Free AB-100 practice test: 20 questions
Twenty questions across the three AB-100 skill areas, weighted roughly as the real exam is: six on Plan, five on Design and nine on Deploy. These are original scenario questions written against the official skills outline. Find the constraint in each scenario before you read the options.
Plan AI-powered business solutions
Question 1. A logistics firm wants an agent to help dispatchers. Before any build, the architect must judge whether the underlying shipment data can support grounding. Which set of properties should the review check?
- A. Colour, font and layout of the source documents
- B. Accuracy, relevance, timeliness, cleanliness and availability
- C. Licence price and seat count
- D. Model size and token limits
Show answer
Answer: B
The outline names accuracy, relevance, timeliness, cleanliness and availability as the properties to review for grounding data. Colour, font and layout are presentation concerns. Licence price and seat count are cost factors. Model size and token limits concern the model, not the data.
Want more questions like this? Full AB-100 practice tests →
Question 2. A sales team needs help drafting follow-up emails from CRM data. A prebuilt capability exists inside the tools they already use. The sponsor wants value in weeks, not months. What is the best first recommendation?
- A. Adopt the prebuilt capability and review gaps afterwards
- B. Build a custom agent on Microsoft Foundry
- C. Train a custom model on past emails
- D. Wait until a more complete product is released
Show answer
Answer: A
Using a prebuilt capability delivers fastest and cheapest when it meets the need. A custom agent or custom model adds time and cost without a stated gap. Delaying until a perfect product appears ignores the weeks-not-months constraint.
Want more questions like this? Full AB-100 practice tests →
Question 3. A manufacturer wants a high-volume classification of short maintenance tickets into 12 categories, at very low cost per request and low latency. A large general model is accurate but expensive. Which option should you consider?
- A. Use a still larger model for accuracy
- B. Expand the prompt library
- C. Reduce the categories to two
- D. Evaluate a customised small language model for the task
Show answer
Answer: D
A customised small language model can serve a narrow, high-volume task at lower cost and latency. A larger model raises cost. A longer prompt library does not reduce per-request cost. Removing categories changes the requirement instead of meeting it.
Want more questions like this? Full AB-100 practice tests →
Question 4. A company is starting its AI programme and wants shared standards for prompts, reusable assets, training and governance across business units. Which element should the architect propose?
- A. One developer who owns every agent
- B. Each team choosing tools independently with no standards
- C. An AI Center of Excellence
- D. A single kickoff workshop
Show answer
Answer: C
An AI Center of Excellence is the central body for standards, reusable assets, training and governance. A single developer owning every agent does not scale. A per-team tool choice without standards causes fragmentation. A one-off workshop does not provide ongoing governance.
Want more questions like this? Full AB-100 practice tests →
Question 5. A business case for an agent lists licence cost only. The CFO asks whether the benefit justifies the full investment. What should the ROI analysis add?
- A. A larger number of licences
- B. Total cost of ownership set against quantified benefits
- C. Marketing reach of the agent
- D. Developer preferences for tools
Show answer
Answer: B
Total cost of ownership includes licences plus build, integration, consumption, support and governance, set against quantified benefits. Adding only more licences extends the same gap. Marketing reach is not an ROI input for an internal agent. Developer preferences are not a financial criterion.
Want more questions like this? Full AB-100 practice tests →
Question 6. An agent's traffic is 85% simple lookups and 15% complex reasoning. Using one premium model for everything makes costs unsustainable. What should be implemented?
- A. A model router that selects a model per request
- B. A single smaller model for every request
- C. Disabling all complex requests
- D. More concurrent users
Show answer
Answer: A
A model router sends simple requests to a cheaper model and complex ones to a stronger model, cutting cost without losing capability where it matters. A single smaller model would hurt complex requests. Disabling reasoning requests removes needed function. More concurrent users does not reduce cost per request.
Want more questions like this? Full AB-100 practice tests →
Design AI-powered business solutions
Question 7. A procurement process must be monitored continuously, and an agent must act when a threshold is crossed, with no user starting it. Actions above a value limit still need approval. Which design fits?
