Domain 1 · Plan and manage an Azure AI solution
Every AI project starts with a set of choices that are hard to reverse later: which model, which service, which retrieval strategy, and what the surrounding infrastructure looks like. This domain covers those decisions and the operational discipline that follows them. It is the second-largest domain on the exam, and its questions are consistently comparative — you will be given a workload and asked which option fits, rather than being asked to recall a definition.
Choosing Foundry services, models, and retrieval
Model choice starts from the shape of the task. LLMs suit broad reasoning and open-ended generation. Small language models (SLMs) trade some capability for materially lower cost and latency, which makes them the right answer for high-volume, narrow, or edge scenarios. Multimodal models are required whenever the input includes images or audio rather than text alone. Alongside these sit Foundry Tools, the capabilities for grounding, vision, speech, translation, and content understanding that you compose with a model rather than replace it with.
Service choice follows the same logic one level up: generative tasks, grounding, vector search, agent workflows, and multimodal processing each route to the appropriate Foundry capability, with Azure AI Search carrying retrieval and indexing through its vector, hybrid, and semantic modes. When a scenario emphasises finding the right documents rather than generating text, the answer usually lives in Search.
For agent integration, the decision is which memory, tools, and knowledge the agent needs to meet its goal — function-calling for actions, knowledge stores and search for grounding, custom functions for anything proprietary. An agent given the wrong integration fails in ways that look like model quality problems but are not.
Setting up AI solutions in Foundry
Design begins with the Azure infrastructure that will host the application and its agents, and with the deployment option for each model — serverless or standard model deployments, chosen against the workload's traffic profile and isolation requirements.
Model and agent deployments are then configured inside a Foundry project, which is the unit of organisation the rest of the platform hangs off. Integrating that project with CI/CD pipelines is what turns a working prototype into a repeatable release process, and it is the reason projects, rather than individual deployments, are the thing you point automation at.
Manage, monitor, and secure
Once a solution is live, three concerns run in parallel.
Capacity and cost management means handling quotas, scaling, and rate limits for model and agent workloads deliberately rather than discovering them under load — a throttled model endpoint presents to users as an outage.
Monitoring in an AI system is broader than in a conventional one. On the model side you watch performance, drift, safety events, and grounding quality; on the retrieval side you watch data-ingestion quality, search-index health, and relevance. Degraded answers frequently originate in the index rather than the model, which is exactly why both halves are instrumented.
Security consistently favours the keyless path: managed identity, private networking, keyless credentials, and RBAC. Where an option offers a stored key and another offers an identity, the identity is the intended answer.
Responsible AI
Responsible AI is treated here as engineering work with concrete controls, not as a statement of intent.
At the request boundary you configure safety filters, guardrails, risk detection, and content moderation through content filters. Before and during release you apply evaluators and safety evaluations to measure behaviour, and instrument explanation tooling so that outputs can be interpreted rather than merely observed.
Accountability comes from auditing — trace logging, provenance metadata, and approval workflows that record what the system did and on what basis. Agent governance extends this to autonomy itself, through oversight modes, explicit constraints, and tool-access controls that bound what an agent is permitted to do without a human in the loop.
Quiz · Domain 1
A product team needs an agent that reasons over multi-step planning tasks while keeping per-token cost low for high request volume. They are comparing models in the Foundry model catalog. Which catalog feature should they use to objectively compare candidate models on quality and task performance before deploying?