Microsoft AB‑100 Agentic AI Architect Exam Prep & Scenarios

In this lecture, we'll map out the full Microsoft AI landscape, compare the key platforms, and show how they fit together for enterprise agentic solutions.

We zoom in on the different agent types supported in these platforms—this is critical for AB-100 design questions. You'll see "Which agent type for this scenario?" repeatedly on the exam.

Copilot Studio is the heart of no-code agent building in Microsoft. This lecture covers its core concepts—topics, actions, connectors, and knowledge—essential for AB-100 design questions.

explore Azure AI Foundry—the powerhouse for custom models and advanced agent orchestration. This is key for AB-100 scenarios requiring scale, domain-specific intelligence, or tight governance.

As agentic solutions get more complex, open standards become essential. This lecture covers MCP and Agent2Agent—key for AB-100 questions on extensibility and multi-agent design. These standards matter for governance and scalability.

When and how to propose custom models in Azure AI Foundry.
This directly maps to AB-100's critical "build vs. buy vs. extend" decisions and the exam subtopic on proposing Foundry tools for requirements.

Learn the first step in the Plan domain: identifying business goals, assessing constraints, determining whether agentic AI is the right fit vs. traditional automation, and applying a risk assessment framework with likelihood, impact, and mitigation strategies

Assess whether your organization's data is ready to ground AI agents. Covers the three dimensions of data readiness — quality, relevance, and availability — along with warning signs, recommended fixes, and Microsoft grounding sources like Copilot Studio Knowledge, Azure AI Foundry, and Power Platform connectors.

Walk through Microsoft's Cloud Adoption Framework phases (Strategy, Plan, Ready, Adopt, Govern) applied to agentic AI. Covers building an AI Center of Excellence, designing structured pilots with success metrics and go/no-go criteria, and scaling responsibly.

Apply a practical decision framework for choosing between building custom agents, buying prebuilt solutions like Copilot for Sales, or extending vendor platforms with custom topics and connectors. Includes trade-off analysis across cost, speed, control, and maintenance, with real-world scenarios.

Quantify the business case for agentic solutions. Covers TCO cost categories (development, model/inference, infrastructure, governance, change management), mitigation strategies for each, and a worked ROI calculation example for customer support automation.

Master prompt structure techniques (specificity, chain-of-thought, few-shot examples), behavior controls (role prompting, guardrails), and how to build a governed prompt library with ownership, version control, review processes, and cross-agent consistency.

Learn when dynamic model routing is sufficient vs. when custom or fine-tuned models are justified. Covers routing by query type (simple → small models, complex → premium), escalation triggers (domain accuracy gaps, compliance isolation, high-volume ROI), and the custom model decision process.

Architect data for multi-agent and cross-system use. Covers semantic layers, data lakes with vector search, connectors and APIs for standardized access, governance with Microsoft Purview, and a recommended four-layer architecture (Sources → Catalog → Common Platform → Agent Access).

Identify when to use prebuilt agents (Copilot for Sales, Service, M365 Copilot) vs. customized small language models like Phi-3. Covers SLM use cases (edge devices, high-volume tasks, privacy-sensitive scenarios) and the exam decision rule: start prebuilt, move to SLMs for cost/privacy/edge, custom large models only as last resort.

Explore the six core components every agentic system requires — perception, reasoning, memory, tools, execution runtime, and guardrails — how they map to Microsoft platforms, and how to choose the right level of agent autonomy (copilot, semi-autonomous, fully autonomous) based on business risk.

Master the five orchestration patterns (sequential, parallel, hierarchical, dynamic, event-driven) for coordinating multi-agent workflows. Covers agent-to-agent communication approaches (A2A, MCP-wrapped tools, shared Dataverse), and reliability patterns like circuit breakers, retry with backoff, timeouts, and fallbacks.

Build agents using Copilot Studio's three core elements — topics, tools, and flows — and understand the three orchestration methods the exam tests by name: Classic NLU, CLU, and Generative. Covers prompt actions for LLM-powered transformations and the three agent types: prompt-and-response, task, and autonomous.

Design with prebuilt Dynamics 365 Copilots. Covers Copilot for Sales capabilities (lead scoring, meeting prep, email drafting, win/loss analysis), Copilot for Service capabilities (case summarization, resolution suggestions, KB authoring, intelligent routing with Contact Center), key business terms (pipeline, opportunity, SLA, CSAT, FCR), and the four-level extension model.

Design AI solutions across Dynamics 365 Finance (invoice matching, anomaly detection, compliance checking) and Supply Chain Management (stockout prediction, quality flagging, logistics optimization). Covers cross-module orchestration through Dataverse and extending F&O agent chat with custom knowledge sources.

