AIAI Agent Dev Labaiagentdev-lab.com
Sign in to save progressSign in
Menu

AI-103 · LESSON 1

Plan and manage an Azure AI solution25–30% · 4 topics

Learn how to choose Microsoft Foundry services, set up AI solutions, manage, monitor and secure them, and implement responsible AI.

1.1Choose the appropriate Foundry services for generative AI and agents

Concept

Microsoft Foundry includes models (Foundry Models: Azure OpenAI models plus partner and open models), Foundry Agent Service (building and running agents), Foundry Tools (AI services such as Vision, Language, Speech and Content Understanding) and Azure AI Search (search and RAG). Combine the ones that fit your requirements.

How it works / steps

  1. Break the problem into “generation,” “agents (work using tools),” “analysis (images, documents, audio)” and “search”
  2. Choose Foundry Models for generation and chat, and Agent Service for autonomous processing with tools
  3. Prefer prebuilt AI services (Language, Vision, Speech, Document Intelligence and so on) for routine analysis
  4. Combine answers based on internal data with Azure AI Search (or agent knowledge)
  5. Compare performance, cost and regional availability in the model catalog before choosing

Workplace example

Invoice processing extracts fields with Content Understanding, while for inquiries an agent searches policies with AI Search and calls internal APIs as tools when needed.

Common confusions and exam tips

Mini quiz

Q1. You want to classify tens of thousands of inquiries a day into “billing,” “delivery” and “other.” Quality requirements are moderate, and you want to keep cost and latency low. Which model should you evaluate first?

  1. A small language model (SLM)
  2. The largest reasoning-specialized model
  3. An embedding model
  4. A fine-tuned large model
Show answer and explanation

Answer: A A small language model (SLM)

High-volume routine classification can often meet requirements with a small model, which is favorable for cost and latency. [Other options] Reasoning models are high quality but costly and slow, excessive for simple classification. An embedding model alone does not output class labels. A fine-tuned large model incurs training and operating costs and is not the first candidate to evaluate.

Q2. Which is the most recommended authentication method when an app calls Foundry models?

  1. Store a service principal’s client secret in a configuration file
  2. Sign in with a developer’s Entra ID account and share that token
  3. Keep the API key in Azure Key Vault and load it when the app starts
  4. Use a managed identity and grant the minimum roles with RBAC
Show answer and explanation

Answer: D Use a managed identity and grant the minimum roles with RBAC

Keyless authentication removes key leakage and rotation management issues. Keep permissions minimal. [Other options] Keys in Key Vault still require leak and rotation management and are less recommended than keyless. Secrets in configuration files or sharing a developer’s personal token are not safe.

Reference links

1.2Set up AI solutions in Foundry

Concept

Create a Foundry resource and project, deploy a model and call the endpoint from an SDK (Foundry SDK, OpenAI SDK). Microsoft Entra ID (managed identity) is the recommended authentication, and external services are linked through connections.

How it works / steps

  1. Create a Foundry resource and project
  2. Choose a model from the catalog, decide the deployment type (Standard, Global Standard, Provisioned and so on) and deploy
  3. Call it from the app using the project endpoint and Microsoft Entra ID authentication
  4. Register AI Search, Storage and so on as connections
  5. Verify behavior in the playground before moving to code

Workplace example

Prototype a chat in the playground, settle the system message and parameters, then build it into a web app that authenticates with a managed identity.

Common confusions and exam tips

Mini quiz

Q1. During month-end closing, model calls frequently fail with 429 errors (too many requests). What is the appropriate first response?

  1. Resend failed requests immediately without delay and repeat until they succeed
  2. Check quotas and rate limits, and implement retries with exponential backoff and capacity planning
  3. Add deployments to the same resource in the same region to increase total quota
  4. Reduce max_tokens to shorten responses and adjust until 429 stops appearing
Show answer and explanation

Answer: B Check quotas and rate limits, and implement retries with exponential backoff and capacity planning

429 means the rate limit was exceeded. Space out retries and, if needed, review capacity (quota, deployment type). [Other options] Immediate resends make 429 worse. max_tokens affects TPM consumption but is insufficient without checking the cause. Adding deployments in the same resource does not increase total quota.

Q2. You are designing Azure infrastructure for an app using a Foundry project. Which is the most appropriate basic configuration for production?

