AI-103 Practice Test

Sample AI-103 Practice Questions

10 real questions from the free AI-103bank. Pick an answer in your head, then open “Show answer” for the explanation.

1. 1. Your Hub has a single provisioned‑throughput Azure OpenAI deployment capped at 50 000 tokens per minute (TPM). Three Projects share this deployment. During peak load, Project A consumes the majority of capacity, starving Projects B and C. What change guarantees each Project a minimum TPM allocation without reducing Project A's peak capacity when the other Projects are idle?

   • A. Increase the deployment's PTU count, expecting Foundry to distribute tokens evenly across the Projects
   • B. Apply an Azure Policy that caps Project A's token consumption at the subscription scope
   • C. Create a separate provisioned-throughput deployment per Project and point each Project's connection to its own
   • D. Add a token‑budget instruction to the system prompt of each Project's base agent to limit per‑session consumption
   Show answer

   Correct answer: C

   Provisioned Throughput Units (PTU) are scoped to a deployment. Creating separate deployments per Project gives each an independent, guaranteed TPM allocation. Increasing PTUs on a shared deployment (A) does not enforce per‑project minimums. Azure Policy (B) applies subscription‑level governance, not per‑deployment allocation. System‑prompt token budgets (D) are logical instructions that do not affect infrastructure‑level throughput limits.

2. 2. Your company policy requires that every AI‑generated response include a confidence score. The model does not natively produce confidence scores. How can you implement this compliantly?

   • A. Prompt the model to express its own uncertainty using phrases such as 'I am confident' or 'I am not certain'
   • B. Use a separate small model to evaluate the response and assign a confidence score post‑generation
   • C. Ask the user to rate each response and calculate an average
   • D. Use Azure AI Content Safety's harm-category confidence thresholds as the response confidence score
   Show answer

   Correct answer: B

   A post-generation evaluator emits a numeric confidence consistently for every response, satisfying the policy without requiring model changes. Model-prompted uncertainty produces inconsistent natural-language hedges, not a programmatic score. User ratings happen after delivery rather than alongside the response. Content Safety confidence scores measure harm-category likelihood, not answer correctness, and cannot substitute for response confidence.

3. 3. The 95th percentile latency of your agent spikes intermittently. Tracing shows that the vector search takes 200 ms normally but occasionally 3 seconds. What should you investigate?

   • A. Model provider outage
   • B. Vector index size and the number of vector dimensions configured per document
   • C. Azure OpenAI deployment quota throttling at the model layer
   • D. Search service scaling and partition configuration
   Show answer

   Correct answer: D

   Intermittent latency on the vector search step - not the model call - points to AI Search infrastructure. Replica count handles concurrent query load and partition count handles index size; both are the documented scaling levers for AI Search. A model provider outage would fail the call rather than slow it. Vector dimensions affect baseline latency uniformly, not intermittent spikes. Model-layer quota throttling would slow the LLM call, not the vector search step shown in the trace.

4. 4. Your skillset includes a slow custom Web API skill that limits overall indexer throughput. How do you prevent the custom skill from blocking the rest of the skillset on re-indexing runs?

   • A. Enable parallel skill execution so the custom skill runs concurrently with other skills rather than blocking the pipeline
   • B. Increase the indexer timeout to allow more processing time per document
   • C. Deploy the custom skill in a faster Azure region with lower network latency to the indexer
   • D. Enable skillset execution caching so the indexer reuses stored skill outputs for documents whose inputs have not changed
   Show answer

   Correct answer: D

   Skillset execution caching stores skill outputs in Azure Storage and reuses them on re-indexing runs when inputs are unchanged, eliminating the slow skill's cost for unchanged documents. Parallel execution does not help re-runs because the slow skill still executes for each document. Increasing timeout permits longer execution but does not avoid the cost. Region selection affects only network latency, not the underlying skill compute.

5. 5. When should you prefer a user-assigned managed identity over a system-assigned managed identity for Azure AI Foundry access?

   • A. When multiple Azure resources need to share the same identity with centralized lifecycle management
   • B. When the resource has a short lifespan and the identity should be automatically deleted with it
   • C. When the workload requires Microsoft Entra B2B federation with a partner organization's identity provider
   • D. When cross-subscription RBAC assignments are required
   Show answer

   Correct answer: A

   User-assigned managed identities are designed for scenarios where multiple resources share one identity, enabling centralized RBAC assignment and independent lifecycle management. System-assigned identities are tied to one resource and deleted when that resource is deleted.