AI-102 sample questions with answers
The other 151 questions are in the practice tool, with answers revealed as you go.
Start 20 questionsPlan and manage an Azure AI solution
- A healthcare platform runs a read-text service as containers in a Kubernetes cluster. Release managers want repeatable deployments and quick reversion when a build misbehaves. Which approach meets this?
- A credit union uses a generative model to draft explanations of loan eligibility for its underwriters. The compliance team wants the design to reflect responsible AI practices. Which choice best does that?
- A fintech runs a Kubernetes scoring service and a serverless function, and both call the same Azure AI resource. The security team wants each workload to have distinct, auditable permissions with no shared credentials. Which approach fits?
Implement generative AI solutions
- A nightly job feeds model output into a billing parser that crashes whenever a field is missing or renamed. Which prompt adjustment would most reduce these failures?
- An enterprise search assistant gives wrong answers for about 5% of questions. The team cannot tell whether the search layer returned the wrong passages, the prompt template dropped them, or the model misread them. Which capability would isolate the failing stage?
- A bank wants customers to ask questions about their scanned mortgage statements and receive answers with amounts pulled from specific form fields. The statements arrive as image-based PDFs. Which pipeline is most appropriate?
- An enterprise wants several generative AI solutions to share one set of approved model connections and governance policies, while each solution keeps its own datasets, evaluations and deployments. What structure fits in Microsoft Foundry?
- A knowledge-base team evaluates a grounded question-answering pipeline and wants a score that shows whether each claim in a generated reply is backed by the retrieved context. Which metric should they report?
- A maintenance assistant for technicians must answer from the latest service bulletins and also explain a photo of a worn gearbox the technician uploads. Which architecture is the best fit?
Implement an agentic solution
Implement computer vision solutions
- A retail app needs a ready-made service that describes a single photo of a shelf with a sentence and detects the items in it, with no model training. Which option fits?
- One still photo must produce a descriptive caption, content tags, detected objects, and any printed text, all from a single analysis request. Which option handles this?
- A bakery's Custom Vision model recognizes its current cake boxes. The company introduces a new seasonal box design, and the model misses it entirely. What should the team do?
Implement natural language processing solutions
- A podcast network needs transcripts for 8,000 archived episodes stored in cloud storage by Monday. No live captions are required. Which approach fits best?
- A pharmaceutical company's leaflets must keep drug names and mandated warning sentences identical when translated from English into Spanish and German. Generic machine translation alters them. What should the company use?
- A railway announcement system must insert a short silence before a safety warning and deliver that warning at a slower speaking rate. The rest of the announcement should keep default timing. Which mechanism should the developer put in the synthesis input?
- A logistics company's voice ordering system misrecognizes spoken SKU codes and warehouse abbreviations. The team has recordings of staff saying these terms along with verified transcripts. What should they do to improve recognition of these terms?
- An English tutoring app asks students to read a passage aloud and must score how accurately each sound was produced, how smoothly the student spoke, and whether every word was said. Which Speech feature should the developer use?
- A telecom quality team has 50,000 recorded support calls to transcribe overnight. The calls contain heavy use of network equipment model names that generic recognition often gets wrong. Which approach best meets these requirements?
Implement knowledge mining and information extraction
- A bank must send each loan file to its internal underwriting scorer during indexing and store the returned grade as a searchable field. Built-in skills cannot provide this grade. What should the team use?
- A claims department needs structured fields pulled from dashcam videos, scanned forms, call recordings, and damage photos, all mapped to one shared business schema. Which service fits best?
- A recipe site wants one query to match exact ingredient names and also find recipes whose embeddings are close to a shopper's free-form description. Which query mode should the developer use?
- A bookkeeping service receives thousands of standard vendor invoices each month and needs the vendor name, total amount, and invoice date extracted. The team has no labeled samples and wants to start immediately. Which Document Intelligence option is most appropriate?
- A retail site's search results must include a short highlighted passage from the best matching product page and a direct answer line above the results. Which Azure AI Search setup delivers this?