When out-of-box isn't enough
Microsoft Foundry lets organizations build, fine-tune, evaluate, deploy, and operate custom models when prebuilt copilots fall short. Reach for a custom model for domain-specific language, high-impact decisions, data-sovereignty mandates, unique workflows, or cost optimization at high inference volume (a small specialized model can beat a large one on both accuracy and cost).
Define business objectives
Measurable targets — accuracy, time saved, cost efficiency — mapped to use cases where a custom model beats a standard copilot.
Assess data requirements
Evaluate proprietary datasets; find gaps in labeling, quality, diversity, structure; confirm governance permits training use.
Select the custom-model path
Fine-tune a foundation model, train a domain-built small model, or use a hybrid (custom model + prebuilt copilot).
Integrate with enterprise systems
Dynamics 365, Azure Functions/Logic Apps, Foundry agent workflows, Azure AI Search, automation pipelines.
Validate & evaluate
Scenario evaluations, safety/bias analysis, latency & scale tests, and ROI/business validation.
Key concept · Standard copilot vs custom model
Lean custom (Foundry) as domain specificity, compliance restrictions, performance needs, data confidentiality, and workflow complexity rise — and for inference-cost optimization via small language models. Operationalize with monitoring/observability, governance & compliance controls, versioning & lifecycle management, and MLOps/GenAIOps deployment automation.
Four layers of agent extensibility
In Copilot Studio, extensibility operates at four layers — start at the cheapest that solves the problem and climb only as needed.
| Layer | Function | How |
|---|---|---|
| Instruction-level | Behavior, tone, boundaries, reasoning authority | Prompt modification — add rules without rebuilding |
| Skill / capability | What the agent can do | Retrieval skills, Power Platform connectors, custom APIs |
| Integration | Connect to systems of record | Power Automate flows, external APIs, event publishing |
| Pro-code | Advanced custom logic & orchestration | Visual Studio Code, source control, custom tools |
Agents in Microsoft 365 Copilot are built with Agent Builder (Define → Add instructions/guardrails → Connect data → Configure actions → Test → Publish) and run inside Teams, Outlook, Loop, SharePoint, and line-of-business apps. Each agent should solve a single high-value need, operate under least-privilege / zero-trust, and — when collaborative — follow explicit handoff rules across patterns like sequential workflow, parallel evaluation, feedback-loop iteration, and orchestrated interaction.
Key concept · Architecture guidance — three pillars
Microsoft's agent architecture framework is organized around Fit for Purpose (designed for its use case), Operability (runs effectively in production), and Trust, Traceability & Transparency (reliable, auditable, explainable). It complements — not duplicates — the Azure and Power Platform Well-Architected Frameworks and NIST.
Standardized context and UI automation
The Model Context Protocol (MCP) is a standardized contract defining what context an agent can access, how to interpret it, and how it retrieves business context from enterprise systems. In Dynamics 365 Finance & Operations it exposes business entities (customers, vendors, products), relationships, metadata, and domain objects — giving agents consistent business semantics.
Key concept · Why MCP matters
- Ensures consistent business semantics across agents.
- Reduces incorrect information by grounding in real system context.
- Enables multi-app interoperability and shared enterprise logic.
- Improves explainability and governance of agent decisions.
- Accelerates extensibility by standardizing how agents consume context.
Computer Use lets an agent control mouse and keyboard, read on-screen elements with vision + reasoning, and perform multi-step tasks across apps and websites — the fallback for when no API or connector exists (legacy systems, vendor portals, desktop-only apps). Design discipline is essential: specific goal-oriented tasks, explicit allowed apps and prohibited actions, step-by-step workflows, descriptive element references ("click the blue Submit button"), validation after each step, and graceful error handling — because UI changes can break automation.
- MCP
- A governed contract for consistent, explainable access to enterprise context across agents.
- Computer Use
- Vision-and-reasoning UI automation for systems with no API — use only as a fallback, with strict guardrails.
Reasoning depth, voice, and grounding
Agent behavior is shaped by four components — instructions (purpose, persona, tone, guardrails), knowledge (content access), actions & tools (execution), and policies & constraints (compliance, privacy, responsible AI). Reasoning depth is a deliberate trade-off:
| Mode | Best for | Cost |
|---|---|---|
| Standard reasoning | Conversational responses, summaries, simple calculations, high volume | Fast, low compute |
| Deep reasoning (preview) | Multi-step tasks, complex rules, analysis, trade-offs — finance, compliance, engineering | Slower, higher precision |
Voice mode / IVR agents demand short, clear phrasing, confirmation steps, and design for interruptions and backtracking, following the flow: speak → interpret intent → confirm/clarify → execute → concise spoken output. Because voice is real-time, prioritize responsiveness, fallback handling, and friendly error messages.
