network restrictions
    Show answer & explanation

    Correct answer: D — Authentication with Microsoft Entra credentials or a valid resource key, plus Azure virtual network restrictions

    • A. Encryption protects data, but it does not establish who is calling. A caller's region is not an identity check, and Foundry Tools require each request to be authenticated with a resource key or Microsoft Entra credentials.
    • B. Foundry Tools do not use a shared administrator login for the subscription to authenticate calls. Requests are authenticated with a resource key or Microsoft Entra credentials. Virtual network restrictions are a real layer, but the identity layer described here is wrong.
    • C. Foundry Tools do not allow anonymous access by default. Each request must carry a resource key or Microsoft Entra credentials. Azure Policy governs resource configuration and does not authenticate calls, and security is not left to the calling application.
    • D. Foundry Tools documentation describes authentication with Entra credentials, a valid resource key, and Azure Virtual Networks. These layers work together.

    2.2: Identify benefits and capabilities of Foundry Tools

    16.A bank wants a spoken, real-time assistant for branch kiosks and does not want to build its own voice orchestration infrastructure. Which capability matches?

    1. A.Azure AI Content Understanding audio analyzers, applied to kiosk recordings once each session ends
    2. B.Azure AI Search semantic ranking over the bank's FAQ content, with answers shown as on-screen text
    3. C.A voice-enabled prompt agent in Foundry Agent Service, connected to the kiosk through Voice Live
    4. D.A text-based prompt agent in Foundry Agent Service, plus voice orchestration code the bank hosts itself
    Show answer & explanation

    Correct answer: C — A voice-enabled prompt agent in Foundry Agent Service, connected to the kiosk through Voice Live

    • A. Content Understanding audio analyzers extract structure from recorded audio after the fact. They do not hold a live spoken conversation, so they cannot serve as a real-time voice assistant.
    • B. Semantic ranking improves retrieval of FAQ content, but the answers appear as text on screen. Nothing in this design speaks to or listens to the visitor, so the spoken requirement is not met.
    • C. Foundry Agent Service offers voice-enabled prompt agents that connect through Voice Live. The platform runs the real-time speech handling, so the bank does not build or host a voice orchestration layer.
    • D. A text-based agent has no speech capability of its own, so the bank would write and host the voice orchestration code. That is the infrastructure it wants to avoid.

    2.2: Identify benefits and capabilities of Foundry Tools

    17.A retailer needs to read text from standard receipts and wants value within weeks. When is a prebuilt Foundry Tools model the better choice over training a custom model?

    1. A.When the layouts are specific to the company and no prebuilt model covers them, and the team can label samples
    2. B.When the documents are a common type that a prebuilt model already covers, and no labeled samples are needed
    3. C.When policy bars sending documents to a managed cloud service, and the team must build and host the model itself
    4. D.When the company already holds thousands of labeled examples and wants full control over how the model is trained
    Show answer & explanation

    Correct answer: B — When the documents are a common type that a prebuilt model already covers, and no labeled samples are needed

    • A. Company-specific layouts that no prebuilt model covers are the case for custom training. Standard receipts are a common document type that prebuilt models already read, so this does not describe the retailer.
    • B. Prebuilt models work out of the box on common document types such as standard receipts. The retailer can start extracting text without collecting or labeling training data, which fits a timeline of a few weeks.
    • C. A policy against managed cloud services rules out the prebuilt cloud API as well as managed training. Building and hosting a model in-house also works against the retailer's goal of value within weeks.
    • D. A large labeled dataset and a wish for training control point toward a custom model. The retailer reads standard receipts and wants quick results, so neither factor favors a prebuilt model.

    2.2: Identify benefits and capabilities of Foundry Tools

    18.An HR assistant built on Azure AI Search must never show a manager's compensation documents to employees who lack permission. Which design principle addresses this?

