AB-620: AI Agent Builder Associate Practice Exam

What the AB-620 exam covers

• Plan and configure agent solutions258 questions
• Integrate and extend agents in Copilot Studio344 questions
• Test and manage agents180 questions

Free AB-620 practice test questions

A sample of 10 questions with answers and explanations. Sign up free to practice all 782.

1. Question 1Integrate and extend agents in Copilot Studio

   What is a knowledge source in Copilot Studio?

   • AA scheduled Power Automate flow
   • BContent the agent searches for groundingCorrect
   • CA topic with exact trigger phrases
   • DAn environment variable holding a value
   ✓ Correct answer: B

   A knowledge source is content, such as SharePoint, a website, an uploaded file, or Dataverse, that the agent's generative answers feature searches and cites to ground its responses. The concept tested is what counts as knowledge versus an agent's other building blocks, automations, dialog logic, and stored values. A scheduled flow automates a process rather than holding searchable text, a topic reacts to trigger phrases with a scripted dialog, and an environment variable stores one configuration value. Remember that if content is searched and cited for an open-ended answer, it is knowledge; if it runs an action or stores a value, it is not.

   Why the other options are wrong

   • AA scheduled Power Automate flow automates a business process on a trigger or timer; it holds no searchable text and is never cited in a generative answer.
   • CA topic is a scripted dialog that fires on matched trigger phrases, not content the agent searches, so it does not qualify as a knowledge source.
   • DAn environment variable stores a single configuration value used by flows or topics, not a body of content the agent can retrieve and cite.
2. Question 2Integrate and extend agents in Copilot Studio

   When an agent retrieves from indexed SharePoint content, whose permissions are evaluated at answer time?

   • ANo identity is checked at all
   • BThe tenant's global admin, always
   • CThe end user asking the questionCorrect
   • DThe maker who authored the agent
   ✓ Correct answer: C

   Security trimming is evaluated against the identity of the person having the conversation at query time, not the maker who built the agent and not a fixed admin account, so each user only ever sees content they are personally authorized to view. The key concept is that this per-requester evaluation is what makes trimming meaningful; evaluating against a fixed high-privilege identity instead would defeat its purpose entirely. Grounding always reflects the asker's own access, not the builder's.

   Why the other options are wrong

   • AIdentity checking is precisely the mechanism by which security trimming enforces per-user access control, so claiming none occurs is factually wrong.
   • BA tenant's global admin identity is not substituted for every user's query by default; each user's own access is what gets checked.
   • DThe maker's permissions during authoring do not carry over to every end user's runtime access when the agent is actually used.
3. Question 3Integrate and extend agents in Copilot Studio

   A tool call fails every time with an authorization error, even though the tool is selected correctly by the model. What does this most likely indicate?

   • AThe tool's description text has quietly become far too long and confusing
   • BA problem with the tool's authentication or connection, not tool selectionCorrect
   • CThe model must have picked the entirely wrong tool for this request
   • DThe agent currently has far too many separate topics published live
   ✓ Correct answer: B

   The scenario states the model is choosing the right tool every time, which rules out selection as the problem, so a failure that happens consistently once the call is actually made points downstream, to expired, invalid, or misconfigured credentials or the underlying connection. This distinction matters because it tells the builder where to look next: at the authentication and connection configuration, not at the tool's description or how the orchestrator reasons about which tool to pick. Takeaway: separate selection problems from execution problems, an authorization error after correct selection is always a credentials or connection issue.

   Why the other options are wrong

   • AA long description affects whether the tool is chosen, not whether an already-selected call is authorized.
   • CThe scenario states the tool is selected correctly, so concluding the model chose the wrong tool contradicts the facts.
   • DThe number of published topics has no relationship to whether one specific tool call is authorized successfully.
4. Question 4Integrate and extend agents in Copilot Studio

   Which of the following is an accurate distinction between a Fabric data agent and a Foundry agent when both are used as connected agents in Copilot Studio?