- A. A prompt-and-response agent
- B. A task agent that users start manually
- C. An autonomous agent with an approval step above the value limit
- D. An autonomous agent with no approval step
Show answer
Answer: C
An autonomous agent triggered by events fits continuous monitoring, and an approval step for high-value actions keeps the required control. A prompt-and-response agent only reacts to questions. A user-triggered task agent does not meet the no-one-starts-it condition. Removing approval breaks the stated control.
Want more questions like this? Full AB-100 practice tests →
Question 8. A Copilot Studio agent receives requests it was never designed for and currently replies with irrelevant answers. What should you design?
- A. Hundreds of additional topics
- B. A computer-use step
- C. Silence for unmatched requests
- D. A fallback topic that rephrases or hands off
Show answer
Answer: D
A fallback topic defines the response when no topic matches, such as asking the user to rephrase or handing off. More topics cannot cover unknown requests. Computer use is irrelevant to unmatched conversation. Disabling replies leaves users with nothing.
Want more questions like this? Full AB-100 practice tests →
Question 9. A team needs an agent to reach several internal tools through a standard interface rather than building a custom integration for each. Which approach should you propose?
- A. Computer use against each tool
- B. Model Context Protocol
- C. A longer system prompt
- D. A new custom model
Show answer
Answer: B
Model Context Protocol provides a standard way to expose tools to an agent. Computer use is for applications without APIs and is less robust. A larger prompt cannot connect tools. A new custom model has no bearing on tool access.
Want more questions like this? Full AB-100 practice tests →
Question 10. A finance director wants employees to ask questions about finance and operations processes and receive answers that also draw on a company policy repository. Which design aim fits?
- A. Add the policy repository as a knowledge source for the agent chat
- B. Build an unrelated standalone chatbot
- C. Train a custom model on the policies
- D. Print and distribute the policies
Show answer
Answer: A
Finance and operations agent chats can be designed to use additional knowledge sources, so adding the policy repository as a knowledge source meets the need. Building a separate unrelated chatbot duplicates function. A custom model is not needed to read documents. Printing the policies is not a solution.
Want more questions like this? Full AB-100 practice tests →
Question 11. An existing scenario already uses Dynamics 365 customer service. The business wants customers to be able to chat with an agent in the contact centre channels. Which area do you design?
- A. A Power BI dashboard only
- B. A paper form for customers
- C. An agent integrated with Dynamics 365 Contact Center channels
- D. A finance-only agent
Show answer
Answer: C
The outline includes designing agents for integration with Dynamics 365 Contact Center channels. Power BI dashboards report data but do not host conversations. Manual forms lack the chat channel. A finance-only agent does not apply to a customer service channel.
Want more questions like this? Full AB-100 practice tests →
Deploy AI-powered business solutions
Question 12. An agent is live. Management asks how to know whether it is working. Which monitoring plan is most complete?
- A. Count the number of logins
- B. Monitor only monthly cost
- C. Wait for complaints
- D. Track usage, answer quality, latency, errors and feedback, and review the backlog
Show answer
Answer: D
A good plan covers usage, quality, reliability and feedback, and acts on a backlog. Counting logins alone ignores quality. Monitoring only cost ignores outcomes. Waiting for complaints is reactive and misses silent failures.
Want more questions like this? Full AB-100 practice tests →
Question 13. You prepare to test an agent that answers HR policy questions. Outputs vary slightly between runs. Which approach gives reliable regression testing?
- A. Chat informally and judge by impression
- B. A fixed test set with expected behaviours and pass criteria, re-run after every change
- C. Test once at launch
- D. Test only the greeting message
Show answer
Answer: B
A fixed set of inputs with expected behaviours and pass criteria, re-run after each change, gives repeatable evidence despite output variation. Ad hoc chatting is not repeatable. Testing once at launch misses later regressions. Testing only the greeting ignores the main function.
