CertArea
All Microsoft exams

Free AB-100 practice test: 30 questions with answers

30 original questions for Agentic AI Business Solutions Architect, drawn from every topic area. Answer each one to see whether you got it right and why. No sign-up.

Free questions
30
Full bank
100 questions

AB-100 exam guide: topics, cost and format

All 30 AB-100 questions with answers

The same questions as the test above, grouped by exam topic. Open any question to see the answer and explanation.

Design AI-powered business solutions (10 questions)

  1. Question 1

    An architect is designing a task agent that books field service appointments in a business application. Which characteristic defines a task agent?

    1. AIt is a static web page with no actions that displays product details and FAQ content to visitors
    2. BIt answers general knowledge questions from a broad public corpus, without access to tenant data, business systems or apps
    3. CIt runs indefinitely and chooses its own goals without bounds across business units, reprioritising work and widening scope
    4. DIt completes a defined set of actions to achieve a specific business task, with defined inputs and outcomes
    Show answer and explanation
    Correct answer: D

    A task agent completes a defined task using specific actions and outputs. Unbounded goal selection, general Q&A only and static pages describe other patterns.

  2. Question 2

    An architect is designing an autonomous agent that monitors supplier delays and reorders stock within approved limits. Which design element is essential to keep the agent safe?

    1. AExplicit thresholds, approval steps above limits, and audit logging of each action
    2. BUnlimited spending authority with no logs for orders placed with each supplier and region
    3. CA single shared password for the agent that is rotated manually and reused across teams in production
    4. DDisabling notifications to people when orders are placed in the audit dashboard or review queue
    Show answer and explanation
    Correct answer: A

    Autonomous agents need explicit limits, escalation above thresholds and audit trails. Unlimited authority, shared passwords and hidden actions remove necessary control.

  3. Question 3

    A Copilot Studio agent must handle questions it does not understand. Which topic design element ensures a graceful outcome?

    1. AA topic that ends the conversation without a message and closes the session in Dynamics 365 Customer Service
    2. BA topic that exposes internal error logs to the user including stack traces and connector IDs
    3. CA fallback topic that asks clarifying questions and hands off to a person when needed
    4. DA topic that repeats the same question forever, each time the user responds, with no change in routing
    Show answer and explanation
    Correct answer: C

    A fallback topic recovers gracefully by clarifying or escalating. Silent endings, loops and exposed logs damage the user experience and security.

  4. Question 4

    An agent must call an external system whose tools are exposed through the Model Context Protocol. Which design decision is most relevant to make this integration manageable?

    1. AAvoid documenting the tool contracts and rely on informal notes for payloads
    2. BGive the agent unrestricted access to every system across the company
    3. CHard-code the external server's password in each prompt and reuse it in tests
    4. DDefine the tool inputs, outputs, permissions, and error handling the agent can rely on
    Show answer and explanation
    Correct answer: D

    Clear tool contracts, permissions and error handling make MCP integrations safe and maintainable. Unrestricted access, hard-coded passwords and undocumented contracts are weaknesses.

  5. Question 5

    A business user wants an agent that reads a web portal and fills in a form that has no API. Which Copilot Studio capability is designed for this kind of task?

    1. AA static FAQ page that lists common questions and answers for customers in a fixed layout
    2. BA standard keyword search topic that matches phrases in articles
    3. CA calendar reminder that notifies staff of upcoming meetings, deadlines and shift changes
    4. DComputer use, which lets an agent operate apps and websites to complete tasks
    Show answer and explanation
    Correct answer: D

    Computer use allows an agent to operate apps and websites when no API exists. Keyword topics, static FAQs and reminders do not interact with the interface.

  6. Question 6

    A company is choosing how a customer-facing bot should understand intent. The bot must handle a fixed set of well-known requests reliably with predictable responses. Which approach fits best in Copilot Studio?