    1. A.Filter results by a department value the employee picks in the chat window before the query is sent
    2. B.Move compensation files into a separate index that the assistant queries with its own service identity
    3. C.Add a system prompt rule telling the model to omit compensation files for users who are not in HR
    4. D.Filter results on the caller's group memberships, matched to permission fields stored with each document
    Show answer & explanation

    Correct answer: D — Filter results on the caller's group memberships, matched to permission fields stored with each document

    • A. A department value chosen in the chat window is controlled by the employee, not tied to their identity or entitlements. Anyone could pick a different value, so unauthorized employees could still reach compensation documents.
    • B. A separate index only isolates the files. The assistant reads it under its own identity, which has access regardless of who asks, so any employee could surface compensation documents through it. No per-user permission check is applied.
    • C. A prompt rule is an instruction to the model, not access control. The restricted documents are still retrieved into the model's context and can leak through rephrased questions or prompt injection.
    • D. Security trimming adds a query-time filter that compares the caller's group memberships with permission fields stored on each indexed document. Restricted documents are never returned to the assistant for users who lack access.

    2.2: Identify benefits and capabilities of Foundry Tools

    19.A bank is deciding when a Foundry agent is more appropriate than a single model call. Which TWO situations favor an agent?(Select 2)

    1. A.The app must translate a customer's message into another language and return one reply per request
    2. B.The assistant must call tools and act on external systems, such as retrieving account data before replying
    3. C.The bank wants each reply written in a fixed tone and format that a system message can set per request
    4. D.The workflow involves multi-step reasoning, with memory and retrieved knowledge persisting across turns of a chat
    5. E.The app fills a prompt template with form fields and shows the model's summary in a single pass
    Show answer & explanation

    Correct answers: B, D — The assistant must call tools and act on external systems, such as retrieving account data before replying; The workflow involves multi-step reasoning, with memory and retrieved knowledge persisting across turns of a chat

    • A. Translating one message into one reply is a single-step task with no external actions. One model call, or a dedicated translation service, handles it without an agent.
    • B. An agent combines a model with instructions and tools, so it can reach external systems and act on them. Fetching account data before replying needs that.
    • C. Fixing tone and format only needs a system message, which a single model call accepts. Nothing here requires tools, memory or multi-step orchestration, so an agent adds no benefit.
    • D. Agents support multi-step reasoning with memory and retrieval, so context and retrieved knowledge persist across a conversation. A one-off model call keeps neither.
    • E. Filling a template and showing one summary is a single-pass task with no tools or orchestration. A single model call suffices, so an agent is unnecessary.

    2.1: Identify benefits and capabilities of Microsoft Copilot

    20.A field sales director travelling frequently wants to ask Copilot work questions from a phone and keep enterprise-grade data protection on prompts and responses. What should they do?

    1. A.Use the Microsoft Copilot mobile app with a personal Microsoft account, which carries the same contractual data protection terms as a work sign-in.
    2. B.Use the Microsoft Copilot mobile app with their Entra work account, which places their chats under the tenant's Enterprise Data Protection.
    3. C.Open the Copilot website in a private browser tab on the phone without signing in, which keeps prompts outside tenant retention and auditing.
    4. D.Have IT publish a Copilot Studio agent to the field team's phones, since Copilot Chat is not offered as a mobile app for work accounts.
    Show answer & explanation

    Correct answer: B — Use the Microsoft Copilot mobile app with their Entra work account, which places their chats under the tenant's Enterprise Data Protection.

    • A. Incorrect. A personal Microsoft account gets consumer Copilot handling, not the contractual enterprise protection tied to Entra work sign-in. Work questions would sit outside what the director's organization governs.
    • B. Correct. The Microsoft Copilot mobile app supports Entra work sign-in, and Enterprise Data Protection applies to prompts and responses when users are signed in with their work identity.
    • C. Incorrect. A signed-out session has no work identity, so Enterprise Data Protection cannot apply. Skipping sign-in removes the organization's protection and governance instead of adding to it.
    • D. Incorrect. Copilot Chat is available in the Microsoft Copilot mobile app for Entra work accounts, so no custom Copilot Studio agent is needed. Building and publishing one would add effort for a capability that already exists.

    2.1: Identify benefits and capabilities of Microsoft Copilot

    21.A company uses a third-party ticketing system and wants Microsoft 365 Copilot to include ticket content in answers, only for people allowed to see each ticket. Which build-buy-extend choice is right?

