Free AB-620 practice test

Azure AI Foundry is the correct platform for this scenario. It is purpose-built for pro-developers who need full control over agentic orchestration, dynamic tool/function selection, custom retry logic, integration with Azure infrastructure (including Service Bus), and the ability to scale to enterprise workloads. It supports frameworks like Semantic Kernel and Prompt Flow for complex multi-step reasoning pipelines. Copilot Studio (B) is a low-to-pro-code platform optimized for conversational agents; while it can call Power Automate flows, it does not natively support Service Bus triggers, dynamic tool registries, or custom orchestration at the code level required here. M365 Copilot extensibility (D) is designed for augmenting the M365 Copilot user experience with grounding and plugins — it is not an autonomous background-process runtime. Copilot Studio's enterprise knowledge source connectors (A) (ServiceNow, SAP) are for querying knowledge at conversation time, not for Service Bus integration or dynamic multi-API orchestration chains.

Copilot Studio uses Retrieval-Augmented Generation (RAG) when knowledge sources such as SharePoint are configured. At query time the system searches the indexed knowledge (via semantic/vector search), retrieves the most relevant passages, and injects them into the language model's context window so it can generate a grounded answer. Fine-tuning (B) would permanently update model weights with training data — Copilot Studio does not fine-tune models per customer; it retrieves at runtime. Supervised classification (D) describes traditional intent-recognition NLU, not knowledge retrieval. Direct SQL queries (A) are not used; SharePoint content is indexed and searched via Microsoft Search / Dataverse search infrastructure.

Enabling generative AI orchestration (also called 'generative answers' or 'generative mode') allows the agent to automatically use configured knowledge sources to answer questions that do not match a specific topic, providing broad coverage without authoring every possible topic. Classic NLU mode (A) relies solely on authored topics and trigger phrases; unmatched queries would fall through to the fallback system topic without consulting knowledge sources. Routing all unmatched queries to a human (C) defeats the purpose of the knowledge base and provides no self-service value. A wildcard catch-all topic (D) would intercept all messages and require manual authoring of every response; it does not leverage knowledge sources dynamically.

ServiceNow (and other enterprise knowledge sources such as SAP) are external systems that require an authenticated connector configuration — typically via OAuth 2.0 or service account credentials — so Copilot Studio can securely query the external system at runtime. SharePoint knowledge sources use the agent's existing Microsoft Entra identity/Graph connection, so no separate auth setup is needed. Enabling generative AI orchestration (A) is required for both SharePoint and enterprise knowledge sources equally, so it is not a differential requirement. Uploading PDFs (C) would be a workaround, not an additional configuration step specific to the live ServiceNow connection. Creating topics per category (B) is not required; the enterprise knowledge source surfaces articles dynamically without per-category topic authoring.

Topic misrouting most commonly occurs when trigger phrases are too broad or overlap across topics. If 'New Employee Onboarding' contains a trigger phrase like 'I need help', it will incorrectly match queries intended for other topics. The fix is to audit trigger phrases on competing topics, make them more specific, and add clear distinguishing phrases to the 'Password Reset' topic. Enabling generative AI orchestration (A) affects how unanswered questions are handled but does not resolve trigger-phrase conflicts between authored topics. A system topic fallback redirect (C) would only affect completely unmatched utterances, not misrouted ones. Clearing global variables (D) addresses data persistence issues, not intent routing.

Prompt engineering best practices for agent instructions include clearly stating the agent's persona/role, what topics it should and should not address (scope and constraints), the desired tone, and any formatting expectations. This gives the model a reliable operational framework. Listing all possible user questions (A) is not how instructions work — that is what topics and knowledge sources are for; instructions set behavior, not enumerate queries. Vague instructions (B) lead to inconsistent, off-topic, or unsafe responses. Sample Q&A pairs (D) are better suited to few-shot examples in a fine-tuning or prompt template context, not as the primary vehicle for agent instructions in Copilot Studio.

Global variables in Copilot Studio are session-scoped and accessible across all topics within the same conversation session, as well as passable to external actions such as Power Automate flows. This makes them the right choice when a value must span multiple topics. Topic variables (A) are scoped only to the topic in which they are created; they are not accessible in other topics. System variables (C) are predefined by the platform (e.g., conversation ID, user name) and cannot be created by authors to store custom data. Power Platform environment variables (D) store static configuration values (like connection strings) that are set at deployment time, not dynamic runtime values collected during a conversation.

Option A is correct because Copilot Studio supports importing OpenAPI specification files to auto-generate custom connectors, and the security definition in the spec maps directly to supported authentication schemes including OAuth 2.0 Bearer tokens — making this the fastest, most maintainable path. Option B would work functionally but hardcoding a Bearer token in a flow is a security anti-pattern and is far less efficient than using a connector with proper auth configuration. Option C is incorrect because Copilot Studio does not have a native 'Azure API Management action' that proxies arbitrary REST calls without connector setup. Option D is incorrect because the knowledge sources section is designed for document/FAQ-style retrieval, not live transactional API calls that require authentication headers.

