ExamSprintLab

AI-103 plan and manage an Azure AI solution explained

Updated September 20, 2026

Plan and manage an Azure AI solution is worth 25–30% of your AI-103 score, and it is the domain candidates underestimate most. It contains very little coding. Instead it asks whether you can pick the right model and service for a job, deploy it sensibly, keep it secure and inside its budget, and show that it behaves responsibly. If you come from a developer background, this is usually where you lose points.

The four groups inside the domain

Choosing the right services and models

You are expected to match a requirement to a model class: a large language model, a small language model, a multimodal model, or one of the Foundry tools. The exam rarely asks “which model is best” in the abstract. It gives you a constraint — latency, cost, on-device deployment, image input, a long context — and wants the cheapest option that still meets it. The same logic applies to services: generation, grounding, vector search, agent workflows and multimodal processing each have a natural home, and you need to know which.

Retrieval and indexing choices sit here too. So does the question of which memory, tool and knowledge integrations an agent should get.

Setting a solution up in Foundry

This covers the infrastructure around the model: how a Foundry project is laid out, which deployment option fits (and what each one implies for throughput and cost), how model and agent deployments are configured, and how a Foundry project is wired into a CI/CD pipeline. Expect at least one question that treats an AI project like any other application deployment, because that is exactly how Microsoft wants you to treat it.

Managing, monitoring and securing

Three recurring themes:

  • Capacity and cost. Quotas, scaling, rate limits and the cost footprint of model and agent workloads. Know what happens when you hit a rate limit and what your options are.
  • Observability. Monitoring model performance, drift, safety events and grounding quality, plus the health of your ingestion pipeline and search index. Relevance is something you measure, not something you hope for.
  • Security. Managed identity, private networking, keyless credentials and role-based policies. The exam’s preferred answer is almost always the one with no secret in it.

Responsible AI

Safety filters, guardrails, risk detection and content moderation; evaluators and safety evaluations; auditing through trace logging, provenance metadata and approval workflows; and governing agent behaviour with oversight modes, constraints and tool-access controls. Responsible AI on AI-103 is operational, not philosophical — it is a set of switches you configure and evidence you can produce later.

How these questions are framed

Almost every question in this domain is a scenario with a constraint attached. A company forbids stored keys. A workload must not traverse the public internet. A budget cap must not be exceeded. A regulator will ask who approved an agent action. Read for the constraint first, then eliminate every option that violates it — that usually leaves one or two answers.

The other pattern is the “least amount of change” or “least administrative effort” wording. When you see it, a working-but-heavy answer is wrong on purpose.

Sample questions

Question 1. A customer support agent must summarise short tickets in under 300 milliseconds and will run at high volume. Quality requirements are modest. Which model choice best fits?

  • A. A large language model with a long context window
  • B. A small language model deployed for high throughput
  • C. A multimodal model so images can be added later
  • D. A fine-tuned large language model
Show answer

Answer: B

A small language model gives the lowest latency and cost per call, which is what the requirement is built around. A large or multimodal model adds capability the task does not need and makes the latency target much harder to hit. Fine-tuning is a separate decision and does not by itself solve latency.

Question 2. Your Foundry project must not be reachable over the public internet, and the application that calls it runs on Azure. Which combination meets the requirement?

  • A. A public endpoint with an IP allow-list and a rotated API key
  • B. A public endpoint with the key stored in a key vault
  • C. Private networking plus a managed identity with a role assignment
  • D. A shared access signature scoped to the calling application
Show answer

Answer: C

Private networking keeps traffic on the Azure backbone, and managed identity removes the stored credential. Firewall IP allow-lists still expose a public endpoint, and rotating keys or storing them in a vault leaves a key in the flow that private networking was meant to eliminate.

Question 3. After a model version upgrade, an internal team reports that answers have quietly drifted in tone and occasionally contradict the source documents. You need evidence of when this started. What should you have in place?

  • A. Monitoring of grounding quality and safety events with trace logging retained over time
  • B. A stricter content safety filter applied to all responses
  • C. A higher quota on the model deployment
  • D. A lower temperature setting on every request
Show answer

Answer: A

Continuous monitoring of grounding quality and safety events, backed by trace logging, gives you a timeline you can compare before and after the upgrade. Raising the content filter, increasing quota or lowering temperature might change behaviour but none of them tells you when the change began or why.

What to practise

Build one Foundry project and deliberately do the unglamorous parts: give it a managed identity instead of a key, lock the network down, set a quota and watch what happens when you exceed it, turn on tracing and read an actual trace. An afternoon of that is worth more than a week of reading, because the exam’s scenarios are written from the perspective of someone who has had to operate the thing.