Explore / Microsoft Azure AI Apps and Agents Developer (AI-103) Exam Prep / Practice questions
Microsoft Azure AI Apps and Agents Developer (AI-103) practice questions
10 questions from the bank behind this course's timed mock. Free to read and to answer, no account needed, and each has its answer and an explanation. They belong to Microsoft Azure AI Apps and Agents Developer (AI-103) Exam Prep.
Answer them one at a time →Read lesson 1 free
Question 1
A developer has 60 study hours before their AI-103 exam. They want to spend the hours in proportion to the skills measured as of April 16, 2026. Which plan best fits the published weightings?
- Spend most of the hours on information extraction, because document processing questions are the hardest to guess
- Spend about 35 hours on planning and managing an Azure AI solution plus implementing generative AI and agentic solutions, and share the rest across computer vision, text analysis and information extraction
- Spend about 35 hours on computer vision, text analysis and information extraction, because those are three areas against two
- Spend 12 hours on each of the five skill areas, since every area appears on the exam
Show the answer
Answer: B. Spend about 35 hours on planning and managing an Azure AI solution plus implementing generative AI and agentic solutions, and share the rest across computer vision, text analysis and information extraction
Plan and manage (25–30%) and generative AI and agents (30–35%) together make up 55–65% of the exam, so about 35 of 60 hours fits the weighting. An equal split looks fair, but it ignores that the three other areas are worth only 10–15% each.
Question 2
A platform team is building a Toolbox so that several agents, running on different MCP-capable runtimes, share one MCP-compatible endpoint with central authentication and versioning. The candidate tools are a remote MCP server, an OpenAPI tool for the HR REST API, an Azure AI Search index, and a function tool that reads a SQLite database inside one client application. Which tool cannot be included in the Toolbox?
- The Azure AI Search index
- The function tool that reads the client application's SQLite database
- The OpenAPI tool for the HR REST API
- The remote MCP server
Show the answer
Answer: B. The function tool that reads the client application's SQLite database
A Toolbox can't include function calling, because function tools run client-side in your own app and not behind a shared endpoint. The OpenAPI tool is the closest distractor, but Foundry calls that API itself from the spec, so it can be bundled in a Toolbox.
Question 3
You are an AI engineer supporting a chat app built on an Azure OpenAI deployment. Over the past week, Time to Last Byte has gone up from about 2 seconds to about 4 seconds. Before you open an incident, you want to know whether the slowdown simply comes from larger requests. Which metric should you chart next to Time to Last Byte, filtered to the same deployment?
- Processed Prompt Tokens (Sum)
- Azure OpenAI Requests split by StatusCode
- Blocked Volume (RAIRejectedRequests)
- The Latency metric under Cognitive Services - HTTP Requests
Show the answer
Answer: A. Processed Prompt Tokens (Sum)
If prompt tokens doubled along with latency (for example, someone added a long system prompt), the slowdown is expected. If tokens stayed flat, something is actually wrong. Requests split by StatusCode shows errors and throttling, but it can't tell you whether requests got bigger.
Question 4
You are building an approval step for a Foundry agent that uses an MCP tool. A reviewer approves a pending refund_order call. Your code sends an McpApprovalResponse with the correct approval_request_id and approve=True by calling responses.create. The agent never continues the original task. What is the most likely cause?
- refund_order was not in the allowed_tools list, so the approved call was silently dropped
- require_approval was left at its default of "always", which blocks every call even after a reviewer approves it
- Tracing was not turned on, and approval responses are only processed when Application Insights is connected
- The responses.create call did not include previous_response_id, so the decision was not linked to the paused response
Show the answer
Answer: D. The responses.create call did not include previous_response_id, so the decision was not linked to the paused response
You send the approval decision back with previous_response_id set to the paused response's ID. Leave it out and the run stalls. The default "always" only means each call needs a decision; it doesn't block calls that have been approved.
Question 5
You are an AI engineer evaluating a support RAG app. The latest run shows a low average retrieval score and a low groundedness score, while relevance is moderate. The failed rows' reasons say the retrieved chunks are about unrelated products. You can make one change before the next run. What should you do first?
- Switch the app to a stronger model, because low groundedness means the model is fabricating.
- Fix the search index: adjust chunking, enable hybrid or semantic search, and tune top-k.
- Tighten the system prompt so the model answers only from the sources or says it does not know, and lower the temperature.
- Replace GroundednessEvaluator with F1ScoreEvaluator so that answers are scored against reference answers.
Show the answer
Answer: B. Fix the search index: adjust chunking, enable hybrid or semantic search, and tune top-k.
When retrieval and groundedness are both low, search returned the wrong chunks, so the index is what needs fixing. Tightening the prompt is the fix when retrieval is good but groundedness is low. In this run the model never saw the right context.
Question 6
You are building a claims workflow in the Foundry portal. The intake agent node extracts policy_no, incident and amount. A later Logic (if/else) node must branch on the amount. Right now the next node receives only free text, and the branch condition can't read the amount reliably. What should you do?