Key concept · M365 grounding & permission inheritance
Agents in Teams and SharePoint ground on enterprise content through Microsoft Graph and inherit the user's permissions — they can never surface content the user couldn't access themselves. Optimize with well-structured, tagged SharePoint sources; concise contextual answers; version control; and AI Center-of-Excellence governance. Common patterns: SharePoint Knowledge Assistant, Teams Project Assistant, Organizational Policy Assistant, Site Owner Support Agent.
Module 6 quiz
Fifteen questions on custom models in Foundry, the four extensibility layers, MCP and Computer Use, and reasoning/voice/grounding behaviors. Answer first, then expand Show answer.
Part A · Multiple choice
Which scenario most justifies a custom model in Microsoft Foundry over a prebuilt copilot?
- A A generic FAQ chatbot
- B Domain-specific reasoning with strict data-sovereignty and compliance needs
- C A one-off demo
- D Sending a templated email
Show answer
B. Custom models suit high domain specificity, compliance restrictions, data confidentiality, and inference-cost optimization.
Which is the cheapest extensibility layer to modify an agent's tone and boundaries without rebuilding it?
- A Pro-code in Visual Studio Code
- B Integration via Power Automate
- C Instruction-level (prompt modification)
- D Skill/capability
Show answer
C — Instruction-level. Prompt modification adds custom rules and behaviors without rebuilding the agent.
What does the Model Context Protocol (MCP) primarily standardize?
- A Screen resolution for UI automation
- B What context an agent can access and how it retrieves/interprets enterprise business context
- C The pricing of tokens
- D The font used in adaptive cards
Show answer
B. MCP is a governed contract for consistent, explainable access to business entities, relationships, and metadata.
When is Computer Use the appropriate automation choice?
- A Whenever a clean REST API is available
- B When no API or connector exists — legacy systems, vendor portals, desktop-only apps
- C For high-frequency, latency-critical tasks
- D To reduce audit logging
Show answer
B. It's a fallback for UI-only systems; prefer APIs when they exist because UI changes can break automation.
Which is the correct order of the four extensibility layers, cheapest to most advanced?
- A Pro-code → Integration → Skill → Instruction
- B Instruction → Skill/capability → Integration → Pro-code
- C Skill → Instruction → Pro-code → Integration
- D Integration → Pro-code → Skill → Instruction
Show answer
B. Instruction-level, then skill/capability, then integration, then pro-code (Visual Studio Code).
Which are the three pillars of Microsoft's agent architecture guidance?
- A Speed, Cost, Scale
- B Fit for Purpose; Operability; Trust, Traceability & Transparency
- C Fairness, Privacy, Accountability
- D Ingest, Train, Serve
Show answer
B. They complement the Azure/Power Platform Well-Architected Frameworks and NIST.
Which tool is used to build agents in Microsoft 365 Copilot?
- A Agent Builder
- B Visual Studio Installer
- C Power BI Desktop
- D Azure Portal only
Show answer
A — Agent Builder. Define → add instructions/guardrails → connect data → configure actions → test → publish.
When should deep reasoning (preview) be enabled instead of standard reasoning?
- A For simple greetings and summaries
- B For multi-step tasks with complex rules, trade-offs, and high-precision logic
- C To lower compute cost on high-volume chat
- D Never
Show answer
B. Deep reasoning suits analytical, multi-stage work (finance, compliance, engineering); standard reasoning is fast and cheap for high volume.
Part B · True or false
A small, specialized custom model can outperform a large foundation model on both cost and performance for high-volume inference.
Show answer
True. Cost optimization at scale is an explicit reason to build custom SLMs.
Agents in Teams and SharePoint can surface content the current user does not have permission to see.
Show answer
False. Agents inherit user permissions via Microsoft Graph and never expose content the user can't access.
Voice-mode agents should use short, clear phrasing and design for interruptions and confirmation steps.
Show answer
True. Real-time voice prioritizes responsiveness, fallback handling, and friendly error messages.
Because Computer Use can reason about the UI, no guardrails or step design are needed.
Show answer
False. Define allowed apps, prohibited actions, clear steps, per-step validation, and error handling — UI drift breaks brittle automation.
Part C · Short answer
Name the four layers of agent extensibility.
Show answer
Instruction-level, skill/capability, integration, and pro-code.
Give three benefits MCP provides to enterprise agents.
Show answer
Any three of: consistent business semantics, reduced incorrect information via real grounding, multi-app interoperability, better explainability/governance, and faster extensibility.
List the four components that shape an agent's behavior in Copilot Studio.
Show answer
Instructions, knowledge, actions & tools, and policies & constraints.