Domain 3 · Deploy AI-powered business solutions
At 40–45%, this is the heaviest-weighted domain on the exam — close to half the questions, and more than Domains 1 and 2 carry between them. It is worth reading that weighting as a statement of intent: Microsoft is testing whether you can keep an agentic solution running, correct, and defensible after it ships, which is a harder problem than designing it. Prioritise this domain when your study time runs short.
Analyze, monitor, and tune
Monitoring an agent is partly a tooling question and partly a process one, and the objectives ask for both: recommending the process and tools for monitoring agents, and analyzing the backlog and user feedback that tell you what is going wrong in terms users actually experience.
From there you apply AI-based tools to identify issues and tune, monitor agent performance and metrics, and interpret telemetry to guide that tuning. The loop is deliberate — telemetry and feedback identify the problem, tuning addresses it, and monitoring confirms the fix — and questions often hinge on which signal you would consult first.
Manage testing
Testing agentic systems is harder than testing deterministic ones, which is why it gets its own section. You recommend the process and metrics for testing agents and create validation criteria for custom AI models — criteria being the crux, since "the answer looked good" does not scale.
Beyond that, you validate Copilot prompt best practices, design end-to-end test scenarios spanning multiple Dynamics 365 apps rather than testing each app in isolation, and build test cases using Copilot itself. That last one is a genuine objective, not an aside: generating test coverage with AI is part of the expected practice.
Design the ALM process
Application lifecycle management is examined across the full breadth of what an agentic solution contains, and the list is the point — each item has its own promotion path between environments:
- the data used by AI models and agents,
- Copilot Studio agents, connectors, and actions,
- the Microsoft Foundry Agents service,
- custom AI models, and
- AI in Dynamics 365 finance/supply-chain and customer-experience apps.
Data is the one most easily overlooked. Grounding data needs versioning and a controlled path to production exactly as code does, because promoting an agent without its data promotes a different agent.
Responsible AI, security, governance, risk, compliance
The final section is where an architect's answer differs most from a developer's.
Design covers agent security and governance and model security, including analyzing vulnerabilities and their mitigations — with prompt manipulation called out specifically as the attack class unique to these systems.
Compliance runs alongside: reviewing adherence to responsible AI principles, and validating data residency and movement compliance, which frequently constrains architecture more than any technical factor does.
Two concrete controls close the domain. Access controls on grounding data and model tuning determine who can influence what the system knows and how it behaves, and audit trails for model and data changes make that influence reconstructable after the fact. Together they are how you answer "who changed this, and when" — the question that arrives immediately after any incident.
Quiz · Domain 3
An architect must enable daily and monthly active user (DAU/MAU) metrics on the Analytics page for a customer-facing Copilot Studio agent. What configuration is a prerequisite for these active-user metrics to populate?