   • AA Fabric data agent targets Fabric data; Foundry agents are more general-purposeCorrect
   • BA Fabric data agent can only ever be used inside a spreadsheet like Excel
   • CAn Azure AI Foundry agent can never be added as a connected agent at all
   • DThere is no real difference at all; both terms simply mean the same thing
   ✓ Correct answer: A

   A Fabric data agent is scoped specifically to answering natural-language questions grounded in Fabric or OneLake data such as a Lakehouse or Warehouse, while a Foundry agent is a more general-purpose, often custom-built agent capable of implementing arbitrary pro-code logic, tools, and orchestration well beyond structured-data queries; both can be added to Copilot Studio as connected agents, but they solve different problems. Knowing which one fits a given need, structured data versus general pro-code logic, is the key distinction to remember.

   Why the other options are wrong

   • BA Fabric data agent is a data-grounded conversational agent, not something limited to use inside a spreadsheet.
   • CAzure AI Foundry agents can be added as connected agents in Copilot Studio, so this claim is inaccurate.
   • DThe two are distinct, structured Fabric data querying versus general pro-code logic, not interchangeable terms.
5. Question 5Integrate and extend agents in Copilot Studio

   A development team wants to build a highly custom, code-first AI agent with bespoke orchestration logic that goes beyond what Copilot Studio's low-code tools support. Which platform is designed for that pro-code scenario?

   • AMicrosoft Foundry agent developmentCorrect
   • BThe Copilot Studio topic canvas only
   • CThe adaptive card designer
   • DThe environment variable list
   ✓ Correct answer: A

   Microsoft Foundry agent development is the platform designed for exactly this: a highly custom, code-first agent with bespoke orchestration logic that goes beyond what Copilot Studio's low-code topics and flows support. The key concept is the low-code/pro-code split across the platform - Copilot Studio's topic canvas is for conversational authoring with guided, visual tools, while Foundry is where a development team writes custom orchestration code directly. That split is precisely what this scenario is testing, since the team explicitly needs bespoke logic beyond low-code limits. Takeaway: when a team needs custom code-level orchestration, that's a Foundry pro-code project, not a Copilot Studio low-code build.

   Why the other options are wrong

   • BThe topic canvas is a low-code authoring surface, not a pro-code agent framework.
   • CThe card designer builds response UI, not agent orchestration logic.
   • DEnvironment variables store configuration values, unrelated to orchestration logic.
6. Question 6Plan and configure agent solutions

   A financial services agent must avoid ever giving specific investment advice, only general information. Which combination of controls best enforces this at the platform level, beyond topic authoring alone?

   • AIncreasing the number of published channels
   • BAdding more connectors for market data
   • CGroundedness to approved content plus guardrailsCorrect
   • DDisabling moderation for natural responses
   ✓ Correct answer: C

   Groundedness restricted to approved content plus explicit guardrails is the combination that enforces this at the platform level: restricting generative answers to only vetted, approved investment-related content keeps the model from fabricating specific advice, while guardrail topics intercept and redirect any out-of-scope request for personalized recommendations, together going well beyond what careful topic authoring alone can guarantee. Takeaway: groundedness plus guardrail topics is the platform-level control for keeping an agent within an approved scope.

   Why the other options are wrong

   • AAdding more channels affects reach, not what kind of content the agent is allowed to produce.
   • BMore market data connectors could increase, not reduce, the risk of specific advice being generated.
   • DDisabling moderation for a naturalness benefit works against the safety goal here.
7. Question 7Plan and configure agent solutions

   After fixing a bug in a failing agent flow step, what is the recommended way to confirm the fix before relying on live traffic?

   • ARun the flow and check run historyCorrect
   • BDelete the existing run history
   • CRename the flow's trigger name
   • DTurn off all flow monitoring
   ✓ Correct answer: A

   Triggering a manual test run of the flow and then opening its entry in run history to confirm the previously failing action now succeeds is the direct way to validate a fix before real conversations depend on it. Takeaway: always verify a fix with a real test run before trusting it in production.

   Why the other options are wrong

   • BDeleting the existing run history only removes evidence of past behavior; it proves nothing about whether the current fix actually works.
   • CRenaming the flow's trigger is a purely cosmetic change with no effect on whether the underlying bug was actually fixed.
   • DTurning off flow monitoring removes visibility into future runs, which is the opposite of what is needed right after applying a fix.
8. Question 8Plan and configure agent solutions

   A custom GPT prompt is configured to return a structured output with fields 'category' and 'priority'. How can the topic use these individually after the prompt runs?