    1. A.Export tickets weekly to a OneDrive folder shared with the support department, so Copilot can ground its answers on the files.
    2. B.Turn on Copilot's web search grounding, so it retrieves ticket pages from the ticketing system when users ask about them.
    3. C.Register a Microsoft Graph connector or Copilot agent for the ticketing system, so Copilot returns its tickets to users with access.
    4. D.Fine-tune a custom model on the ticket history in Azure AI Foundry, so support staff query it instead of Copilot for ticket questions.
    Show answer & explanation

    Correct answer: C — Register a Microsoft Graph connector or Copilot agent for the ticketing system, so Copilot returns its tickets to users with access.

    • A. Incorrect. Sharing the exported files with a whole department gives every member access to tickets they may not be allowed to see. The ticketing system's per-ticket permissions are lost, and the weekly export leaves the content stale.
    • B. Incorrect. Web search grounding draws on public web content. It cannot sign in to a private ticketing system, so ticket content would not be retrieved, and it does not apply the system's per-ticket permissions.
    • C. Correct. Graph connectors and agents are the supported ways to extend Copilot to third-party systems. External ticket data then appears in responses for users who have permission to see it.
    • D. Incorrect. A model fine-tuned on ticket history absorbs the content without per-user access trimming, so users could see tickets they are not allowed to open. It also builds a new model where an extension of Copilot would do.

    2.1: Identify benefits and capabilities of Microsoft Copilot

    22.A 50-person firm only needs meeting summaries, email drafting and document creation. It has no proprietary systems to integrate and no developers. What should it do?

    1. A.Assign Microsoft 365 Copilot licenses to staff and use the Copilot features that ship inside Teams, Outlook and Word for these tasks.
    2. B.Author a Copilot Studio agent that summarizes Teams meetings and drafts Outlook emails, then publish it to staff in Microsoft 365 Copilot.
    3. C.Develop a custom engine agent in Microsoft Foundry that handles meeting summaries, email drafting and documents, then deploy it into Teams.
    4. D.Set up Microsoft Graph connectors to index the firm's external data sources, then roll out Copilot to staff for these tasks.
    Show answer & explanation

    Correct answer: A — Assign Microsoft 365 Copilot licenses to staff and use the Copilot features that ship inside Teams, Outlook and Word for these tasks.

    • A. Correct. Meeting summaries, email drafting and document creation are built into Microsoft 365 Copilot in Teams, Outlook and Word, so licensing it meets the need without building anything.
    • B. Incorrect. Copilot Studio would rebuild summarizing and drafting that Copilot already does in the apps, and the firm has nobody to author or maintain such an agent.
    • C. Incorrect. A custom engine agent is a development effort, and the firm has no developers. It would also reproduce features that are already built into Copilot.
    • D. Incorrect. Graph connectors index external or proprietary data, and the firm has none. Copilot already reaches Microsoft 365 content through Graph, so the connectors would add nothing.

    2.1: Identify benefits and capabilities of Microsoft Copilot

    23.A telecom wants an agent that answers customer questions and resolves issues over voice and chat, escalating to a human representative when it cannot resolve a request. Which Copilot Studio option is designed for this?

    1. A.The Microsoft Copilot harness, built for reasoning-heavy work across internal business processes and for employees' complex multi-step tasks.
    2. B.The Researcher agent in Microsoft 365 Copilot, given the contact center's call transcripts so it can answer customer calls as they come in.
    3. C.The Customer Conversation harness, designed around a primary agent that works with specialized sub-agents and can pass a case to a live person.
    4. D.GitHub Copilot Chat, connected by developers to the telephony system so it can answer customers and pass conversations to support staff.
    Show answer & explanation

    Correct answer: C — The Customer Conversation harness, designed around a primary agent that works with specialized sub-agents and can pass a case to a live person.