  1. Managed identity, private endpoints and role assignments
  2. An API key from Key Vault, public access and the Contributor role
  3. Resources shared between production and development, key authentication and the Reader role
  4. A service principal secret, IP restrictions and the Owner role
Show answer and explanation

Answer: A Managed identity, private endpoints and role assignments

Base production on keyless authentication (managed identity), isolation (private endpoints) and least-privilege role assignments. [Other options] Keys and secrets in Key Vault leave more leak risk than keyless. Contributor and Owner grant excessive permissions for inference, and sharing production with development violates separation.

Reference links

1.3Manage, monitor, and secure AI systems

Concept

Operating AI systems combines access management (RBAC, managed identities), network protection (private endpoints), data protection (customer-managed keys), usage and cost monitoring (Azure Monitor, token counts) and tracing to make behavior visible.

How it works / steps

  1. Assign least-privilege RBAC and disable key authentication (disableLocalAuth)
  2. Block public network access with private endpoints
  3. Send logs and metrics to Log Analytics with diagnostic settings
  4. Set alerts on token usage, 429 errors and latency
  5. Enable AI threat protection in Microsoft Defender for Cloud

Workplace example

One department’s usage spikes and causes frequent 429s; an Azure Monitor alert detects it, deployment quotas are reviewed and per-department token limits are set in API Management.

Common confusions and exam tips

Mini quiz

Q1. In an internal document Q&A app, answers have started drifting away from the source documents. Which metric is most appropriate to monitor continuously?

  1. The number of content filter detections
  2. Token usage and daily inference cost
  3. Quality metrics such as groundedness
  4. Trends in response latency
Show answer and explanation

Answer: C Quality metrics such as groundedness

Declining groundedness often indicates degradation of the search index or data ingestion. [Other options] Latency, cost and filter detections matter operationally but do not show that answers are departing from their grounding.

Q2. What is the most appropriate governance when an agent has a payment execution tool?

  1. Write the payment limit in the prompt and have it reconfirm when exceeded
  2. Restrict the tool’s use and require human approval before execution
  3. Instruct in the system message to “judge large payments carefully”
  4. Log the payment tool’s results and review them together monthly
Show answer and explanation

Answer: B Restrict the tool’s use and require human approval before execution

Control high-impact tools with access control and approval. [Other options] Instructions or limits in prompts are not reliable controls. Monthly after-the-fact reviews cannot prevent incorrect payments.

Reference links

1.4Implement responsible AI across generative AI and agentic systems

Concept

Make Microsoft’s responsible AI principles (fairness, reliability and safety, privacy and security, inclusiveness, transparency, accountability) concrete with content filters, Prompt Shields (detecting direct and indirect prompt attacks), groundedness detection, safety evaluations and human review.

How it works / steps

  1. Identify risks (harmful output, hallucination, prompt injection, information leaks)
  2. Configure content filter severity thresholds and Prompt Shields
  3. Add safety instructions (a safety metaprompt) to the system message
  4. Measure with safety evaluations and red teaming
  5. Tell users they are dealing with AI and its limits, and require human approval for high-impact operations

Workplace example

In a public-facing chat, Prompt Shields detect instructions embedded in referenced web pages (indirect attacks), and the agent’s email sending requires approval by staff.

Common confusions and exam tips

Mini quiz

Q1. To prevent harmful output, what should you configure first for model inputs and outputs?

  1. Write “do not output harmful content” in the system message
  2. Lower temperature to reduce output variation
  3. Limit the maximum number of response tokens
  4. Content filters (safety filters) and guardrails
Show answer and explanation

Answer: D Content filters (safety filters) and guardrails

Safety filters and guardrails on inputs and outputs are the basic responsible AI measures. [Other options] Instructions in the system message alone cannot reliably prevent harmful output. Temperature and maximum tokens are not safety controls.

Q2. You have finished verifying an agent’s behavior in the development environment. What is the most appropriate next step before production deployment?

  1. Use the development environment’s API key and endpoint in production too
  2. Release to production users first and fix problems as they come up
  3. Delete development traces and leave production logs empty
  4. Evaluate quality and safety with test data and confirm the thresholds
Show answer and explanation

Answer: D Evaluate quality and safety with test data and confirm the thresholds

Before production, evaluate quality (groundedness, relevance and so on) and safety, and confirm the acceptance criteria are met. [Other options] Reusing development keys violates environment separation. Deleting traces loses evaluation material. Releasing before evaluation leaves quality and safety unverified.

Reference links

Lessons are written by this site based on public information and do not guarantee exam content. Product specifications change, so check the latest information at the reference links.

Notes (unofficial, original questions, no pass guarantee)