    1. AConversational language understanding with defined topics for those requests
    2. BA free-form generative approach with no topics, letting the model compose each answer
    3. CNo intent handling, with every message sent to a person for manual review and triage
    4. DRandom responses to keep the bot interesting by varying tone and wording each session
    Show answer and explanation
    Correct answer: A

    For a fixed set of known requests, defined topics with conversational language understanding give predictable behaviour. Free-form generation is less predictable, and random responses are unsuitable.

  7. Question 7

    An architect designs a Dynamics 365 Contact Center integration where an agent must hand off to a human agent during an active chat. Which design element is required for a smooth handover?

    1. AHiding the transfer from the human agent and showing the chat as a new session
    2. BEnding the chat and asking the customer to call back later through the IVR
    3. CAsking the customer to repeat everything after the transfer to the agent
    4. DPassing the conversation context and customer details to the human agent with the transfer
    Show answer and explanation
    Correct answer: D

    Passing context with the transfer avoids repetition and speeds resolution. Repeating everything, ending the chat and hiding the transfer all degrade service.

  8. Question 8

    A custom model in Microsoft Foundry will ground answers on internal documents. Which design step ensures the grounding data is appropriate for use?

    1. ADefine data processing steps that apply access controls, cleaning, and chunking aligned with the data's sensitivity
    2. BUpload every file the company owns without review from shared drives, email archives and intranet sites, regardless of sensitivity labels
    3. CRemove the metadata so the model cannot see document authors or access labels and retention tags
    4. DCopy the data to a public location for easy access on an anonymous share with open read permissions so that partners can retrieve files
    Show answer and explanation
    Correct answer: A

    Grounding data needs processing that respects access controls and quality, and chunking helps retrieval. Uploading unreviewed files, public copies and stripping metadata that governs access create risk.

  9. Question 9

    A Power Apps canvas app needs an AI step that drafts a customer reply for a human to review before sending. Which business process design is most appropriate?

    1. ASend the AI draft to customers automatically without review through the Power Apps app
    2. BGenerate the draft with an AI component, then route it for human approval before the reply is sent
    3. CStore drafts in personal notes on individual laptops and in unshared email folders
    4. DRemove the approval step to save time and let drafts reach customers faster at peak times
    Show answer and explanation
    Correct answer: B

    A draft-then-approve flow keeps a human accountable for customer communication. Automatic sending and removing approval remove that control.

  10. Question 10

    A business wants Copilot to act on behalf of users across Microsoft 365 with access limited to what each user can already see. Which principle must the design follow?

    1. AAgents inherit the signed-in user's permissions and do not gain broader access than the user
    2. BAgents are granted tenant-wide administrator rights in Microsoft 365 and Azure so they can change content
    3. CAgents ignore access controls for speed when retrieving SharePoint, Teams and Outlook content during peaks
    4. DAgents share one account for every user across the Copilot pilot group, with credentials stored in a shared file
    Show answer and explanation
    Correct answer: A

    Agents should respect the user's permissions to prevent privilege escalation. Admin-wide, shared or ignored access controls create serious exposure.

Deploy AI-powered business solutions (10 questions)

  1. Question 11

    Which TWO practices help monitor a deployed agent's behaviour over time? (Choose two.)

    1. AAssuming the agent is stable once it is published
    2. BReviewing telemetry on successful resolutions, escalations, and errors
    3. CTracking response quality through user feedback and review samples
    4. DDisabling logs to reduce storage costs
    5. EIgnoring usage patterns because they are not important
    Show answer and explanation
    Correct answer: B, C

    Telemetry and feedback reveal how the agent performs and where it fails. Assuming stability, ignoring usage and disabling logs remove the means to detect problems.

  2. Question 12

    Which TWO controls help protect an AI solution against prompt manipulation? (Choose two.)

    1. AInclude the system instructions in plain text on the public web page
    2. BTrust all external content as instructions
    3. CValidate and constrain user inputs and tool use within defined boundaries
    4. DAllow the agent to execute any instruction found in user-supplied documents
    5. ERestrict tool permissions to the minimum needed for each task
    Show answer and explanation
    Correct answer: C, D

    Input constraints and minimal tool permissions limit what a manipulated prompt can cause. Executing instructions from documents and trusting external content invite injection.