Option C is correct. Microsoft Fabric's AI Skill (also referred to as OneLake AI Skill) is specifically designed to enable natural language querying over structured data in a Fabric Lakehouse or Warehouse. Copilot Studio supports connecting to a Fabric AI Skill as a knowledge source, enabling semantic retrieval without requiring exact keyword matches. Option A is incorrect because exporting to a SharePoint list introduces data staleness, loses semantic indexing, and is an unnecessarily complex pipeline for what Fabric AI Skill handles natively. Option B is incorrect because a KQL queryset requires a predefined, hardcoded query and cannot dynamically interpret natural language to generate queries — it would not satisfy the 'natural language' requirement. Option D is incorrect because the scheduled Notebook approach introduces latency, requires Dataverse licensing, and does not provide real-time or on-demand natural language querying of the Lakehouse data.

Options D and E are correct. To connect a Copilot Studio agent to an MCP server: (1) you must supply the MCP server's endpoint/manifest URL so that Copilot Studio can discover the available tools — this is the primary identifier for the server connection; and (2) you must configure the appropriate authentication for that server, which for API key auth means selecting 'API Key' and specifying the header name and key value in the connection settings. Option B is incorrect because MCP server connections do not require Azure AD app registration or a special 'MCP.ReadWrite' role; this conflates Azure resource permissions with MCP authentication. Option A is incorrect because Copilot Studio does natively support MCP server connections as a first-class integration; there is no need to export and re-import as an OpenAPI custom connector. Option C is incorrect because there is no 'Allow agent to call unauthenticated MCP servers' tenant toggle in the Power Platform admin center; this is a fabricated setting. Option F is a plausible future pattern, but currently Copilot Studio's MCP connection settings accept the key value directly in the connection configuration rather than by referencing an Azure Key Vault secret name.

Option B is correct and reflects the proper layered architecture: (1) a Generative Answers node retrieves relevant SharePoint documents as grounding context; (2) an action node invokes the Azure AI Foundry-connected model with both conversation history and the retrieved context as explicit inputs; (3) a subsequent Power Automate flow action post-processes the Foundry model's output; and (4) the Send Message node presents the final result. This uses each component for its designed purpose. Option C is incorrect because Azure AI Foundry is not configured as a 'knowledge source' in Copilot Studio — knowledge sources are data stores (SharePoint, websites, files), not AI models. Foundry models are connected as AI model capabilities or called via actions. Additionally, Generative Answers does not automatically route through Power Automate. Option A is incorrect because swapping the tenant-level default AI model in Power Platform admin center applies globally and does not support per-agent fine-tuned model targeting via Foundry. Azure OpenAI deployments used this way do not automatically inherit Foundry-specific features. Option D is incorrect because while technically achievable, routing all AI inference through a single Power Automate flow bypasses Copilot Studio's native orchestration capabilities, conversation context management, and topic logic — it is an anti-pattern that creates an unmaintainable, brittle solution and does not leverage the platform's strengths.

Option A is correct. Configuring the custom connector with OAuth 2.0 delegated permissions through Azure AD means that when an employee invokes the connector, the access token represents that specific user's identity — satisfying the IT security mandate against shared service accounts. SAP S/4HANA supports Azure AD-based OAuth for OData services through SAP's BTP integration. Option C is incorrect because 'No authentication' sends unauthenticated requests; relying on SAP-side security alone while the connector passes no identity defeats the delegated identity requirement and creates an unauthenticated surface. Option D is incorrect because using a hardcoded service account in an environment variable is exactly what the security team prohibited — a shared credential — regardless of the audit logging added as a compensating control. Option B is incorrect because importing an OData metadata document as a knowledge source only gives the agent schema information (field names, entity definitions), not live purchase order data; it cannot execute queries against the live SAP system.

Option B is correct. The Agent-to-Agent (A2A) protocol is a Microsoft-supported open standard that enables agents to publish an 'agent card' describing their capabilities, and allows orchestrator agents — including those built in Azure AI Foundry — to discover and invoke them dynamically without hardcoded endpoints. Option A is incorrect because MCP (Model Context Protocol) is designed for exposing tools and data sources to a single agent, not for agent-to-agent orchestration and discovery across platforms. Option C is incorrect because a shared Power Automate flow is a workaround that creates tight coupling and does not provide the capability discovery, task routing, or protocol-level interoperability that A2A offers. Option D is incorrect because a Service Bus pattern would require custom development for message schema, error handling, and routing, and is not a standard integration pattern for agent-to-agent communication in Copilot Studio or Azure AI Foundry.

