Free Microsoft Certified: Azure AI Apps and Agents Developer Associate (AI-103) Practice Questions

Sample AI-103 Practice Questions

Browse all 10 free Microsoft Certified: Azure AI Apps and Agents Developer Associate practice questions below.

1. Developers added OpenTelemetry instrumentation to a hosted agent and export spans with the Azure Monitor exporter, but no traces appear on the Foundry project's tracing page. The agent responds normally, and the developers already hold the Log Analytics Reader role. What is the most likely missing configuration?

   • The agent has not been published, and tracing works only for published agents.
   • Content recording is turned off, and that setting suppresses every span the agent emits.
   • The Foundry project has no Application Insights resource connected to it.
   • The developers lack Foundry Agent Consumer, which is required before an agent can emit spans.

2. An airline's rebooking agent uses a language model to understand passenger requests. Fare rules decide who may rebook for free, and refunds above 500 USD need a supervisor. Regulators require these decisions to be reproducible and auditable. Which TWO responsibilities should a deterministic rules engine own instead of the model? (Select TWO)

   • Interpreting a passenger's free-text description of the disruption
   • Writing the confirmation message in the passenger's preferred language
   • Applying the 500 USD threshold that routes refunds to a supervisor
   • Summarizing the conversation for the colleague who takes over a handoff
   • Deciding whether a passenger's fare qualifies for a free rebooking

3. A bank's platform team must stop every project team from creating model deployments that could process prompts outside the European Union, while still allowing Data Zone and Standard deployments in EU regions. Teams create deployments through the Foundry portal, the Azure CLI, and Bicep templates. Which control enforces this across all three paths?

   • An Azure Policy rule that denies deployments with a Global SKU name such as GlobalStandard
   • A private endpoint on each Foundry resource so that inference traffic stays inside the bank network
   • A custom RBAC role that removes the deployments write permission from every developer in the tenant
   • A Foundry guardrail assigned to every model deployment that blocks requests sent from outside the EU

4. A field service company transcribes technician calls recorded in loud plant environments. The calls are full of equipment part numbers, and baseline transcription accuracy is poor. The team will build a custom speech model. Which TWO steps belong in the plan? (Select TWO)

   • Compare the word error rate of the base and custom models on the same test data.
   • Rely on a phrase list alone, since phrase lists also adapt recognition to plant noise.
   • Use custom voice training data so that the recognizer learns the technicians' accents.
   • Train with plant audio and human transcripts, plus text data containing part numbers.
   • Deploy the custom model as a Voice Live avatar so it can transcribe the calls.

5. A research agent in Foundry Agent Service calls web search and an OpenAPI tool that returns supplier documents. During red teaming, a supplier PDF contained hidden text telling the model to forward the full conversation to an external address, and the agent tried to comply. The agent's guardrail scans only the user input and output intervention points. Which change addresses this attack path?

   • Add an indirect attack control at the tool response intervention point of the agent's guardrail.
   • Add a protected material for text control at the output intervention point of the model deployment.
   • Enable Spotlighting on the agent's guardrail so documents are base64 encoded before the model reads them.
   • Set the user prompt attack control on the user input intervention point to its strictest threshold.

6. A healthcare company builds prompt agents in Foundry Agent Service. Its policy requires that agent conversation history, files uploaded by users and the vector stores that agents create are stored at rest in Azure resources that the company owns, so that it controls encryption, retention and network access. Which setup meets this requirement?

   • A hosted agent whose own code writes each conversation turn to a company storage account
   • Standard agent setup with the company's own Storage, Azure AI Search and Cosmos DB resources
   • Foundry IQ knowledge bases that index the conversation history in the company's search service
   • Basic agent setup, with customer-managed keys turned on for the Foundry resource

7. A support agent retrieves from an agentic retrieval knowledge base in Azure AI Search. Most user queries are short lookups, such as an error code or a product name, and the latency budget leaves no room for LLM query planning on each request. Which retrieval reasoning effort should the team set?

   • medium
   • low, which is the default
   • high
   • minimal

8. A bank deploys Foundry Agent Service with private networking and network injection so that agent traffic to its tools stays inside the bank's virtual network. The network team must prepare the subnet that the agent runtime is injected into before the deployment starts. Which subnet configuration is required?

   • A subnet delegated to Microsoft.CognitiveServices/accounts with an address range of /29
   • A subnet delegated to Microsoft.App/environments with an address range of /27 or larger
   • The GatewaySubnet of the virtual network, which already routes traffic to the bank's offices
   • The subnet that hosts the private endpoints for the Storage, Azure AI Search and Cosmos DB resources

9. An insurer builds a Content Understanding analyzer for photos of damaged cars. A field for the license plate number uses the extract method, and creating the analyzer fails. The team still needs the plate number and a damage severity label from a fixed list for each photo. How should the fields be defined?

   • Use extract for both fields and set enableSegment to true.
   • Keep extract for the plate number and set the workflow to agentic.
   • Run the OCR skill and map its text output to the plate field.
   • Use generate for the plate number and classify for the severity label.

10. A pharmaceutical company has 60,000 human-approved English to German sentence pairs from past regulatory submissions. It wants translations that follow this terminology and style, from a dedicated model that it trains once and reuses for months. Which approach fits best?

    • Build an adaptive custom translation index from the sentence pairs.
    • Prompt a general LLM with five approved pairs as examples.
    • Train a Custom Translator model on the parallel sentence pairs.
    • Use synchronous document translation with a glossary of key terms.