AI-103 practice test: 1300 questions with full explanations

Questions on the exam

about 60 — vendor indicates, no fixed count published

Time allowed

120 minutes format →

Passing score

700 of 1000 — vendor, checked September 14, 2026 detail →

1300 practice test questions for Microsoft Azure AI Apps and Agents Developer Associate AI-103, grouped by exam domain. Every question below shows all four options, which one is correct, and why each of the other three is not — the wrong answers are where most candidates lose marks.

Not sure where you stand? Take the free 5-min AI-103 readiness check →

AI-103 exam format →  ·  AI-103 passing score →

Questions by domain

• Implement generative AI and agentic solutions — 429 questions →
• Plan and manage an Azure AI solution — 364 questions →
• Implement computer vision solutions — 169 questions →
• Implement text analysis solutions — 169 questions →
• Implement information extraction solutions — 169 questions →

Sample questions

Retrieve approved passages per question and require: Which implementation best meets these requirements?

1. Retrieve approved passages per question and require supported answers or abstention. ✓

   This uses retrieval-augmented generation to supply current document context without retraining. The application can require supported answers and acknowledge insufficient evidence.
2. Fine-tune the deployed model on current support documents, then request concise answers with citations in the prompt.

   Fine-tuning requires repeating model updates as documents change and does not itself establish that citations support the answer. Retrieval better addresses changing external knowledge.
3. Prompt the model to cite approved documents and report confidence, without supplying those documents as context.

   Prompt instructions alone do not provide current external information. The model may produce fluent but unsupported citations or answers.
4. Return JSON containing an answer, citations, and confidence, then validate fields before displaying the response.

   Structured output or JSON mode can shape response fields, but it does not retrieve current documents or prove factual support for citations.

All 429 Implement generative AI and agentic solutions questions →

Create a hosted agent that runs custom orchestration code: Which implementation should you choose?

1. Use Content Understanding to extract requirements into structured fields, without hosting agent code.

   Content Understanding analyzes supported content into structured results; it does not itself provide hosted agent orchestration.
2. Create a prompt agent with instructions and configured tools, without a custom orchestration-code runtime.

   Prompt agents use instructions and tools, but this choice does not provide the required custom orchestration-code runtime.
3. Create a hosted agent that runs custom orchestration code while Foundry manages its runtime hosting. ✓

   Hosted agents are intended for custom code and orchestration with managed hosting, matching both stated requirements.
4. Use RAG to retrieve orchestration logic from business documents at run time.

   RAG supplies external information as model context; it is not a custom-code orchestration runtime or hosting approach.

All 364 Plan and manage an Azure AI solution questions →

Use a supported image-editing model with a mask covering: Which implementation should the developer choose?

1. Use a supported image-editing model with the full photograph as reference, requesting a regenerated storefront.

   This is a real editing workflow, but regenerating the storefront changes a broader area than the requirement permits.
2. Use a supported image-generation model with the photograph as reference, prompting it to recreate the entire storefront.

   Reference-image generation can guide output, but recreating the entire storefront does not target only the damaged sign.
3. Use a supported image-editing model with a mask covering only the damaged sign, then review the returned image. ✓

   A mask defines the selected editing region, while review validates the generated result and preserves responsible media handling.
4. Use a supported image-generation model with an image prompt, requesting a new storefront that resembles the photograph.

   Generating a similar new image neither selects the damaged region nor reliably preserves the existing scene details.

All 169 Implement computer vision solutions questions →

Use structured outputs with a supported schema: Which implementation best meets both requirements?

1. Ask the model for a concise summary containing invoice details, then map the summary into required fields.

   A summary describes text broadly; it is not the task-specific entity-extraction approach needed for reliably producing the required invoice fields. Mapping a summary does not establish source support.
2. Run PII redaction on the invoice text, then use the redacted text to populate the required invoice fields.

   PII detection and redaction address sensitive-entity handling, not fixed structured extraction or verification of invoice values. Redaction may also remove information needed for extraction.
3. Use JSON mode for valid JSON, then treat successfully parsed fields as supported invoice facts.

   JSON mode targets valid JSON but does not provide the same schema constraint as structured outputs. Parsing success also does not prove that the fields are supported by the source text.
4. Use structured outputs with a supported schema, then validate extracted values against the source text. ✓

   Structured outputs constrain the response to the required JSON schema. Separate validation is still necessary because schema conformance does not establish that extracted values are factually supported by the invoice text.

All 169 Implement text analysis solutions questions →

Use integrated vectorization for indexing and queries: Which implementation best satisfies these requirements?

