Quick Navigation
Cloud Adoption Framework for AI — Adoption ProcessAgent Types and Design Patterns — Copilot StudioA2A Protocol vs. MCP — Integration ProtocolsMicrosoft Foundry — Model Catalog and Model RouterConnector Types — Knowledge vs. Real-Time DataBuild vs. Buy vs. Extend — AI Solution SelectionDynamics 365 AI Features — Orchestration by ModulePower Platform Well-Architected Framework — Intelligent WorkloadsApplication Lifecycle Management (ALM) — By Solution TypeMonitoring, Telemetry, and TestingResponsible AI — Six Principles and Review ProcessSecurity Design — Agents, Models, and Data
Cloud Adoption Framework for AI — Adoption Process
- CAF AI phases: AI Strategy → AI Plan → AI Ready → Govern AI → Manage AI → Secure AI
- The six-phase AI adoption process from the Cloud Adoption Framework for Azure — implement this framework first when designing enterprise AI strategy, not the general CAF.
- AI Strategy phase
- Define business motivations, expected outcomes, and a portfolio of AI use cases aligned to organizational priorities before selecting any technology.
- AI Ready phase
- Establish technical foundations: Azure landing zones, data platforms, networking, identity, and tool access required to host AI workloads securely.
- Govern AI phase
- Define policies, accountability structures, and compliance controls for AI systems — includes responsible AI standards, data residency rules, and AI Center of Excellence governance.
- AI Center of Excellence (AI CoE)
- Organizational body that governs AI strategy, standardizes tooling and practices, manages the AI portfolio, and drives adoption across business units.
- CAF for AI vs. general CAF
- The AI-specific CAF process (Strategy, Plan, Ready, Govern, Manage, Secure) is distinct from the general CAF migration process — do not confuse or combine them on the exam.
Agent Types and Design Patterns — Copilot Studio
- Three agent types: Task Agent | Autonomous Agent | Prompt-and-Response Agent
- Task agents follow predefined workflows; autonomous agents independently plan and execute multi-step tasks with reasoning; prompt-and-response agents answer questions conversationally.
- Task agent
- Executes specific, well-defined workflows triggered explicitly — deterministic, narrow in scope, requires human-initiated invocation for each task.
- Autonomous agent
- Independently makes decisions, plans multi-step workflows, and acts without constant human intervention — requires stronger governance, security controls, and monitoring than task agents.
- Prompt-and-response agent
- Conversational agent that answers questions using knowledge sources — best for Q&A, information retrieval, and guided conversation scenarios.
- Topics and Fallback Topics in Copilot Studio
- Topics define structured conversational flows triggered by user intent; fallback topics handle unmatched inputs when no topic matches — always design a fallback topic to prevent dead ends.
- Orchestration modes: Standard NLP | Azure CLU | Generative AI Orchestration
- Standard NLP is fastest/cheapest for predictable intents; Azure Conversational Language Understanding handles custom entity extraction; Generative AI Orchestration uses LLMs for open-ended context-aware conversations.
- Copilot Studio agent behaviors: Reasoning Mode | Voice Mode
- Reasoning mode enables complex multi-step problem solving; Voice mode enables speech-based interaction — these are configurable agent behaviors, not separate agent types.
A2A Protocol vs. MCP — Integration Protocols
- A2A: Agent-to-Agent Protocol — connects AGENTS to AGENTS
- Open protocol for structured agent-to-agent communication: agents exchange messages, delegate subtasks, and coordinate across systems using shared organizational context.
- MCP: Model Context Protocol — connects AGENTS to TOOLS
- Standardized protocol for connecting AI agents to external tools, services, databases, and enterprise resources with consistent security, authentication, and auditing.
- A2A vs. MCP decision rule
- Use A2A when one agent needs to delegate tasks to or communicate with another agent; use MCP when an agent needs to invoke an external tool, API, or data source.
- A2A GA in Copilot Studio (April 2026)
- Agent-to-Agent protocol is generally available in Copilot Studio, enabling first-party, second-party, and third-party agent coordination — use the A2A connector for Power Automate integration.
