AI-103: Developing AI Apps and Agents on Azure Practice Exam

What the AI-103 exam covers

• Plan and manage an Azure AI solution222 questions
• Implement generative AI and agentic solutions266 questions
• Implement computer vision solutions100 questions
• Implement text analysis solutions100 questions
• Implement information extraction solutions100 questions

Free AI-103 sample questions

A sample of 10 questions with answers and explanations. Sign up free to practice all 788.

1. Question 1Plan and manage an Azure AI solution

   You are building a Python agent in Microsoft Foundry that must extract structured fields from scanned invoices, then summarize the extracted totals in fluent prose. Which combination of models BEST fits this pipeline?

   • AA small language model for extraction and an LLM for the summary
   • BDocument Intelligence for extraction and an LLM for the summaryCorrect
   • CA multimodal LLM for both extraction and summary
   • DAn embedding model for extraction and an LLM for the summary
   ✓ Correct answer: B

   Scanned invoices are semi-structured documents where Azure AI Document Intelligence prebuilt invoice models reliably return typed fields such as InvoiceTotal with high accuracy and low cost. Passing those fields to an LLM produces the fluent narrative, so each task uses the tool built for it.

   Why the other options are wrong

   • AA small language model is text-only and cannot read a scanned image, so it cannot perform the extraction step.
   • CA single multimodal LLM can read images but is far less accurate and more costly for high-volume structured field extraction than Document Intelligence.
   • DEmbedding models produce vectors for similarity search, not structured field values, so they cannot extract invoice fields.
2. Question 2Plan and manage an Azure AI solution

   A financial agent must call an internal accounts API using the caller's own identity so downstream authorization and auditing reflect the real user. Which approach should you use for the tool's access?

   • AOn-behalf-of the signed-in user with Entra ID token exchangeCorrect
   • BA single shared service account key for all users
   • CThe agent's own managed identity for every call
   • DAnonymous access with IP allow-listing
   ✓ Correct answer: A

   Using the Microsoft Entra ID on-behalf-of flow, the agent exchanges the user's token to call the downstream API as that user. This preserves per-user authorization and produces audit trails tied to the actual caller rather than a shared or agent identity.

   Why the other options are wrong

   • BA shared service account key hides who acted and breaks per-user authorization and auditing.
   • CThe agent's managed identity represents the app, not the individual user, so downstream logs lose the caller.
   • DAnonymous access with IP allow-listing removes user identity entirely and cannot audit by user.
3. Question 3Implement generative AI and agentic solutionsSelect all that apply

   You want to continuously monitor a production Foundry agent for both operational health and answer quality. Which TWO signals should you collect to cover both dimensions? (Choose TWO.)

   • ALatency and error rate from trace spans for operational healthCorrect
   • BContinuous evaluation scores such as groundedness on sampled responsesCorrect
   • CThe color theme configured in the Foundry portal
   • DThe number of exam certifications held by the developers
   • EThe alphabetical order of the agent names in the project
   ✓ Correct answer: A, B

   Latency and error rate from trace spans capture operational health, while continuous evaluation scores like groundedness on sampled production responses capture answer quality, together covering both dimensions. Portal theme, developer certifications, and name ordering are irrelevant to monitoring agent behavior.

   Why the other options are wrong

   • CThe portal color theme is a UI preference with no bearing on agent health or quality.
   • DDeveloper certifications are unrelated to runtime monitoring of the deployed agent.
   • EAgent name ordering is cosmetic metadata and carries no operational or quality signal.
4. Question 4Implement text analysis solutions

   You process customer feedback that arrives in several languages, and analysts want a single English sentiment report. Which pipeline order in Microsoft Foundry is the MOST reliable?

   • ATranslate to English with Translator, then run sentimentCorrect
   • BRun sentiment analysis first, then translate the labels
   • CRedact PII, then translate, then discard the text
   • DSummarize each item, then detect its language
   ✓ Correct answer: A

   Normalizing all feedback to English first with Azure Translator gives the sentiment step a single, consistent language to score, producing a unified English report. Translating before analysis avoids inconsistent multilingual sentiment behavior. This ordering is the reliable pattern for cross-language reporting.

   Why the other options are wrong

   • BTranslating only the labels leaves the underlying scoring done inconsistently across many languages before normalization.
   • CDiscarding the text after redaction leaves nothing to run sentiment on, so no report is produced.
   • DSummarizing then detecting language never produces sentiment and puts language detection after the point it is needed.
5. Question 5Plan and manage an Azure AI solution

   A pilot chat feature has bursty, unpredictable traffic during business hours and none overnight, and the team wants to pay only for tokens actually consumed while iterating. Which Azure OpenAI deployment type should you choose FIRST?

   • AStandard (pay-as-you-go) deploymentCorrect
   • BProvisioned throughput units reserved monthly
   • CA dedicated capacity reservation
   • DA batch endpoint scheduled hourly
   ✓ Correct answer: A

   A standard deployment charges only for tokens processed, so an unpredictable pilot that goes idle at night pays nothing while quiet and avoids committing to reserved capacity. Provisioned options bill for reserved throughput whether or not it is used, which wastes money during idle hours.

   Why the other options are wrong

   • BProvisioned throughput units bill for reserved capacity around the clock, wasting spend during idle overnight hours.
   • CA dedicated capacity reservation commits cost regardless of the bursty, low pilot usage.
   • DA batch endpoint is for offline throughput, not an interactive chat feature.
6. Question 6Plan and manage an Azure AI solutionSelect all that apply

   Your Standard GPT-4o deployment intermittently returns 429 responses, and you want to make the client resilient while also reducing the chance of hitting the limit. Which TWO client-side techniques should you implement? (Choose TWO.)