    • A. Incorrect. This harness is aimed at reasoning-heavy internal business processes and employee tasks. The telecom needs a customer-facing agent for voice and chat, which is a different scenario.
    • B. Incorrect. Researcher is a research agent for employees in Microsoft 365 Copilot. Giving it call transcripts does not make it a customer-facing agent that takes live voice and chat conversations or hands them to a representative.
    • C. Correct. The Customer Conversation harness is the Copilot Studio option for customer-facing voice and chat agents. A primary agent coordinates specialized sub-agents and can escalate to a human representative.
    • D. Incorrect. GitHub Copilot Chat is a developer coding assistant, not a Copilot Studio customer service option. Wiring it into telephony would be custom work with no built-in conversation handling or escalation to a representative.

    2.1: Identify benefits and capabilities of Microsoft Copilot

    24.A strategy team lists tasks to give to the Researcher agent in Microsoft 365 Copilot. Which TWO tasks fit Researcher best?(Select 2)

    1. A.Generate unit tests and refactor modules in a developer's Python service, working from the repository open in the IDE.
    2. B.Prepare a go-to-market brief for leadership that pulls together market trends, synthesizes findings and recommends next steps.
    3. C.Triage a flood of security alerts, correlate them with threat intelligence and propose containment steps for the incident.
    4. D.Produce a competitive landscape report that combines public industry news with internal emails and decks, with cited sources.
    5. E.Clean three raw CSV exports, run regression analysis iteratively and chart the residuals to show the data pattern.
    Show answer & explanation

    Correct answers: B, D — Prepare a go-to-market brief for leadership that pulls together market trends, synthesizes findings and recommends next steps.; Produce a competitive landscape report that combines public industry news with internal emails and decks, with cited sources.

    • A. Incorrect. Writing tests and refactoring code inside the developer's IDE and repository is a coding-assistant scenario, which GitHub Copilot covers. Researcher produces research deliverables from web and work content, not code changes.
    • B. Correct. Collecting market information, synthesizing it and turning it into a recommendation brief for leadership is multistep research, which is what Researcher is built for.
    • C. Incorrect. Alert triage is a Security Copilot scenario.
    • D. Correct. A competitive report that draws on both public web sources and internal emails and presentations, with citations, is a multistep research task across web and work data. That is what Researcher does.
    • E. Incorrect. Iterative raw data analysis and charting is an Analyst task.

    3: Identify an implementation and adoption strategy for Microsoft’s AI apps and services

    3.1: Align an AI strategy with Microsoft responsible AI policies

    25.The legal team asks whether prompts and documents processed by Microsoft 365 Copilot are used to train the underlying foundation models. What should the AI strategy lead report?

    1. A.Training use depends on each user, because individuals can opt in to share their prompts with model training pipelines
    2. B.Prompts do train the foundation models unless an administrator buys a premium data-isolation license for the whole tenant
    3. C.Prompts, responses, and data accessed through Microsoft Graph are not used to train the foundation large language models
    4. D.Only documents carrying sensitivity labels are excluded, while unlabeled documents are used to train the foundation models
    Show answer & explanation

    Correct answer: C — Prompts, responses, and data accessed through Microsoft Graph are not used to train the foundation large language models

    • A. Incorrect: there is no per-user opt-in that feeds tenant prompts into foundation model training.
    • B. Incorrect: no extra license is required, because the commitment not to train on this data applies by default.
    • C. Correct: Microsoft states that prompts, responses, and Graph-accessed data are not used to train the foundation LLMs behind Copilot.
    • D. Incorrect: labels control protection and access, but they are not a switch deciding whether data trains the models.

    3.2: Plan for AI adoption across the organization

    26.Microsoft Copilot Chat is available to users with eligible Microsoft 365 subscriptions without buying a separate per-user Copilot license.

    1. A.True
    2. B.False
    Show answer & explanation

    Correct answer: A — True

    • A. True: Copilot Chat is included with eligible subscriptions for Entra-signed-in users, while the paid add-on provides the in-app Copilot experience.
    • B. False would be wrong because Copilot Chat is part of eligible Microsoft 365 subscriptions rather than a separate per-user purchase.

    3.2: Plan for AI adoption across the organization

    27.An AI champions program removes the need for a formal adoption team, because peer advocates can cover technical readiness, licensing and governance on their own.