- Remote MCP server integration in Copilot Studio
- Connect agents to remote MCP servers to expose tools and data sources — recommended over custom connectors when the external system supports the MCP standard.
- Computer Use agents
- Agents that automate tasks by interacting directly with application UIs (mouse, keyboard simulation) — use when the target system has no API or MCP interface available.
Microsoft Foundry — Model Catalog and Model Router
- Microsoft Foundry model catalog: 1800+ models
- Curated catalog of LLMs from OpenAI, Anthropic, Meta, Mistral, DeepSeek, Grok, and Phi families — select models based on task requirements, cost, latency, and compliance.
- Model Router routing modes: Balanced | Cost | Quality
- Balanced (default): dynamically picks the most cost-effective model within a small quality band (~1-2%); Cost: aggressively favors cheaper models accepting larger quality trade-off; Quality: always selects highest-quality model regardless of cost.
- Model Router — Balanced mode (default)
- Selects the most cost-effective model within a small quality tolerance — the right choice when you want automatic cost optimization without significantly compromising response quality.
- Model Router — Cost mode
- Aggressively routes to cheaper models, accepting a larger quality band — appropriate when cost minimization is the primary constraint and quality variation is acceptable.
- Model Router — Quality mode
- Always selects the highest-quality model for every prompt regardless of cost — appropriate for high-stakes decisions, compliance-critical responses, or when quality must never be compromised.
- Foundry Agent Service vs. Copilot Studio
- Foundry Agent Service is code-first (SDK, custom models, full control over logic); Copilot Studio is low-code (topics, actions, connectors) for business users and citizen developers.
- Foundry Tools: fine-tuning, evaluation, deployment, monitoring
- Foundry Tools provide the development pipeline for custom model creation — define solution rules and constraints when using Foundry Tools for model training and deployment.
- Small Language Models (SLMs) — use cases
- Prefer SLMs (Phi family) over LLMs when the task is domain-specific, latency is critical, or cost efficiency is required — SLMs can be fine-tuned for specialized vocabulary and tasks.
Connector Types — Knowledge vs. Real-Time Data
- Copilot connectors — indexed knowledge for grounded Q&A
- Index enterprise content (SharePoint, websites, structured data) for search-based retrieval at scale — use when the agent needs to answer questions across a large knowledge corpus.
- Power Platform connectors — real-time transactional data
- Connect to live business systems (Dynamics 365, SAP, Salesforce) for real-time data reads and writes — use when the agent needs current inventory, pricing, or order status.
- Copilot connectors vs. Power Platform connectors — decision rule
- Copilot connectors: search and Q&A at scale over indexed content; Power Platform connectors: live transactional queries against business systems — never use them interchangeably.
- Data grounding quality factors: accuracy, relevance, timeliness, cleanliness, availability
- All five quality dimensions must be assessed before designing an AI solution — poor grounding data quality directly degrades agent reliability regardless of model capability.
- Prompt Library
- Organizational collection of curated, tested, and approved prompts for standardizing AI interactions — provide guidelines for prompt library creation to ensure consistent, effective agent behavior.
Build vs. Buy vs. Extend — AI Solution Selection
- Decision order: Extend existing → Buy prebuilt → Build custom
- Always evaluate whether existing Copilot features or prebuilt agents meet requirements before recommending custom development — building custom AI adds cost, time, and risk.
- Extend existing: configure Dynamics 365 Copilot or Microsoft 365 Copilot
- When standard capabilities meet business requirements, extend and configure prebuilt Copilot features — fastest time to value and lowest total cost of ownership.
- Buy prebuilt: use prebuilt agents from Microsoft catalog
- Microsoft provides ready-to-use agents across Microsoft 365, Dynamics 365, and Power Platform for common business scenarios — configure and orchestrate before considering custom builds.
- Build custom: Microsoft Foundry custom models or Copilot Studio agents
- Custom development is justified only when no prebuilt solution meets unique requirements, proprietary data demands specialized training, or competitive differentiation requires it.