1. Apply semantic ranking to indexed text, then add query embeddings without rebuilding documents or checking vector dimensions.

   Document and query vectors require compatible dimensions and embedding space; changing representations requires appropriate reindexing or migration.
2. Use keyword retrieval with metadata filters, then apply semantic ranking to improve relevance for natural-language questions.

   Semantic ranking reranks text results but does not replace vector retrieval for semantic similarity across differently worded content.
3. Generate embeddings during ingestion and query execution, then use vector-only retrieval with document chunking and access metadata.

   Vector-only retrieval can support semantic similarity but does not provide keyword matching optimized for exact identifiers.
4. Use integrated vectorization for indexing and queries, then combine vector and keyword retrieval with metadata access filters. ✓

   This supports semantic similarity, exact identifier matching, and application access controls in one retrieval pipeline.

All 169 Implement information extraction solutions questions →

Configure the request to use structured outputs: Which implementation change addresses the missing prerequisit

1. Configure the request for JSON mode and parse the returned JSON before applying application validation.

   JSON mode targets valid JSON but does not provide the same schema-conformance guarantee required here. Parsing and validation do not replace structured-output configuration.
2. Define a function tool with the schema and return its call result to the model.

   Function calling describes tool parameters and enables application execution, but a tool schema does not itself impose the required response schema or guarantee conformance.
3. Add a retrieval step that supplies current business documents as additional model context.

   Retrieval can ground responses in external information, but it does not constrain the response to a JSON schema. Retrieval also requires separate inspection of relevance and support.
4. Configure the request to use structured outputs with the application’s supported JSON schema. ✓

   Structured outputs constrain supported model responses to the specified supported JSON schema. The application must still validate business rules and handle refusals or incomplete responses.

All 429 Implement generative AI and agentic solutions questions →

Select a supported model and configure it as the prompt: What should the developer do first?

1. Add Content Understanding to convert policy documents into structured business records instead of configuring the model.

   Content Understanding analyzes source content into structured output, but it does not provide the model required by a prompt agent. The scenario already has document retrieval configured.
2. Replace it with a hosted agent and add custom orchestration code for document answers.

   Hosted agents are appropriate when custom code or orchestration is required, but the scenario requires neither. This changes the runtime instead of supplying the missing model configuration.
3. Select a supported model and configure it as the prompt agent’s response model. ✓

   A prompt agent requires a model in addition to instructions and tool configuration. Selecting the supported model resolves the stated missing prerequisite; document retrieval and permissions are already complete.
4. Add retrieval sources so the agent can answer from documents despite lacking a configured model.

   Retrieval supplies external information for model context; it cannot replace the model that generates the agent response. Adding sources does not address the reported prerequisite.

All 364 Plan and manage an Azure AI solution questions →

Select an image-generation model explicitly supporting: Which action satisfies the requirement?

1. Run image safety filtering on the prompt and output, then resubmit the same reference image and mask.

   Safety filtering is an appropriate media-pipeline control, but it does not change the selected model's generation-only capability into image editing.
2. Select an image-generation model explicitly supporting mask-based editing, then submit the prompt, reference image, and mask. ✓

   A model supporting the requested editing modality is the missing prerequisite. The mask can then identify the region to replace, subject to model-specific constraints.
3. Keep the selected model and refine the prompt to describe the desired replacement while retaining the supplied reference image and mask.

   Prompt refinement can improve instructions, but it cannot add image-editing capability to a generation-only model. The required modality remains unsupported.
4. Apply a watermark and branding policy to the generated output, while retaining the supplied prompt, reference image, and mask.

   Watermarking and branding address governance and provenance signaling; they do not enable mask-based editing or selected-region replacement.

All 169 Implement computer vision solutions questions →

AI-103 exam: the facts

How many questions are on the AI-103 exam?

Around 60. The vendor does not publish a fixed count for AI-103, so this is the figure it indicates rather than a guaranteed number.

How long is the AI-103 exam?

120 minutes. Across 60 questions that is about 120 seconds per question.

What topics does the AI-103 exam cover?

5 domains: Implement generative AI and agentic solutions, Plan and manage an Azure AI solution, Implement computer vision solutions, Implement text analysis solutions, Implement information extraction solutions. Weights: Implement generative AI and agentic solutions 0.325%, Plan and manage an Azure AI solution 0.275%, Implement computer vision solutions 0.125%, Implement text analysis solutions 0.125%, Implement information extraction solutions 0.125%.

How many AI-103 practice test questions does Certsqill have?

1300, spread across 5 exam domains. Every one shows all options, which is correct, and why each of the others is not.