   • ARetry with exponential backoff that honors the Retry-After headerCorrect
   • BBatch multiple small prompts into fewer larger requests where feasibleCorrect
   • CRetry immediately in a tight loop until the request succeeds
   • DLower the deployment's TPM quota to smooth traffic
   • EIncrease max_tokens so fewer total requests are needed
   ✓ Correct answer: A, B

   Exponential backoff that respects the Retry-After header spaces retries so the client recovers gracefully from transient throttling instead of amplifying it. Consolidating many small prompts into fewer larger requests reduces request-rate pressure against the RPM limit. Together they improve resilience and lower the likelihood of tripping the limits.

   Why the other options are wrong

   • CA tight immediate-retry loop hammers the throttled endpoint and worsens the overload rather than recovering from it.
   • DLowering the TPM quota shrinks the available ceiling, making throttling more likely, not less.
   • ERaising max_tokens lets responses grow longer and consumes more tokens, which does not reduce request count for distinct prompts.
7. Question 7Implement generative AI and agentic solutions

   For a regression test suite you need a gpt-4o deployment to return the same completion for the same prompt as often as possible across runs. You already set temperature to 0. Which additional step gives the BEST chance of reproducible outputs and lets you confirm the backend state matched?

   • APass a fixed seed value and compare the system_fingerprint returned across runsCorrect
   • BPass top_p of 0 alongside temperature 0 to remove all sampling
   • CSet logprobs to true so the token probabilities force identical output
   • DCache the first response and replay it for every later identical request
   ✓ Correct answer: A

   Supplying a seed requests best-effort deterministic sampling for identical parameters. The response includes system_fingerprint, which identifies the backend configuration; if it changes between runs, determinism is not guaranteed even with the same seed. Together seed and fingerprint let you both request and verify reproducibility.

   Why the other options are wrong

   • Btop_p 0 further narrows sampling but does not provide the deterministic seed mechanism or any way to verify backend consistency.
   • Clogprobs only reports probabilities for emitted tokens and has no effect on determinism.
   • DReplaying a cached answer sidesteps the model entirely and does not test the deployment's actual reproducibility.
8. Question 8Implement generative AI and agentic solutions

   A regulated agent must answer only from an approved internal knowledge base and must never fabricate or pull external content. Which combination of tool choices BEST enforces this?

   • AAttach only the Azure AI Search tool over the approved index and instruct the agent to answer solely from retrieved resultsCorrect
   • BAttach both the Bing grounding tool and the Azure AI Search tool and let the model freely choose which to use per query
   • CAttach only the Code Interpreter tool and upload the entire approved knowledge base as files for the sandbox to search
   • DAttach an OpenAPI tool pointed at a public knowledge API and rely on that external endpoint for all grounded answers
   ✓ Correct answer: A

   Limiting the agent to a single knowledge tool over the approved index removes any path to external content, and grounding instructions constrain answers to retrieved passages. This enforces the closed-corpus, no-fabrication requirement. Restricting the tool set is the primary control for regulated grounding.

   Why the other options are wrong

   • BIncluding Bing grounding gives the agent a route to external web content, violating the no-external-content rule.
   • CCode Interpreter runs sandboxed code and does not perform grounded retrieval or citation over a knowledge base.
   • DA public knowledge API returns external content, which the closed-corpus requirement explicitly forbids.
9. Question 9Implement computer vision solutions

   You must caption dense scenes so that each distinct region gets its own localized description with coordinates, for example describing several people and objects separately rather than one sentence for the whole image. Using Azure AI Vision Image Analysis, which feature returns this?

   • ADense Captions, which returns up to ten region captions each with a bounding boxCorrect
   • BCaption, which returns one sentence for the whole image
   • CTags, which returns content labels without descriptions or regions
   • DSmart Crops, which returns cropping regions optimized for thumbnails
   ✓ Correct answer: A

   The Dense Captions feature generates a separate human-readable caption for each detected region of interest and returns a bounding box for each one, up to ten regions. That matches the need for localized, per-region descriptions rather than a single global sentence. It is the feature designed specifically for granular scene description.

   Why the other options are wrong

   • BCaption produces only a single sentence summarizing the entire image, not per-region localized descriptions.
   • CTags lists content labels with confidence scores but provides neither descriptive sentences nor region coordinates.
   • DSmart Crops returns suggested crop rectangles for thumbnail generation, not descriptive captions of regions.
10. Question 10Implement text analysis solutions

    A safety monitoring agent must transcribe factory-floor audio and flag non-speech sounds such as alarms, breaking glass, and machinery faults alongside any spoken words. Which Azure AI Speech capability should you enable to capture these non-speech signals?

    • AAudio event detection in the transcription outputCorrect
    • BSpeaker diarization with participant labels
    • CA custom speech model trained on factory jargon
    • DProfanity masking on the recognized text
    ✓ Correct answer: A

    Audio event detection identifies non-speech acoustic events such as alarms, breaking glass, or mechanical sounds and emits them in the transcription results. This is distinct from transcribing words, which is what the safety scenario needs to flag. Enabling it surfaces the acoustic signals a text-only transcript would ignore.

    Why the other options are wrong

    • BSpeaker diarization attributes spoken sentences to different people and does not detect non-speech sounds.
    • CA custom speech model improves recognition of spoken domain terms, not detection of acoustic events.
    • DProfanity masking only sanitizes recognized words and captures no non-speech audio.

Related Microsoft resources

• AI-103 study guideKey concepts
• Microsoft practice examsAll Microsoft
• Certification pathWhere this fits
• Certification exam guides & tipsBlog
• Plans & pricingFree & paid
• SC-200 practice examRelated
• AZ-140 practice examRelated
• AZ-900 practice examRelated

AI-103 practice exam FAQ