Free AI-103 practice test: 20 questions
Twenty questions spread across the five AI-103 skill areas in roughly the same proportion as the real exam. Work through them without notes, then check your answers. If you miss more than six, go back to the domain guides before booking.
Plan and manage an Azure AI solution
Question 1. You are selecting a region for a new Azure AI Foundry deployment. The application must use a specific generative model version, and legal requires that data stays inside the EU. What should drive the region choice?
- A. The region with the lowest latency to your office
- B. The region with the lowest compute price
- C. A region that offers the required model version and lies within the EU
- D. The region where your existing storage account lives
Show answer
Answer: C
Model availability differs per region, so the region must offer the exact model version you need as well as satisfy the data residency requirement. Picking on latency or price alone risks choosing a region where the model is unavailable.
Question 2. A team reports HTTP 429 responses from an Azure AI service during a nightly batch job. The job must finish inside its window. What is the most appropriate first response?
- A. Retry immediately in a tight loop until the call succeeds
- B. Implement exponential backoff and smooth the request rate across the window
- C. Switch the resource to a different region
- D. Disable content filtering to reduce processing time
Show answer
Answer: B
429 means the request rate exceeded the provisioned quota. Implementing retry with exponential backoff and spreading the load respects the limit and is the standard client-side handling. Raising quota may also be needed, but uncontrolled retries make throttling worse.
Question 3. Your organisation needs to trace which employee triggered each generative AI request for a compliance audit, without storing prompt content. What should you configure?
- A. Diagnostic settings that export request metadata to a log workspace
- B. Full prompt and completion logging to blob storage
- C. Client-side logging inside the browser
- D. No logging, relying on the platform’s built-in retention
Show answer
Answer: A
Diagnostic logging to a log sink records request metadata such as identity, timestamp and operation without capturing prompt bodies. Storing full prompts would conflict with the requirement not to retain content.
Question 4. You must give a containerised workload running outside Azure access to an Azure AI service, with credentials that rotate automatically. Which approach fits best?
- A. A service account with an API key in an environment variable
- B. A shared key checked into the container image
- C. A connection string distributed by email
- D. Workload identity federation with Microsoft Entra ID
Show answer
Answer: D
Workload identity federation lets an external workload exchange its own token for an Entra ID token, so no long-lived secret is stored anywhere. Keys and connection strings would have to be stored and rotated manually.
Question 5. Cost for a generative AI workload is rising faster than usage. Investigation shows prompts include the entire knowledge base on every call. What is the correct fix?
- A. Lower the temperature setting
- B. Retrieve only the relevant passages and pass those to the model
- C. Cap the maximum response length
- D. Move the workload to a cheaper region
Show answer
Answer: B
Retrieving only the relevant passages at query time cuts input tokens dramatically while preserving answer quality. Lowering temperature or shortening replies does not address the input token volume, which is where the cost sits.
Implement generative AI and agentic solutions
Question 6. An agent must call an internal ordering API, but only after a human approves the order. How should the workflow be designed?
- A. Let the agent call the API and email a notification afterwards
- B. Have the agent write the order to a file that someone checks weekly
- C. Have the agent propose the call and require explicit approval before it executes
- D. Give the agent the API key and log the calls
Show answer
Answer: C
The agent proposes the tool call and pauses for explicit approval before execution. Letting the agent call the API directly removes the control the requirement demands, and approving afterwards is too late.
Question 7. Your RAG application returns passages that are topically related but rarely answer the question. Chunk size is 4,000 tokens per document section. What is the most likely improvement?
- A. Reduce chunk size and add overlap between chunks
- B. Increase the number of chunks returned to 50
- C. Raise the model temperature
- D. Remove the system prompt
Show answer
Answer: A
Very large chunks dilute the embedding, so retrieval matches the general topic rather than the specific answer. Smaller, overlapping chunks produce sharper embeddings and more precise retrieval.
Question 8. A customer-facing agent must never discuss competitors' products. Which control belongs in the design?
- A. Ask users politely not to raise the subject
- B. Lower the temperature to zero
- C. Fine-tune on competitor documentation so the model recognises it
- D. A scoping system instruction plus an output filter for prohibited topics
Show answer
Answer: D
A system instruction defining scope, combined with output filtering, enforces the boundary at both ends. Relying only on training data or user instructions leaves the behaviour to chance.
Question 9. You need an agent to answer from a 12,000-page manual that is updated weekly. Which architecture is most appropriate?
- A. Fine-tune a model on the manual each week
- B. Index the manual and retrieve relevant sections at query time
- C. Paste the manual into the system prompt
- D. Summarise the manual once and use the summary
Show answer
Answer: B
RAG over an index keeps answers current as the manual changes, with no retraining. Fine-tuning would need repeating weekly, and no context window holds 12,000 pages.
Question 10. Two agents must cooperate: one gathers data, another writes a report. The report agent sometimes starts before data collection finishes. What should you add?