  3. Question 13

    Which THREE items should an ALM process for Copilot Studio agents include? (Choose three.)

    1. AEnvironment promotion from development to test to production
    2. BSharing one maker account among all developers
    3. CEditing production agents directly without a record of changes
    4. DVersioned solution packages for agents, connectors, and actions
    5. ETest scenarios run before each promotion
    Show answer and explanation
    Correct answer: A, D, E

    Environment promotion, versioned solution packages and pre-promotion tests form a sound ALM process. Direct production edits and shared maker accounts undermine traceability.

  4. Question 14

    Which TWO are appropriate ways to verify that a custom AI model is ready for production? (Choose two.)

    1. ADefine validation criteria for accuracy, fairness, and performance on representative data
    2. BUse a holdout dataset and compare results against agreed thresholds
    3. CDeploy first and collect validation later
    4. DAccept the model because its training loss is low
    5. ETest only on the data used for training
    Show answer and explanation
    Correct answer: A, E

    Validation criteria on holdout and representative data, compared with thresholds, show readiness. Training data alone, deploy-first and low training loss do not prove generalization.

  5. Question 15

    A team is designing the lifecycle for data used by AI models and agents. Which practice ensures that changes to training or grounding data are controlled across environments?

    1. ADo not track changes to grounding data or indexed content once it is published, since the index updates itself
    2. BVersion and promote data sets through environments with approval and traceable change records
    3. CEdit the production data sets directly during business hours in the live tenant, unreviewed
    4. DKeep data changes in personal folders on individual laptops and shared USB drives as working spreadsheets
    Show answer and explanation
    Correct answer: B

    Versioned, approved and traceable data promotion keeps AI behaviour reproducible and auditable. Direct edits, personal folders and untracked changes break traceability.

  6. Question 16

    A security architect reviews an agent that reads external emails and summarizes them for a user. Which vulnerability is the main concern?

    1. ASlow email delivery times caused by mailbox throttling, transport rule latency, and message queue backlogs at peak
    2. BStorage of attachments in a folder with a different name than the mailbox display name, which is unrelated to summarization
    3. CFont rendering problems in summaries caused by non-breaking spaces, HTML entities, and inconsistent character encodings
    4. DIndirect prompt injection, where malicious email content attempts to change the agent's instructions
    Show answer and explanation
    Correct answer: D

    Email content is untrusted input that can carry instructions aimed at the model. Delivery speed, fonts and folder names are not the main security risk.

  7. Question 17

    A governance board requires an audit trail for changes to a deployed model and to the data used to ground it. Which control best provides this?

    1. AKeeping changes only in the developer's memory and in personal notes, with the rationale for each update recorded later
    2. BLogging of model version changes and data source changes with the user or pipeline identity and timestamps
    3. CA monthly verbal update from the team lead during the steering meeting, summarizing deployment activity and data refreshes
    4. DOverwriting logs when storage is full, with the oldest entries rotated out by the retention job, erasing earlier records
    Show answer and explanation
    Correct answer: B

    Logs of version and data changes with identities and timestamps satisfy audit requirements. Verbal updates, overwritten logs and memory do not provide evidence.

  8. Question 18

    An organization must validate that data used by an agent stays within the required geographic region. Which action is most appropriate?

    1. ACheck where the data is stored and processed, and confirm the service configuration keeps it in the approved region
    2. BAsk the agent to report its region in a chat message, then accept its self-reported residency status
    3. CAssume all cloud services store data in the same region by default across every tenant and workload
    4. DCopy the data to every region for redundancy and disaster recovery without reviewing data residency rules
    Show answer and explanation
    Correct answer: A

    Validating storage and processing location against the configuration is the reliable way to prove residency. Assumptions, self-reporting by the agent and uncontrolled copies do not establish compliance.

  9. Question 19

    A team wants to test a Copilot Studio agent before release. Which test strategy gives the best coverage of real use?