- ROI analysis: TCO, measurable business impact, cost-benefit evaluation
- Always justify AI solution decisions with ROI analysis — total cost of ownership includes model compute, data infrastructure, ALM overhead, monitoring, and governance costs.
- Microsoft 365 Copilot extension vs. custom Copilot Studio agent
- Extend M365 Copilot (declarative agents, plugins) when users already work in Microsoft 365 and need domain-specific help within Teams/Word/SharePoint; build custom agents when you need full deployment control or non-M365 channels.
Dynamics 365 AI Features — Orchestration by Module
- Dynamics 365 Finance — AI features
- Copilot in Finance provides cash flow forecasting, customer payment predictions, vendor invoice automation, and intelligent budget proposals — extend with additional knowledge sources via agent chats.
- Dynamics 365 Supply Chain Management — AI features
- Copilot in SCM provides demand forecasting, supplier risk assessment, inventory optimization, and agent chats that can be extended with additional knowledge for in-app guidance.
- Dynamics 365 Customer Service — Copilot features
- AI drafts responses, summarizes cases, provides real-time agent assistance, and integrates with Contact Center channels — orchestrate Copilot for Service configuration for consistent CX.
- Copilot for Sales vs. Copilot for Service
- Copilot for Sales: role-specific AI in Dynamics 365 Sales and Outlook/Teams for lead insights, meeting summaries, and pipeline guidance; Copilot for Service: AI for case resolution, agent assist, and contact center operations.
- Dynamics 365 Contact Center — agent integration
- Supports multi-channel agent deployment (voice, chat, email, social) — design agents specifically for Contact Center channel integration, not generic Copilot Studio agents.
- Business terms for Dynamics 365 customer experience
- Configure domain-specific business terms to improve Copilot's understanding of industry vocabulary in Customer Service and Sales apps — required for accurate, contextually relevant responses.
- Finance and Operations agent chats
- In-app agent chat experiences in Dynamics 365 Finance and Supply Chain that can be extended with additional knowledge sources beyond the out-of-the-box capabilities.
Power Platform Well-Architected Framework — Intelligent Workloads
- Five pillars: Reliability | Security | Operational Excellence | Performance Efficiency | Experience Optimization
- Apply all five pillars specifically to intelligent application workloads — AI-specific considerations within each pillar differ from general app workload guidance.
- Reliability pillar — intelligent workloads
- Design for agent failover, model fallback, and graceful degradation when AI services are unavailable — LLM outages must not cause complete application failure.
- Security pillar — intelligent workloads
- Protect AI inputs (prompt injection defenses), outputs (content filtering), grounding data (access controls), and model endpoints (network isolation) — AI introduces attack surfaces not present in traditional apps.
- Operational Excellence pillar — intelligent workloads
- Define ALM processes, deployment pipelines, monitoring strategies, and telemetry for each AI component — agents, models, and D365 AI features each require separate operational processes.
- Performance Efficiency pillar — intelligent workloads
- Select appropriate model size, use model router for cost/quality optimization, apply SLMs where suitable, and configure caching for repeated prompts to reduce latency and token costs.
- Experience Optimization pillar — intelligent workloads
- Design agent personas, conversation flows, and response formats that align with user expectations — measure user satisfaction alongside technical performance metrics.
Application Lifecycle Management (ALM) — By Solution Type
- Three separate ALM processes required: Copilot Studio | Foundry | Dynamics 365 AI
- Each solution type has distinct deployment mechanisms, versioning approaches, and environment strategies — design separate ALM processes, never a single unified process.
- ALM for Copilot Studio agents
- Package agents, topics, connectors, and actions as Power Platform solutions; deploy across dev/test/prod environments using Pipelines in Power Platform with environment variables for configuration.
- ALM for Microsoft Foundry agents and custom models
- Use model versioning, staged deployment (shadow/canary/full rollout), and rollback capabilities in Foundry — custom model deployment requires separate CI/CD pipelines from Copilot Studio agents.
- ALM for Dynamics 365 AI features
- Manage AI configuration (Copilot settings, business terms, agent chats) as part of D365 solution packages deployed through standard D365 ALM — separate from both Copilot Studio and Foundry ALM.
