AB-100: Agentic AI Business Solutions Architect Practice Exam

What the AB-100 exam covers

• Plan AI-powered business solutions209 questions
• Design AI-powered business solutions209 questions
• Deploy AI-powered business solutions341 questions

Free AB-100 sample questions

A sample of 10 questions with answers and explanations. Sign up free to practice all 759.

1. Question 1Plan AI-powered business solutions

   A finance team asks for an agent to approve every purchase requisition automatically. Which question settles whether an agent is the right answer?

   • AWhether the approval rule is already deterministic and written downCorrect
   • BWhether the team has licences for Microsoft Copilot Studio
   • CWhether the requisition volume exceeds a hundred per day
   • DWhether the finance team has used an agent before
   ✓ Correct answer: A

   An approval governed by thresholds and a written policy is a deterministic decision, and automating it with a Power Automate flow is cheaper, faster and auditable in a way a generative agent is not. Licence availability, daily volume and the team's prior experience are three considerations that affect how a solution is delivered rather than whether an agent is the right shape for it.

   Why the other options are wrong

   • BWhether the team has licences decides how to build rather than what to build.
   • CVolume affects the return on automating rather than the shape of it.
   • DWhether the team has used an agent before affects adoption rather than fit.
2. Question 2Plan AI-powered business solutionsSelect all that apply

   Which two are legitimate reasons to reject an otherwise valuable agent scenario during planning?

   • AThe content the agent would need cannot lawfully be used this wayCorrect
   • BNo one in the business will accept accountability for its answersCorrect
   • CThe scenario would need a connector the team has not built before
   • DThe scenario would need its own environment and deployment pipeline
   ✓ Correct answer: A, B

   Both of these are absolute: content that may not lawfully be used this way cannot be used however valuable the outcome, and an answer nobody will stand behind has no route to being corrected when it is wrong. An unfamiliar connector and a dedicated environment are both work, and work is a cost to be planned rather than a reason to reject a scenario that is worth doing.

   Why the other options are wrong

   • CA connector the team has not built before is work to plan rather than a blocker.
   • DNeeding its own environment and pipeline is a cost, not grounds for rejection.
3. Question 3Plan AI-powered business solutions

   A plan proposes measuring adoption by weekly active users. What should it pair that with?

   • AWhether those users came back after their first unsuccessful answerCorrect
   • BWhether those users are distributed evenly across the departments
   • CWhether those users are using the agent during working hours
   • DWhether those users opened the agent from Teams or from the web
   ✓ Correct answer: A

   Anybody will try a new tool once, so the number that matters is whether they came back after it let them down - that is the difference between curiosity and reliance, and it is the earliest signal that the thing has a future. Departmental spread, time of day and entry channel describe who and how rather than whether the tool earned another chance.

   Why the other options are wrong

   • BEven distribution across departments describes who is trying it, not whether it held.
   • CUse during working hours describes when rather than whether it was trusted.
   • DWhether they opened it from Teams or the web is an entry channel detail.
4. Question 4Design AI-powered business solutions

   What does a design gain by separating the retrieval step from the generation step behind an interface?

   • AEither can be replaced or tested without disturbing the otherCorrect
   • BThe two steps can be billed to different cost centres
   • CThe generation step will consume fewer tokens per request
   • DThe retrieval step can be executed on cheaper infrastructure
   ✓ Correct answer: A

   Retrieval quality and generation quality move independently and are improved by different work, so a seam between them lets you evaluate the retriever against known-good passages and swap the model without re-testing the search. Cost allocation, token usage and cheaper infrastructure are incidental benefits of the same boundary.

   Why the other options are wrong

   • BBilling the two steps to different cost centres is an accounting benefit.
   • CToken usage in generation is unchanged by where the boundary sits.
   • DRunning retrieval on cheaper infrastructure is an incidental saving.
5. Question 5Design AI-powered business solutions

   A design needs an agent to work when its primary model is unavailable. What is the soundest fallback?

   • AA second model, with the fallback path tested as often as the primaryCorrect
   • BA cached copy of the most recent answer to each common question
   • CA message telling users the agent is unavailable and to try later
   • DA rules-based responder that handles the most frequent questions
   ✓ Correct answer: A

   A second model keeps the agent answering, and testing that path as routinely as the primary is what stops it failing the first time it is needed - untested fallbacks fail precisely when everything else already has. Cached answers go stale and are wrong for anything that moves, an unavailability message is honest but is not a fallback, and a rules responder is a second system to build and maintain.

   Why the other options are wrong

   • BCached answers to common questions go stale and are wrong for anything that moves.
   • CA message saying the agent is unavailable is honest but is not a fallback.
   • DA rules-based responder is a second system to build, test and maintain.
6. Question 6Design AI-powered business solutions

   A design must let an agent act on behalf of a user in Microsoft 365. What is the correct pattern?

   • ADelegated permissions acquired through the user's own consentCorrect
   • BApplication permissions granted to the agent across the tenant
   • CA privileged account the agent signs in as for every request
   • DThe agent owner's own credentials, stored and reused per request
   ✓ Correct answer: A

   Delegated permissions confine the agent to what that specific person may already do, carry their consent and expire - which is exactly the semantics of acting on somebody's behalf. Application permissions grant tenant-wide access that no individual user has, a privileged account abandons per-user authorisation, and reusing the owner's credentials makes every action look like theirs.