   • AThey are combined into one unformatted string with no access
   • BEach output field can be mapped to its own variableCorrect
   • COnly the first field returned can ever be used
   • DThe topic must call the prompt twice for both fields
   ✓ Correct answer: B

   A structured prompt output keeps its declared fields, category and priority, distinct rather than merging them, so when the topic adds the prompt as an action, it maps each field separately to its own variable, letting later nodes reference category and priority independently. The key concept is that structured outputs are a schema of named values, not a single blob of text, and mapping preserves that structure all the way into the topic. This applies directly because the scenario needs both fields usable individually, which is only possible if each is captured on its own. Remember: structured outputs let every declared field be mapped and used on its own.

   Why the other options are wrong

   • AStructured outputs preserve individual fields rather than collapsing them into one blob.
   • CAll declared output fields, not just the first, can be mapped and used.
   • DA single prompt call can return multiple structured fields at once; a second call isn't needed.
9. Question 9Test and manage agents

   A test case's expected answer could be phrased many valid ways, such as 'the refund takes 5-7 days' versus 'refunds are processed in five to seven business days.' Which evaluation method should the builder choose?

   • AExact string match evaluation
   • BSemantic AI-assisted evaluationCorrect
   • CALM solution checker validation
   • DDeployment publish validation
   ✓ Correct answer: B

   When the expected answer could be worded many valid ways, such as '5-7 days' versus 'five to seven business days,' a meaning-based evaluation method is the right choice, since it judges whether the meaning matches the expected answer rather than requiring identical text. In Copilot Studio's agent evaluation, that method is named Compare meaning, which scores how well the meaning of a test case's answer matches the expected answer out of 100% against a pass score the builder sets. The key concept is that an exact-match method would fail every one of those valid rewordings, because it only passes when the answer matches the expected answer exactly, so a meaning-based evaluator is needed whenever multiple correct phrasings are expected. The remaining options check ALM packaging or deployment readiness, neither of which judges response wording quality. Takeaway: match the evaluation method to whether wording is fixed or flexible.

   Why the other options are wrong

   • AExact string match would flag both valid rewordings as failures, since it requires identical text rather than matching meaning.
   • CThe ALM solution checker validates packaging and component references, not the quality or correctness of a generated response.
   • DDeployment publish validation confirms an app is ready to ship, unrelated to judging whether a response's wording is correct.
10. Question 10Test and manage agents

    An import of a solution containing an agent fails with a missing dependency error for a custom connector. What is the most direct fix?

    • AChange orchestration mode
    • BIncrease solution version
    • CAdd the connector, then re-importCorrect
    • DConvert connector to knowledge source
    ✓ Correct answer: C

    The correct fix is to add the connector, then re-import. The key concept is that a missing-dependency error at import time means the target environment received a reference to a component, the custom connector, that was not actually packaged, so the resolution is to go back to the source solution, add that connector as a component, re-export, and re-import so the dependency is now satisfied. This is more direct than adjusting orchestration mode or bumping the version number, neither of which addresses the actual missing component, or trying to convert the connector into an unrelated component type. Takeaway: a missing-dependency import error is fixed by adding the missing component, not by changing unrelated settings.

    Why the other options are wrong

    • AOrchestration mode controls how the agent selects topics or tools at runtime; it has no effect on whether a connector exists in the solution package.
    • BIncreasing the solution version number identifies the package as an update but does not by itself add the missing connector component.
    • DA custom connector cannot be converted into a knowledge source; those are entirely different component types serving different purposes.

Who this AB-620 practice exam is for

This practice set is for anyone preparing for the AB-620: AI Agent Builder Associate exam at the intermediate level - 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-620 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-620 exam.

Related Microsoft resources

• AB-620 study guideKey concepts
• Microsoft practice examsAll Microsoft
• Certification pathWhere this fits
• Certification exam guides & tipsBlog
• Plans & pricingFree & paid
• How these questions are written and reviewedMethodology
• Report a problem with a questionCorrections
• AB-730 practice examRelated
• AB-731 practice examRelated
• AB-900 practice examRelated

AB-620 practice exam FAQ