    1. A.True
    2. B.False
    Show answer & explanation

    Correct answer: B — False

    • A. This is not true: champions help with peer enablement, whereas readiness, licensing, security and governance need a formal cross-functional team.
    • B. Correct: champions complement the adoption team by driving peer uptake and feedback, but they do not replace its readiness and governance duties.

    3.1: Align an AI strategy with Microsoft responsible AI policies

    28.A procurement team evaluating an AI vendor asks what a service is intended for, its limitations, and how to deploy it responsibly. Which Microsoft artifact answers these questions for Microsoft's AI services?

    1. A.Transparency Notes, which describe a service's purpose, performance boundaries, and guidance for deploying it safely
    2. B.Responsible AI Standard, which sets the internal requirements Microsoft's engineering teams follow when building AI systems
    3. C.Service Trust Portal audit reports, which document a service's security controls and compliance certifications for customers
    4. D.Service-level agreements, which commit uptime percentages and service credits if availability falls below the stated target
    Show answer & explanation

    Correct answer: A — Transparency Notes, which describe a service's purpose, performance boundaries, and guidance for deploying it safely

    • A. Correct: Transparency Notes are Microsoft's per-service documents covering what an AI service is for, where its performance has limits, and what to consider when deploying it responsibly.
    • B. Incorrect: the Responsible AI Standard is a company-wide framework governing how Microsoft's own teams build AI. It does not describe the intended uses or limitations of a particular service a customer is evaluating.
    • C. Incorrect: Service Trust Portal reports attest to security and compliance posture through audits and certifications. They do not explain what an AI service is meant for, where it falls short, or how to deploy it responsibly.
    • D. Incorrect: an SLA covers availability commitments and says nothing about intended use or limitations.

    3.1: Align an AI strategy with Microsoft responsible AI policies

    29.Security leaders worry that users or documents may embed instructions that trick a company's generative AI app into ignoring its rules. Which Azure AI Content Safety capability targets this risk?

    1. A.Groundedness detection, which compares generated answers with the supplied source documents and flags statements they do not support
    2. B.Prompt Shields, which scans user prompts and attached documents for jailbreak attempts and hidden directives before the model responds
    3. C.Protected material detection, which scans generated text for matches to known copyrighted works such as song lyrics or news articles
    4. D.Custom categories, which let teams train filters for organization-specific content that the predefined categories do not cover
    Show answer & explanation

    Correct answer: B — Prompt Shields, which scans user prompts and attached documents for jailbreak attempts and hidden directives before the model responds

    • A. Incorrect: groundedness detection checks whether an answer is backed by the supplied sources, which addresses fabricated content. It does not inspect inputs for attempts to override the application's rules.
    • B. Correct: Prompt Shields analyzes user prompts for jailbreak attempts and documents for hidden directives, which are the direct and indirect attacks that try to make the model disregard its rules.
    • C. Incorrect: protected material detection examines model output for copyrighted text. It reduces copyright exposure but does nothing to detect attempts to subvert the application's rules.
    • D. Incorrect: custom categories add organization-specific content types to be filtered. They are not designed to recognize jailbreak attempts or instructions hidden in documents.

    3.1: Align an AI strategy with Microsoft responsible AI policies

    30.Which TWO responsibilities fit an AI council that guides strategy, oversight, and cross-functional alignment?(Select 2)

    1. A.Connect AI priorities to the company's business objectives and coordinate the departments adopting AI
    2. B.Perform the contract review for AI vendor agreements in place of the legal department's specialists
    3. C.Assess high-risk AI use cases against internal policy and decide if and on what terms they proceed
    4. D.Assign Microsoft 365 Copilot seats to individual employees and reconcile license usage each month
    5. E.Author and maintain the instructions for the Copilot agents that business teams build
    Show answer & explanation

    Correct answers: A, C — Connect AI priorities to the company's business objectives and coordinate the departments adopting AI; Assess high-risk AI use cases against internal policy and decide if and on what terms they proceed

    • A. Correct: linking AI priorities to business objectives and coordinating the departments that adopt AI is the council's strategy and cross-functional role.
    • B. Incorrect: a council guides and coordinates with legal on AI risk, but it does not stand in for the legal department's specialist contract review.
    • C. Correct: assessing high-risk use cases against policy and deciding whether, and under what conditions, they proceed is a core oversight function of the council.
    • D. Incorrect: assigning seats and reconciling license usage is routine IT administration, not strategy, oversight, or cross-functional alignment.
    • E. Incorrect: writing and maintaining agent instructions is hands-on build work for the teams that own each agent, not a duty of a governing council.