    1. AOnly a single greeting message sent from the Teams test pane, confirming that the welcome topic renders
    2. BTests that the agent can start and load its topics, with no checks on answer accuracy
    3. CA review of the agent's icon, colour palette, and avatar alignment in the Copilot Studio authoring canvas
    4. DEnd-to-end scenarios built from real user intents, including edge cases and escalation paths
    Show answer and explanation
    Correct answer: D

    End-to-end scenarios built from real intents, edge cases and escalations test what users will do. Greeting-only and icon reviews miss most failures.

  10. Question 20

    A company wants a sound environment strategy for AI solutions built on Power Platform and Microsoft Foundry. Which approach is most appropriate?

    1. AEach maker keeps a personal production environment in the same tenant, with its own connectors
    2. BProduction changes made directly in the Power Platform admin console without promotion
    3. CSeparate development, test, and production environments with controlled promotion of solutions
    4. DOne environment shared by all makers, test users, and production workloads in one region
    Show answer and explanation
    Correct answer: C

    Separated environments with controlled promotion protect production and support testing. Shared or personal production environments and direct edits remove that protection.

Plan AI-powered business solutions (10 questions)

  1. Question 21

    A retailer wants agents to answer product availability questions. The inventory data is refreshed nightly, and some items are updated several times a day. What should the architect assess first when reviewing data for grounding?

    1. AThe colour of the company logo on the storefront header banner image
    2. BTimeliness of the inventory data relative to how fast answers must reflect stock
    3. CThe number of agents on the existing tier-one support team across regional service centres
    4. DThe font used for headings and body text in the product catalogue PDF
    Show answer and explanation
    Correct answer: B

    Grounding quality depends on accuracy, relevance, timeliness, cleanliness and availability, and here the nightly refresh conflicts with intraday changes. Logo, staffing and fonts do not affect grounding quality.

  2. Question 22

    A company is adopting AI and wants a structured approach aligned to Microsoft guidance. Which framework should guide the adoption process?

    1. AThe AI adoption process in the Cloud Adoption Framework for Azure
    2. BA generic project charter template with no defined stages, owners or gates
    3. CA vendor's sales playbook with pricing, discounts and quota targets
    4. DThe company's annual holiday planning guide for office staff
    Show answer and explanation
    Correct answer: A

    The Cloud Adoption Framework for Azure includes an AI adoption process that guides strategy, readiness and governance. The other options are not structured adoption guidance.

  3. Question 23

    An architect must decide between building a custom agent and extending Microsoft 365 Copilot for an internal knowledge task. The task uses only content already in SharePoint and Teams, and no custom actions are needed. Which choice is most appropriate?

    1. AExport the content to a spreadsheet for manual lookups by staff
    2. BCreate a custom language model from scratch on dedicated GPU infrastructure and a separate pipeline
    3. CExtend Microsoft 365 Copilot with the existing content and minimal configuration
    4. DBuild a fully custom agent with a separate hosting platform in Azure
    Show answer and explanation
    Correct answer: C

    When the content lives in Microsoft 365 and no custom actions are needed, extending Microsoft 365 Copilot is the lighter option. A custom platform and custom model add cost and complexity with no clear benefit.

  4. Question 24

    A business case for an AI assistant must show return on investment to the finance committee. Which set of inputs is most useful?

    1. AThe count of Teams messages sent during the pilot, grouped by department, user role and hour of day and posted to dashboards
    2. BTotal cost of ownership including licensing, build, run costs, and measured benefit against an agreed baseline
    3. CThe vendor's marketing claims about productivity in published case studies, analyst briefings and promotional webinars
    4. DThe number of demonstrations given to executives during quarterly steering committee sessions, with counts recorded by unit
    Show answer and explanation
    Correct answer: B

    An ROI analysis compares total cost of ownership with measured benefits against a baseline. Demonstrations, marketing claims and message counts do not establish financial return.

  5. Question 25

    An enterprise has several AI workloads that call different models. Costs have risen because simple requests reach the most expensive model. Which design change best addresses this?