Option C is correct. Copilot Studio allows you to inspect the output of a Power Automate flow action (including status codes or error flags returned by the flow) using a Condition node. You can route to a separate topic or message node that presents a user-friendly error message when an error condition is detected. This is the standard, supported error-handling pattern within the agent canvas. Option B is incorrect because Copilot Studio's action nodes do not have a built-in 'Retry on failure' toggle with a custom message capability — this is a fabricated option that reflects a common misconception. Option D is incorrect because Power Automate flows do not have a mechanism to publish events back into a Copilot Studio agent topic mid-conversation; the flow can return structured output, but 'publishing an error event' is not a supported integration pattern. Option A is incorrect because introducing Azure API Management to mask 500 errors as 200 responses hides the failure from all consumers of the API, creates misleading telemetry, and is an architectural anti-pattern for error handling.

Correct answers (A and C): Adaptive Cards are the native interactive card format supported by both Microsoft Teams and Copilot Studio. You define the card JSON (including Action.Submit for buttons and input elements for dropdowns/date pickers) and deliver it using the 'Send a message' or 'Ask with Adaptive Card' authoring node. The 'Ask with Adaptive Card' node specifically captures the submission payload and maps the values to topic variables, allowing the conversation to branch based on the user's choice — this is the standard, supported pattern.

Why B is wrong: Having a Power Automate flow push the card and then polling for a response introduces unnecessary complexity and latency, breaks the conversational flow, and is not the recommended pattern when the agent itself can natively deliver and receive Adaptive Cards.

Why D is wrong: Hosting the form as an HTML page on Azure Static Web Apps and using a webhook is a valid general web-integration pattern but is a heavy over-engineering of the scenario. It bypasses native Copilot Studio card support and introduces significant additional infrastructure and maintenance overhead.

Why E is wrong: Structured Output mode controls how an AI model returns data in a machine-readable format (e.g., JSON schema enforcement); it is not a mechanism for rendering interactive UI forms to end users.

Why F is wrong: While Teams is a primary channel for Adaptive Cards, Copilot Studio supports Adaptive Cards in multiple channels (including the web chat channel). Restricting deployment solely to Teams is not a requirement.

Exporting as a managed solution is the standard ALM pattern for Power Platform and Copilot Studio. A managed solution locks customizations in the target environment (preventing unmanaged edits), and environment variables allow connection strings or endpoint values to be swapped without modifying the solution itself—this is exactly what is needed for a clean dev-to-production promotion. Manually re-creating the agent is error-prone, time-consuming, and not a scalable practice. The Test chat pane export exports conversation transcripts for analysis, not solution artifacts—it cannot migrate an agent. The 'Save as' option creates a copy within the same environment; it does not support cross-environment deployment and does not manage environment variables.

In Copilot Studio's role model, the 'agent contributor' role allows users to collaborate on building and editing agent content (topics, entities, etc.) but does not grant the right to publish an agent to channels. Publishing is reserved for makers and administrators. A separate Azure Bot Service subscription is not required for a custom website channel in Copilot Studio—the platform handles channel connectivity. There is no platform requirement that the Test chat pane must be used before publishing is allowed; it is a recommended practice, not an enforced gate. Exporting as a solution is an ALM practice and has no bearing on whether a channel can be published—these are independent operations.

Resolution rate measures the percentage of sessions in which the agent fully resolved the user's issue without escalation—a direct indicator of self-service effectiveness. Escalation rate measures the inverse: the percentage of sessions where the conversation was handed off to a human agent. Together they give a complete picture of how well the agent handles requests autonomously. Engagement rate measures how many sessions had at least one user-agent exchange beyond the opening greeting, which reflects reach but not resolution quality. CSAT score reflects user satisfaction ratings submitted after a session; while useful for sentiment, it does not directly measure whether the agent resolved the issue. Message capacity consumed is a billing and capacity metric, not a resolution-effectiveness metric.

Managed Environments includes the ability to block unmanaged customizations from being imported or saved in a production environment, which directly closes the loophole of makers deploying directly. Combining this with a formal pipeline (e.g., Power Platform Pipelines or Azure DevOps-based ALM) that includes an admin approval gate ensures all production changes go through a controlled process while sandbox environments remain open for development. Deleting maker licenses in production would prevent legitimate administrative deployments and is an extreme measure that does not solve the root cause. Blocking all connectors via DLP does not prevent unmanaged solution imports—it only controls which connectors can run—and would break legitimate production agent functionality. Using the CoE Starter Kit dashboard to manually delete agents weekly is a reactive, manual process that does not prevent unauthorized deployments from occurring in the first place.

DLP policies in Power Platform allow administrators to classify connectors into tiers (Business, Non-Business, or Blocked). Setting the Salesforce connector to Blocked in the targeted environment prevents any flow or agent in that environment from using it, which is exactly the right control for preventing PII exposure. Removing a connector from the tenant catalog is not a standard administrative action and would affect all environments globally rather than a targeted one. The 'agent contributor' role controls who can edit an agent, not which connectors are permitted—it has no DLP functionality. Managed environment sharing limits restrict how widely an agent can be shared, not which connectors makers can use.