Extend Microsoft 365 Copilot using the four extensibility options: Copilot Studio plugins, declarative agents (JSON manifests), Graph connectors, and message extensions. Covers declarative agent design, agent discovery and admin controls, and M365 governance through sensitivity labels, DLP policies, and audit logging.

Prevent hallucination through grounding and RAG (Retrieval-Augmented Generation). Covers knowledge source selection (Dataverse, SharePoint, Azure AI Search, Graph, MCP), Copilot Studio knowledge configuration, data processing techniques (chunking, embedding, hybrid search), freshness management, and designing for citation and verification.

Master advanced extensibility mechanisms. Covers MCP for secure, auditable tool access (plus REST API tools and Skills as alternatives), UI Automation and RPA for legacy UI automation with strict guardrails, and configurable agent behaviors: reasoning mode, voice mode, autonomous triggers, and generative answers.

Decide when custom models are justified versus prebuilt or fine-tuned alternatives. Covers the Azure AI Foundry model catalog and tools (prompt flow, evaluation, fine-tuning, deployment), the seven-step custom model design workflow, and integration patterns including model routing, multi-model pipelines, and governance.

Decide when custom models are justified versus prebuilt or fine-tuned alternatives. Covers the Azure AI Foundry model catalog and tools (prompt flow, evaluation, fine-tuning, deployment), the seven-step custom model design workflow, and integration patterns including model routing, multi-model pipelines, and governance.

Understand why probabilistic agents require fundamentally different monitoring than traditional software. Covers the five KPI dimensions (reliability, performance, cost, quality, governance), the monitoring closed loop (monitor → alert → investigate → tune → validate), and real-world scenarios showing how unmonitored agents drift, spike costs, and create compliance gaps.

Build a layered testing strategy for agents that produce non-deterministic output. Covers the five-layer testing pyramid (unit, integration, end-to-end, adversarial, human review), why each layer catches different failure types, and practical scenarios for supply chain and customer service agents.

Apply application lifecycle management practices to agentic systems. Covers versioning agents and models, change tracking and approval gates, environment promotion (dev → test → prod) using managed solutions in Copilot Studio and model registries in Foundry, rollback strategies for when deployments go wrong, and data lineage tracking with Microsoft Purview including quality gates that prevent bad data from reaching production.

Implement the ethical backbone of enterprise AI. Covers Microsoft's six Responsible AI principles, governance frameworks with impact assessments, real-world scenarios of bias in hiring agents and unexplainable loan decisions, and how to design fairness thresholds, transparency mechanisms, and human oversight into every agent.

Defend agents against real-world attack vectors including prompt injection, data extraction, and overprivileged access. Covers input validation and sanitization, least-privilege design, row-level security for grounding data, MCP credential management, DLP policies, and building audit trails that satisfy regulatory requirements.

Navigate regulatory frameworks (GDPR, CCPA, HIPAA, SOC 2) that govern where and how agent data is processed. Covers data classification, residency region selection in Azure, right-to-explanation obligations, compliance documentation, and real-world scenarios of residency violations and regulatory consequences.

Turn raw telemetry into actionable insights. Covers root-cause analysis techniques for escalation rate spikes, cost anomalies, and quality drift, with real-world scenarios showing how to diagnose stale grounding data, prompt bloat, and model degradation — and how to apply data-driven tuning decisions.

Treat deployment as the starting line, not the finish. Covers feedback loop infrastructure (user ratings, escalation logging, interaction analysis), iterative tuning through prompt engineering, grounding data updates, and connector integration, plus using Copilot to accelerate tuning cycles. Includes a CSAT transformation case study.

Validate custom models before production using accuracy, fairness, robustness, drift detection, and adversarial testing. Covers Foundry evaluation tools, end-to-end testing across multi-Dynamics 365 workflows, and generating comprehensive test case suites with Copilot to reduce QA effort.

Exam-day tactics: question triage, time allocation across 40-60 questions, flagging and returning to hard items. How to handle preview/beta features on the exam (they're scored but marked). Localization note for non-English test-takers (extra time may be available). Builds on Section 1 orientation with specifics now that students have the full knowledge base.

One complex, realistic scenario taken end to end across all three exam domains. Walk through requirement analysis, platform and architecture decisions, then governance, security, testing, and monitoring -- showing how exam questions chain across skill areas. Reinforces cross-domain reasoning that individual section lectures taught in isolation.

Video walkthrough of 6 representative exam-style questions. Display each question on screen, reason through -- eliminate wrong answers, explain why distractors are tempting, and show the decision pattern that leads to the correct choice. Teaches the thinking process, not just content.