    3.1: Align an AI strategy with Microsoft responsible AI policies

    31.Which TWO are among Microsoft's six responsible AI principles?(Select 2)

    1. A.Explainability
    2. B.Auditability
    3. C.Inclusiveness
    4. D.Sustainability
    5. E.Reliability and safety
    Show answer & explanation

    Correct answers: C, E — Inclusiveness; Reliability and safety

    • A. Incorrect: explainability is not a named principle. Microsoft's six are fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability. Making AI behaviour understandable falls under transparency.
    • B. Incorrect: auditability is not one of the six principles. Microsoft covers oversight and answerability for AI systems under accountability, which is the named principle.
    • C. Correct: inclusiveness is one of the six principles.
    • D. Incorrect: Microsoft has environmental sustainability commitments, but sustainability is not one of the six responsible AI principles.
    • E. Correct: reliability and safety is one of the six principles.

    3.2: Plan for AI adoption across the organization

    32.A 12,000-person manufacturer is about to start its Microsoft Copilot rollout. Leadership asks who should staff the adoption team so that it can handle both technical readiness and user enablement. Which composition fits best?

    1. A.IT administrators, security engineers and a deployment partner's consultants, with each plant's line managers handling usage questions and training as they arise
    2. B.An executive sponsor, IT and security administrators, a compliance representative, and learning and communications staff, with business leads owning the scenarios
    3. C.Data scientists and ML engineers, with an external AI vendor supporting them, fine-tuning models and building pipelines around the company's plant and supply data
    4. D.HR, internal communications and learning and development staff running training sessions and a department champions network, with IT consulted on licensing
    Show answer & explanation

    Correct answer: B — An executive sponsor, IT and security administrators, a compliance representative, and learning and communications staff, with business leads owning the scenarios

    • A. Incorrect: this team covers the technical side, but nobody is accountable for enablement. Line managers answering questions as they arise is not a planned program of training, communications and scenario ownership for 12,000 people, and compliance has no seat.
    • B. Correct: a cross-functional team pairs sponsorship, IT, security and compliance for readiness with learning, communications and business leads for enablement. Together they cover both needs for a 12,000-person organization.
    • C. Incorrect: Copilot adoption centers on readiness and user enablement, not on building or fine-tuning models and custom ML pipelines. This team has no one for training, communications, compliance or business scenarios, so enablement is not covered.
    • D. Incorrect: this team can run training and communications, but it leaves technical readiness to a consulted IT group. Security, data access and configuration work needs administrators on the team itself, and compliance and business owners are missing.

    3.2: Plan for AI adoption across the organization

    33.An organization wants every employee to have secure, work-grounded AI chat at no additional per-user cost, and already holds eligible Microsoft 365 subscriptions. Which offering meets this requirement?

    1. A.Microsoft Copilot Business, a paid add-on for organizations of up to 300 users, bought as an annual commitment per seat
    2. B.Foundry Tools commitment tiers, an Azure pricing plan that discounts API service usage in return for a fixed monthly fee
    3. C.Microsoft 365 Copilot, a paid add-on license assigned in the admin center to the employees who use Copilot in Office apps
    4. D.Microsoft 365 Copilot Chat, included for users who sign in with Entra accounts on qualifying Microsoft 365 plans, with no seat license
    Show answer & explanation

    Correct answer: D — Microsoft 365 Copilot Chat, included for users who sign in with Entra accounts on qualifying Microsoft 365 plans, with no seat license