Want more questions like this? Full AB-100 practice tests →
Question 14. A custom model is being prepared for release. What should be defined before approving it for production?
- A. Validation criteria with measurable thresholds
- B. A new logo for the model
- C. Release-day weather
- D. The number of meetings held
Show answer
Answer: A
Validation criteria set the measurable thresholds the model must meet. A logo does not validate behaviour. Release-day weather is irrelevant. The number of meetings held says nothing about model quality.
Want more questions like this? Full AB-100 practice tests →
Question 15. A process spans Dynamics 365 Sales and Dynamics 365 Customer Service with an agent handing work between them. Individual app tests pass, but live handoffs fail. What should the test strategy add?
- A. More tests of each app in isolation
- B. Removal of the handoff
- C. End-to-end test scenarios that cross both apps
- D. Testing only in production
Show answer
Answer: C
End-to-end scenarios across the apps test the handoff where failures occur. More tests of each app alone repeat what already passes. Removing the handoff changes the process. Testing only in production exposes customers to the failures.
Want more questions like this? Full AB-100 practice tests →
Question 16. A Copilot Studio agent uses connectors and actions that point to different systems in development and production. Which ALM practice avoids breaking the production agent when promoting?
- A. Hard-code development endpoints into the agent
- B. Export screenshots of the agent
- C. Edit production by hand after each release
- D. Keep environment-specific configuration separate from the solution
Show answer
Answer: D
Separating environment-specific configuration such as connections and endpoints from the solution lets the same package work in each environment. Hard-coding development endpoints breaks production. Exporting screenshots does not move the agent. Editing production by hand is uncontrolled.
Want more questions like this? Full AB-100 practice tests →
Question 17. An autonomous agent can read a customer's records and also send email. A security review warns about prompt manipulation from inbound email text. Which mitigation is best?
- A. Make the agent’s replies more polite
- B. Treat inbound text as untrusted, limit tool permissions and filter outputs
- C. Give the agent administrator permissions
- D. Ignore the warning
Show answer
Answer: B
Treat inbound content as untrusted data, limit the agent’s tool permissions, and filter outputs, so injected instructions cannot trigger harmful actions. Greater politeness in the prompt does not block injection. Giving the agent more permissions increases impact. Ignoring the warning leaves the risk.
Want more questions like this? Full AB-100 practice tests →
Question 18. A company must process data for an agent in a specific region and prove it. Which review is required?
- A. Validate data residency and data movement compliance
- B. Check the colour theme
- C. Count the number of prompts
- D. Ask users to self-certify
Show answer
Answer: A
Validating data residency and data movement compliance confirms where data is stored and processed and whether it crosses a boundary. Checking colour themes is irrelevant. Counting prompts does not show location. Asking users to self-certify is not evidence.
Want more questions like this? Full AB-100 practice tests →
Question 19. Auditors ask who changed a tuned model's training data last quarter and when. What should the design have provided?
- A. Faster responses
- B. A larger model
- C. Audit trails for changes to models and data
- D. A shared administrator account for all changes
Show answer
Answer: C
Audit trails for changes to models and data record who changed what and when. Faster responses are a performance matter. A larger model has no bearing on traceability. A shared account hides who made each change.
Want more questions like this? Full AB-100 practice tests →
Question 20. A pre-launch review of a hiring-assistant agent must check it against responsible AI principles. Which concern is most directly part of that review?
- A. Response time of the chat window
- B. Colour scheme of the interface
- C. Number of connectors installed
- D. Whether the agent treats candidate groups fairly
Show answer
Answer: D
Fairness, that the agent does not disadvantage groups of candidates, is a core responsible AI principle and the most direct concern for a hiring scenario. Response time, colour scheme and number of connectors are quality or design matters, not responsible AI principles.
Want more questions like this? Full AB-100 practice tests →
Scoring yourself
Count your score per section. Below 70% on a section, reread its guide: Plan, Design or Deploy. The Deploy block is nine questions because it carries 40–45% of the exam. Then review your plan using the study plan.