- Replace the Logic node with a Group chat template so the agents can discuss the amount in a shared conversation
- Add a Basic chat node after the intake agent and tell it to repeat the amount as its first word
- Turn on the YAML Visualizer View toggle so that agent outputs are typed automatically
- Set the intake agent's text format to JSON Schema, save its output to a variable, and reference it in Power Fx with a scope prefix such as Local.ClaimData
Show the answer
Answer: D. Set the intake agent's text format to JSON Schema, save its output to a variable, and reference it in Power Fx with a scope prefix such as Local.ClaimData
To pass structured data between workflow nodes, you set the agent's text format to JSON Schema, save the output to a variable, and read it in Power Fx with a scope prefix like Local.ClaimData. The YAML Visualizer View is a real toggle, but it only shows the same workflow as YAML. It doesn't add types to agent output.
Question 7
A developer at Contoso has an app that calls a Foundry-hosted agent. Server-side tracing is on, and client-side tracing is set up with configure_azure_monitor() and AIProjectInstrumentor().instrument(). In Application Insights, the app's own 'handle-support-request' spans and the agent's server-side spans appear as two separate traces instead of one nested tree. The developer wants a single end-to-end trace. What should the developer check first?
- That the OpenAI client is created with project.get_openai_client() after instrument() has been called, so trace context is passed to Foundry
- That sampling is turned off in Application Insights, so that no parent spans are dropped
- That the developer has the Log Analytics Reader role on the Application Insights resource
- That OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT is set to true, so the server can link spans by message content
Show the answer
Answer: A. That the OpenAI client is created with project.get_openai_client() after instrument() has been called, so trace context is passed to Foundry
Client spans join server spans only when trace context is passed along, and that depends on getting the client from get_openai_client() after instrument(). Content capture only decides whether prompts and completions are recorded. It plays no part in linking spans.
Question 8
Your team deployed GPT-4o in Microsoft Foundry under the deployment name 'vision-prod'. A developer uses the OpenAI() client with the v1 base_url and keyless Entra ID auth, sets model="gpt-4o" in chat.completions.create, and gets an error saying the deployment can't be found. Image requests to that deployment worked earlier from another app. What should the developer change?
- Replace the Entra ID token provider with an API key, because keyless auth can't resolve deployments by name
- Provision an Azure AI Vision Image Analysis resource, because chat deployments can't take image inputs on their own
- Set model="vision-prod", because the model parameter takes the deployment name, not the model family name
- Set model to the full model version string, such as "gpt-4o-2024-11-20", so the service can resolve the exact model
Show the answer
Answer: C. Set model="vision-prod", because the model parameter takes the deployment name, not the model family name
On Foundry, the model parameter names your deployment, so 'vision-prod' is the value to send. A full version string is still a model identifier, not the deployment name, so it fails the same way, and the auth method and a separate Vision service have nothing to do with this error.
Question 9
A developer at a logistics company extracts ticket data with a gpt-4.1-mini deployment in Foundry. The request sets response_format to {"type": "json_object"}, and the system prompt describes the fields. Every response parses, but about 3% are missing the `category` field or contain an extra field called `notes`, and this breaks the downstream loader. What is the best way to make every response match the required shape?
- Send the text to Azure Language in Foundry Tools entity recognition and map its categories to the loader's fields
- Change response_format to type json_schema, supply the schema with strict set to true, list every field in required, and set additionalProperties to false
- Keep JSON mode and paste the full JSON schema into the system prompt, then set temperature to 0
- Keep JSON mode and add three few-shot examples that show the exact field layout
Show the answer
Answer: B. Change response_format to type json_schema, supply the schema with strict set to true, list every field in required, and set additionalProperties to false
JSON mode only guarantees valid JSON. Strict structured outputs guarantee that the output matches the schema, so required fields are always there and extra fields can't appear. Putting the schema in the prompt and lowering the temperature makes mistakes less likely, but nothing enforces the shape.
Question 10
You are an AI engineer at Contoso. Company policy bans API keys. An index's vector profile uses an azureOpenAI vectorizer with no apiKey, and the search service has a system-assigned managed identity. Indexing works, but queries that send text in vectorQueries fail because the call to the Azure OpenAI embedding deployment is denied. What should you do so the vectorizer can embed queries?
- Give the client app's identity the Search Index Data Reader role on the search service.
- Give the search service's managed identity the Cognitive Services OpenAI User role on the Azure OpenAI resource.
- Give the search service's managed identity the Cognitive Services Contributor role on the Azure OpenAI resource.
- Add the Azure OpenAI apiKey to azureOpenAIParameters in the vectorizer definition.
Show the answer
Answer: B. Give the search service's managed identity the Cognitive Services OpenAI User role on the Azure OpenAI resource.
The search service calls the embedding model itself, so its identity needs the Cognitive Services OpenAI User role on the Azure OpenAI resource. Search Index Data Reader only lets the app read the index and does nothing for the call to Azure OpenAI. Adding an apiKey would work technically, but it breaks the no-keys policy.
The full timed mock
50 questions in 100 minutes, with a pass mark of 70%. Each paper is drawn from a bank of questions written for the mock alone, and these 10 are left out of it. Full access opens it, with every level of the course.
LevelUpWith is independent of the organisation that sets this exam. These questions were written for this course; they are not taken from the exam.