- Environment strategy: Dev → Test → Prod
- Always use separate environments for development, testing, and production — environment variables in Power Platform solutions allow configuration differences across environments without code changes.
- AI Hub in Power Platform
- Centralized hub for discovering and managing AI capabilities, models, and prompts available across Power Platform — propose AI Hub for organizations consolidating their AI component inventory.
Monitoring, Telemetry, and Testing
- Agent performance metrics: resolution rate, escalation rate, CSAT, topic coverage, response latency
- Monitor these key metrics continuously — resolution rate measures how often agents resolve queries without human escalation; CSAT reflects user satisfaction with agent responses.
- Telemetry interpretation vs. monitoring
- Monitoring collects data; telemetry interpretation is the analysis step that identifies root causes and informs tuning decisions — always interpret telemetry before taking corrective action.
- Agent backlog analysis
- Review unresolved and escalated conversations to identify missing topics, poor intent matching, or knowledge gaps — backlog analysis drives topic expansion and knowledge source updates.
- Validation criteria for custom AI models
- Define accuracy thresholds, precision/recall targets, bias checks, and responsible AI compliance tests before a custom model can be promoted to production.
- End-to-end test scenarios for multi-app solutions
- Design test scenarios that span multiple Dynamics 365 apps and agent interactions — unit tests per component are insufficient; integration and end-to-end scenarios must validate cross-system flows.
- Test case creation with Copilot (AI-assisted testing)
- Use Copilot to generate test cases from agent conversation logs and business requirements — AI-generated test coverage supplements but does not replace manually designed critical-path tests.
Responsible AI — Six Principles and Review Process
- Six Microsoft Responsible AI principles: Fairness | Reliability & Safety | Privacy & Security | Inclusiveness | Transparency | Accountability
- All six principles must be actively reviewed in every AI solution — treating responsible AI as optional is an exam trap; the review is required, not aspirational.
- Fairness
- AI systems must treat all people equitably — test for bias across demographic groups, ensure training data is representative, and monitor for disparate impact in production.
- Reliability and Safety
- AI systems must perform reliably across conditions and fail safely — design for graceful degradation, fallback behaviors, and human escalation paths for high-stakes decisions.
- Privacy and Security
- Protect user data throughout the AI lifecycle — apply data minimization, access controls, encryption, and audit trails; AI introduces new attack surfaces (prompt injection, data leakage).
- Transparency
- Users must understand that they are interacting with AI and how decisions are made — disclose AI involvement, provide explanations for AI-driven decisions, and avoid deceptive personas.
- Accountability
- Humans must remain accountable for AI system behavior — design audit trails, human-in-the-loop checkpoints for high-risk decisions, and clear ownership for AI governance.
Security Design — Agents, Models, and Data
- Prompt injection / prompt manipulation attacks
- Malicious user input attempts to override agent instructions or extract unauthorized data — defend with input validation, output filtering, instruction isolation, and ongoing adversarial testing.
- Jailbreaking mitigation
- Prevent users from bypassing safety instructions by implementing system prompt hardening, content filters, and behavioral monitoring for unusual agent output patterns.
- Access controls on grounding data
- Apply role-based access controls to knowledge sources and indexed content — agents must only surface data the authenticated user is authorized to see.
- Data residency vs. data access controls
- Data residency = WHERE data is stored and processed (geographic/sovereign compliance); data access controls = WHO can access data (RBAC, authentication) — these are separate compliance requirements.
- Data residency and movement compliance
- Validate that AI-processed data remains within required geographic boundaries — configure Azure region selection for all AI services (Foundry, Azure OpenAI, AI Search) explicitly.
- Audit trails for AI models and data
- Design audit trails capturing who changed which model or data, when, and why — required for regulated industries and Responsible AI accountability principle compliance.
- Governance framework for agents
- Define agent scope boundaries, escalation thresholds, human-in-the-loop requirements, and monitoring policies — autonomous agents require stricter governance than task agents.