    • A. Incorrect: Copilot Business is a paid, per-seat add-on aimed at organizations of up to 300 users, so it adds a per-user cost. Employees on eligible Microsoft 365 plans can already get chat without buying it.
    • B. Incorrect: Foundry Tools commitment tiers are an Azure pricing plan that discounts usage of specific API services. They are billed as an extra fixed fee and are not an employee chat offering tied to existing Microsoft 365 subscriptions.
    • C. Incorrect: Microsoft 365 Copilot is a paid per-user add-on that unlocks Copilot inside the Office apps and deeper work grounding. Choosing it adds per-user cost, and secure chat does not depend on that license.
    • D. Correct: Microsoft 365 Copilot Chat is available to Entra-signed-in users on qualifying Microsoft 365 plans, with enterprise data protection and no separate seat license, so it adds no per-user cost.

    3.2: Plan for AI adoption across the organization

    34.A legal firm needs Copilot built into Word, Outlook and Teams that reasons over its documents, email and meetings for 150 attorneys, out of 3,000 employees. Which licensing step enables this?

    1. A.Buy a Foundry Tools commitment tier and name the attorney group in the plan, adding Copilot to Word, Outlook and Teams
    2. B.Pin Copilot Chat in the Microsoft 365 apps for the attorney group, using the access already included in the subscription
    3. C.Assign the paid Microsoft 365 Copilot per-user license to the attorney group, added on top of an eligible base subscription
    4. D.Enable pay-as-you-go billing for Copilot on a linked Azure subscription, metering the attorney group's usage in Word and Outlook
    Show answer & explanation

    Correct answer: C — Assign the paid Microsoft 365 Copilot per-user license to the attorney group, added on top of an eligible base subscription

    • A. Incorrect: Foundry Tools commitment tiers discount usage of AI service APIs. Naming users in such a plan does not license anyone for Copilot inside Word, Outlook or Teams.
    • B. Incorrect: Copilot Chat comes with the subscription at no extra cost, but it is not the in-app Copilot that reasons over a user's documents, email and meetings. That needs the paid per-user license.
    • C. Correct: Copilot in the Microsoft 365 apps, grounded in a user's documents, email and meetings, comes from the paid per-user Copilot license added to an eligible base subscription. Only the attorneys need it.
    • D. Incorrect: pay-as-you-go billing meters usage of agents and similar Copilot capabilities through Azure. It does not give users the in-app Copilot experience over their Microsoft 365 work content, which requires the per-user license.

    3.2: Plan for AI adoption across the organization

    35.A university's adoption team has run a pilot and now plans scale-out. Which activity keeps the programme sustainable after launch day?

    1. A.Move user support to the vendor's helpdesk and cut internal messaging to release notes, with vendor documentation covering guidance
    2. B.Standardise on the pilot's approved scenarios and training content from the second quarter, with changes reviewed once a year
    3. C.Keep a champions community trading scenarios and feedback, with refresher training and regular reviews of usage metrics
    4. D.Hand day-to-day adoption to the IT service desk once licences are assigned, with the adoption team stood down after a closing report
    Show answer & explanation

    Correct answer: C — Keep a champions community trading scenarios and feedback, with refresher training and regular reviews of usage metrics

    • A. Incorrect: a vendor helpdesk and documentation cover product questions, not the university's own change management, role-specific coaching and internal communications that a scale-out beyond the pilot needs.
    • B. Incorrect: scenarios and training built for the pilot group do not fit the new roles and departments joining in the scale-out, and a yearly review is too slow to correct low uptake or add new use cases.
    • C. Correct: peer champions who share scenarios and feedback, refresher training and regular usage-metric reviews keep enablement and measurement running beyond the launch and feed improvements back into the programme.
    • D. Incorrect: a service desk resolves tickets but does not drive enablement, peer sharing or measurement; with the adoption team stood down once licences are assigned, nothing maintains usage and value as the rollout widens.

    Want the full experience?

    These are just samples. Practice the full Microsoft AI Transformation Leader (AB-731) - January 11 onwards question bank in quiz mode — free, no signup, with domain practice and exam simulation.