Free Plan and manage an Azure AI solution Practice Questions — Microsoft Azure AI Apps & Agents Developer (AI-103)

Question 1 of 6 · Plan and manage an Azure AI solution

A financial services company is deploying a multi-agent system in Azure AI Foundry that processes loan applications. The system uses three specialized agents: a document extraction agent, a risk assessment agent, and a compliance verification agent. The compliance agent must access an external regulatory API that requires OAuth2 authentication, while the risk agent needs to query a private Azure SQL database. Security policy mandates that no credentials be stored in code or configuration files. Which configuration approach BEST meets these requirements?

User-assigned managed identities with federated credentials enable keyless authentication for both the SQL database (via Azure AD authentication) and the external OAuth2 API (via workload identity federation). This eliminates credential storage entirely, meets the security policy requirement, allows granular RBAC assignment per agent capability, and integrates directly with AI Foundry agent tool authentication configurations. Federated credentials enable OAuth2 flows without client secrets.

Question 2 of 6 · Plan and manage an Azure AI solution

An e-commerce platform uses Azure AI Foundry to power a customer service agent that handles product recommendations and order tracking. Telemetry shows that 40% of conversations require grounding queries against a 500GB product catalog with 50M items, updated hourly via batch ingestion. The agent currently uses a single vector index with 1536-dimension embeddings, but users report 8-12 second response latencies during peak hours (10,000 concurrent users). The index rebuild after each hourly update takes 45 minutes, causing stale results. Cost optimization is a secondary concern. Which solution BEST reduces latency while maintaining index freshness?

Hybrid search combining reduced-dimension vectors (768d from distilled models like MiniLM provide 40-60% dimensionality reduction with minimal accuracy loss) with keyword filtering dramatically reduces search latency by offloading exact-match queries to inverted indexes while maintaining semantic capability. Index aliasing enables atomic index swaps—build the new index in parallel, then switch the alias when ready—achieving zero-downtime updates and eliminating the 45-minute stale window. This directly addresses both the latency and freshness requirements at scale.

Question 3 of 6 · Plan and manage an Azure AI solution

A healthcare organization is implementing an Azure AI Foundry agent that assists radiologists by analyzing medical images and generating diagnostic suggestions. The solution must comply with HIPAA requirements and organizational policies that mandate: (1) all AI-generated content must include provenance metadata showing which training data influenced the output, (2) radiologists must explicitly approve AI suggestions before they're added to patient records, and (3) all agent reasoning steps must be auditable for 7 years. The agent uses GPT-4o for image analysis and a custom fine-tuned model for diagnostic classification. Which implementation BEST satisfies these requirements?

Human-in-the-loop mode (agent oversight mode with approval gates) directly satisfies the explicit approval requirement before diagnostic actions. Azure AI Foundry's trace logging with prompt flow runtime captures all reasoning steps, tool calls, and model interactions—2557 days equals 7 years retention in Log Analytics. GPT-4o multimodal grounding with citations enabled provides provenance metadata showing which image regions and training patterns influenced outputs. Immutable storage for conversation transcripts ensures long-term auditability. This is the only option that properly addresses all three requirements with native Foundry capabilities.

Question 4 of 6 · Plan and manage an Azure AI solution

A multinational company deploys an Azure AI Foundry agent that handles customer inquiries in 15 languages. The agent uses GPT-4o with a token context window of 128K and processes an average of 250,000 requests per day with peak loads of 45 requests per second. Each request averages 8,000 input tokens and 1,200 output tokens. The current deployment uses a single model deployment in East US with 150K TPM quota. Cost analysis shows monthly inference costs of $89,000. Performance monitoring reveals P95 latency of 4.2 seconds during peak hours, and 12% of requests exceed the 10-second timeout. Which optimization strategy BEST reduces both cost and latency while maintaining capacity?

Provisioned throughput (PTU) provides predictable costs and eliminates per-token billing, which at this scale (250K requests/day × 9,200 tokens average = 2.3B tokens/month) typically offers 40-60% cost savings over pay-as-you-go at high volume. 200 PTUs distributed across three regions provides ~66-67 PTU per region with automatic geographic distribution, directly addressing the P95 latency issue through proximity routing and eliminating queue contention. Front Door health probes enable intelligent failover. Prompt optimization to reduce output tokens by 30% (360 fewer tokens × 250K requests = 90M fewer tokens/month) compounds savings without sacrificing capability. This is the only option that addresses cost, latency, and scaling systematically.

Question 5 of 6 · Plan and manage an Azure AI solution

A financial analytics platform uses Azure AI Foundry to implement a research agent that processes quarterly earnings reports and generates investment recommendations. The agent uses a RAG pattern with vector search over 150,000 documents stored in Azure AI Search (15GB index). The organization's compliance team has identified that 8% of generated recommendations over the past month contained hallucinated financial figures not present in source documents. Content evaluator logs show grounding scores averaging 0.72 (target: >0.85). The agent uses GPT-4 with temperature=0.3 and includes search citations in prompts. Which intervention BEST improves grounding quality and reduces hallucinations?

Hybrid retrieval (vector + BM25 keyword) is specifically effective for financial data where exact numerical matching is critical—BM25 ensures precise figures are retrieved even if semantic embedding doesn't capture them. Query expansion with financial synonyms improves recall for domain-specific terminology. Chunk-level attributions force the model to ground specific claims in specific source text, directly addressing the hallucination problem. The verification tool (agent self-validation via re-querying) enables the agent to double-check numerical facts before output—a proven pattern for reducing hallucinations in high-stakes domains. This is the only option that combines retrieval improvement, attribution enforcement, and automated verification.

Question 6 of 6 · Plan and manage an Azure AI solution

A media company operates an Azure AI Foundry agent that generates marketing copy and social media content in real-time during live events. The agent must comply with brand safety policies that prohibit mentions of competitor brands, political content, and inappropriate language. During a recent product launch event processing 2,000 requests over 3 hours, safety monitoring detected 47 policy violations (2.35% violation rate) that were published before detection. Current configuration uses Azure OpenAI content filters at default severity levels (medium threshold) and asynchronous content moderation via Azure Content Safety API called after generation. The compliance team requires violation rate below 0.1% and real-time blocking of violating content. Which approach BEST achieves the compliance requirement?

Annotated content filters in Azure OpenAI allow custom keyword lists (competitors, political terms) to be enforced AT GENERATION TIME—the model's output is filtered token-by-token, blocking violations before the response is completed. This is the only native capability that enables real-time blocking during generation rather than post-hoc detection. The dual-agent safety architecture provides domain-specific policy evaluation (GPT-4 can understand nuanced brand safety rules that keyword matching misses). Streaming with sentence-level checks enables mid-generation abort, preventing partial violations from being published. Dynamic blocklist updates address emerging terms during live events. This combines generation-time filtering, semantic policy evaluation, and adaptive term detection.