- A. A longer timeout on the report agent
- B. A retry loop on the report agent
- C. An orchestrator that sequences the agents and passes state between them
- D. A larger model for the report agent
Show answer
Answer: C
An orchestration layer that sequences the agents and passes completion state prevents the second agent running early. Increasing timeouts or retrying masks the ordering problem rather than fixing it.
Question 11. Your evaluation shows high groundedness but low relevance. What does this indicate?
- A. Answers are faithful to the sources but do not address the question
- B. Answers are invented and unsupported
- C. The model’s language quality is poor
- D. The index is empty
Show answer
Answer: A
Groundedness confirms answers are supported by retrieved sources; low relevance means those answers do not address what was asked. The retrieval is returning supportable but off-target content.
Question 12. A prompt injection attempt hidden inside a retrieved document caused an agent to ignore its instructions. Which mitigation addresses this most directly?
- A. Shorten the system prompt
- B. Block that one document
- C. Raise the temperature
- D. Treat retrieved content as untrusted data, isolated from the instruction channel
Show answer
Answer: D
Treating retrieved content as untrusted data rather than instructions, and separating it from the instruction channel, is the direct defence. Blocking the source or shortening prompts does not generalise.
Implement computer vision solutions
Question 13. You must detect whether safety helmets are worn in factory photos and report the location of each head. Which capability do you need?
- A. Image classification
- B. Object detection
- C. Optical character recognition
- D. Image captioning
Show answer
Answer: B
Object detection returns both a class and a bounding box for each instance, which satisfies the location requirement. Image classification returns a single label for the whole picture with no position.
Question 14. A vision model performs well in testing but poorly on the factory floor, where lighting is uneven and cameras sit at odd angles. What is the most effective response?
- A. Increase the confidence threshold
- B. Use a larger pre-trained model
- C. Add training images captured under real production conditions
- D. Convert images to grayscale
Show answer
Answer: C
Adding training images that reflect real conditions, including the awkward lighting and angles, closes the gap between test and production distributions. A bigger model trained on the same unrepresentative data will not help.
Question 15. You need to read handwritten delivery notes, including printed and handwritten fields, into text. Which service capability applies?
- A. Optical character recognition with handwriting support
- B. Object detection
- C. Image captioning
- D. Face detection
Show answer
Answer: A
OCR with handwriting support extracts both printed and handwritten text. Object detection locates items but does not read them, and captioning describes a scene rather than transcribing it.
Implement text analysis solutions
Question 16. Support tickets must be routed by the product mentioned in the text, using a custom taxonomy of 40 internal product names. Which approach is most appropriate?
- A. Sentiment analysis
- B. Key phrase extraction
- C. Custom named entity recognition trained on labelled tickets
- D. Language detection
Show answer
Answer: C
A custom named entity recognition model trained on your own labelled examples recognises internal product names, which a general-purpose model has never seen. Sentiment and key phrase extraction do not map text to a taxonomy.
Question 17. You must remove names, phone numbers and account numbers from transcripts before they are stored. Which capability fits?
- A. Summarisation
- B. PII detection and redaction
- C. Translation
- D. Sentiment analysis
Show answer
Answer: B
Personally identifiable information detection and redaction identifies and masks those categories directly. Summarisation shortens text without guaranteeing removal, and translation changes language only.
Question 18. A multilingual feedback form must be analysed for satisfaction trends, but you do not know which language each entry uses. What should the pipeline do first?
- A. Detect the language, then run sentiment analysis with the right configuration
- B. Run sentiment analysis and ignore the language
- C. Translate everything to a single language and discard the original
- D. Reject entries that are not in English
Show answer
Answer: A
Language detection determines the language so the correct model or setting is applied downstream. Running sentiment analysis blindly risks poor accuracy on unexpected languages.
Implement information extraction solutions
Question 19. Thousands of supplier contracts must yield party names, dates and payment terms as structured fields. Layouts vary widely by supplier. Which approach is most robust?
- A. Regular expressions over raw OCR output
- B. Manual data entry
- C. Image classification per contract
- D. A document extraction model trained on layout and field labels
Show answer
Answer: D
A trained document extraction model that combines layout understanding with field extraction handles varied formats. Regular expressions on raw OCR text break as soon as layout changes.
Question 20. An extraction pipeline returns a confidence score per field. Finance requires that no incorrect total ever reaches the ledger. How should confidence be used?
- A. Accept all extracted values and correct errors later
- B. Route fields below a confidence threshold to human review before posting
- C. Reject any document containing a low-confidence field
- D. Ignore confidence and rely on the model’s accuracy
Show answer
Answer: B
Routing low-confidence fields to human review before posting protects the ledger while keeping automation for the rest. Accepting everything or rejecting whole documents sacrifices either accuracy or throughput.
How did you do?
Sixteen or more correct suggests you are close to exam-ready. Below fourteen, the domain guides in this section are the fastest way to close the gap — each one explains the concepts behind these questions and adds three more questions of its own.