   Why the other options are wrong

   • BApplication permissions grant tenant-wide access no individual user holds.
   • CA privileged account for every request abandons per-user authorisation.
   • DReusing the owner's credentials makes every action appear to be theirs.
7. Question 7Deploy AI-powered business solutionsSelect all that apply

   Which two are worth telling users explicitly when an agent launches?

   • AThe kinds of question it is expected to answer wellCorrect
   • BWhat happens to a question it cannot answerCorrect
   • CWhich model the agent uses to generate its answers
   • DHow many documents are in the agent's knowledge base
   ✓ Correct answer: A, B

   Setting the subject expectation raises the proportion of questions that land in scope, and describing the fallback removes the fear of a dead end - together they produce a user who asks confidently and is not stranded. The model behind it and the document count are facts about the build that change nobody's behaviour.

   Why the other options are wrong

   • CWhich model generates the answers is a build fact users do not act on.
   • DThe number of documents in the knowledge base changes nobody's behaviour.
8. Question 8Deploy AI-powered business solutions

   Which property of an action most strongly argues for a human approval step?

   • AThat its effects would be difficult or costly to reverseCorrect
   • BThat it is performed more often than the other actions
   • CThat it involves data from more than one source system
   • DThat it takes the agent longer to complete than others
   ✓ Correct answer: A

   Approval buys the chance to catch a mistake before it becomes permanent, so its value scales with how hard the action is to reverse - a payment issued or a message sent to a customer cannot be recalled, while a draft can simply be discarded. Frequency argues against a gate rather than for one, since gating a common action is where approval fatigue starts. Multiple source systems describe complexity rather than consequence, and processing time is a performance property with no bearing on risk.

   Why the other options are wrong

   • BThat an action is performed more often argues against a gate, since gating common actions is where fatigue begins.
   • CThat it involves data from more than one source system describes complexity rather than consequence.
   • DThat it takes the agent longer to complete is a performance property with no bearing on the risk it carries.
9. Question 9Deploy AI-powered business solutions

   Slide decks are being indexed and answering badly. What is the most likely reason?

   • ASlides carry meaning in layout that extraction discardsCorrect
   • BSlides are usually more out of date than written documents
   • CSlides contain images that cannot be indexed as text
   • DSlides are written for a presenter rather than a reader
   ✓ Correct answer: A

   Slides encode relationships spatially - columns that contrast, arrows that sequence, a heading that governs the boxes beneath it - and text extraction produces a flat list of fragments in which those relationships are gone, leaving passages that match poorly and read as disconnected phrases. Currency varies by organisation rather than by format. Unindexed images are a real but partial loss, and being written for a presenter is true and contributes without being the mechanism.

   Why the other options are wrong

   • BSlides being more out of date than written documents varies by organisation rather than following from the format.
   • CSlides containing images that cannot be indexed as text is a real but partial loss next to the structural one.
   • DSlides being written for a presenter rather than a reader contributes without being the actual mechanism.
10. Question 10Deploy AI-powered business solutions

    What should be tested that only appears when memory is enabled?

    • AWhether something remembered becomes wrong laterCorrect
    • BWhether the agent answers the first question well
    • CWhether memory is stored in the right region
    • DWhether the memory can be exported on request
    ✓ Correct answer: A

    Memory introduces a failure mode that a single-turn test cannot produce: a fact captured correctly at one moment and repeated confidently after it has stopped being true, with no error and nothing to alert on. Testing it requires a scenario with elapsed time in it. First-question quality is exactly what stateless testing already covers. Storage region and export capability are configuration and compliance checks that hold whether or not anything has been remembered yet.

    Why the other options are wrong

    • BWhether the agent answers the first question well is precisely what stateless testing already covers.
    • CWhether memory is stored in the right region is a configuration check that holds before anything is remembered.
    • DWhether the memory can be exported on request is a compliance capability rather than a behaviour over time.

Who this AB-100 practice exam is for

This practice set is for anyone preparing for the AB-100: Agentic AI Business Solutions Architect exam - from first-time candidates building a foundation to experienced Microsoft practitioners doing a final review before test day. If you learn best by working through realistic questions and reading why each answer is right or wrong, it is built for you.

How to use this AB-100 practice exam

1. Start with the free sample questions above to gauge your current baseline.
2. Read the full explanation on every question, including why each wrong option is wrong.
3. Track your weak domains and focus your study where you are losing the most marks.
4. Once you are scoring consistently well, take a timed, full-length mock exam.
5. Use your readiness score to decide when you are ready to book the real AB-100 exam.

Related Microsoft resources

• Microsoft practice examsAll Microsoft
• Certification pathWhere this fits
• Certification exam guides & tipsBlog
• Plans & pricingFree & paid
• SC-200 practice examRelated
• AZ-140 practice examRelated
• AZ-900 practice examRelated

AB-100 practice exam FAQ