    1. ATurn off AI features for every user until spending is reviewed and budgets are reset
    2. BImplement a model router that sends requests to the most suitable model based on task complexity
    3. CAsk users to write shorter prompts and keep each request under two hundred characters
    4. DRoute every request to the largest model for consistency so that output style and tone stay uniform across teams
    Show answer and explanation
    Correct answer: B

    A model router sends each request to a model suited to its complexity and cost, which reduces spend without removing capability. The other options either increase cost or remove value.

  6. Question 26

    An architect is deciding whether to build a custom AI model or use an existing managed service for document classification. Which factor most strongly favours building a custom model?

    1. AThe model will be used for a single one-off experiment and then retired after one reporting cycle, with no further updates or retraining expected in the coming years
    2. BThe vendor offers a discount on custom development services if the contract is signed before the end of the fiscal quarter, with a rebate on renewals
    3. CThe classification categories are unique to the business and no available model achieves acceptable accuracy with the organization's data
    4. DThe team wants to add a new project to its roadmap so that the work appears on the planning board for next year's review
    Show answer and explanation
    Correct answer: C

    A custom model is justified when the domain is unique and managed options cannot reach acceptable accuracy with the organization's data. Roadmap, discounts and one-off use do not justify it.

  7. Question 27

    An organization wants to create a shared library of prompts used by teams building agents. Which guideline is most important for keeping the library useful over time?

    1. AVersioned prompts with owners, tested examples, and documented expected outputs
    2. BPrompts written once at project start and left unreviewed as models and business rules change
    3. CPrompts copied from public forums without testing, then pasted into live agents by developers
    4. DPrompts stored in individual personal notes on local laptops and unshared notebooks for staff
    Show answer and explanation
    Correct answer: A

    Versioned, owned and tested prompts keep the library reliable and auditable. Personal notes, unreviewed prompts and untested copies lead to inconsistent behaviour.

  8. Question 28

    A business wants a small language model that runs close to its data for a narrow classification task with strict latency needs. Which evaluation is most important when developing the use case?

    1. AWhether the model is the most popular on social media across developer forums, streaming platforms and professional networks
    2. BWhether the model has the most parameters in its base architecture, as reported in vendor benchmark sheets that remain unaudited
    3. CWhether the model name matches the company's brand in the public marketing materials and internal naming conventions
    4. DWhether a smaller tuned model meets the accuracy and latency targets compared with a larger general model
    Show answer and explanation
    Correct answer: D

    For narrow tasks, a small model is justified when it meets accuracy and latency targets. Parameter count, popularity and naming do not decide suitability.

  9. Question 29

    A company is forming a centre of excellence for AI. Which element should it include to govern agent adoption across business units?

    1. ARemoving documentation requirements for agent designs, prompts, data sources and tool connections
    2. BA ban on AI experimentation in production and pilot environments until a formal risk assessment
    3. CA standard intake process, shared guidelines, and a review board for agent approvals
    4. DEach business unit building agents with no shared standards and publishing them in separate team folders
    Show answer and explanation
    Correct answer: C

    A centre of excellence should provide intake, standards and approval processes so adoption is consistent and governed. Unstructured or banned adoption fails both goals.

  10. Question 30

    A sales organization uses Dynamics 365 Sales together with Copilot features in other Dynamics 365 apps. The architect must design a solution spanning these apps. What should the design consider first?

    1. AThe number of users in the marketing department and their licence tier assignments for Dynamics 365 Sales
    2. BThe timing of the company's holiday schedule for finance and HR teams across Europe, North America and Asia
    3. CData flows and shared definitions across the apps so that agents see consistent customer records
    4. DThe colours of each app's home screen and the font sizes and icon sets on desktop devices
    Show answer and explanation
    Correct answer: C

    Consistent shared data definitions and flows across Dynamics 365 apps are the foundation of a multi-app AI design. The other items do not affect the architecture.

Want more AB-100 questions?

The full bank has 100 questions with answers and explanations, as a PDF, an online practice exam, or both.

Full question bank · 100 questions
$89$124

Both formats together. The best value.

Instant access after payment, with 90 days of updates. Checkout is on our store, AIExamCerts.