Sidecar, Embedded, and Outside
Copilot appears in Dynamics 365 Finance & Supply Chain in three orchestration shapes. The architect's job is to choose the right one per scenario and keep data governance, role-based access, and responsible AI consistent across modules.
Sidecar
Copilot alongside the app for natural-language chat — generative help, workflow summaries (journals, approvals, status), and chat with finance & operations data. Optimize prompts with domain vocabulary; apply role-based access; orchestrate cross-module queries via standardized entity metadata.
Embedded
Copilot inside workspace pages and operational views — PO change analysis, collections summaries, demand-planning insights, supplier drafting. Surfaces recommended actions where work happens; must use authoritative data entities and honor business rules.
Outside (external orchestration)
External agents act on F&SC data beyond the UI, orchestrating across apps and tasks — cross-app automation, role-specific copilots in Teams, workflow routing. Use Dataverse or custom APIs for governed access; honor data-residency and approval constraints.
Copilot for F&SC extends through custom scripts/extensions, prompt-defined sidecar behaviors, custom data sources, business-event triggers (Power Automate / Azure Functions), and custom actions Copilot can call — including client plugins built with X++ in the finance-and-operations context.
Proposing Microsoft 365 agents
Microsoft 365 agents retrieve, reason, and act across apps and data to achieve a specific outcome. Unlike ad-hoc prompts, a well-designed agent has a defined mission and scope, grounding data and tools, operational guardrails (identity, authorization, DLP, reviewability), and telemetered outcomes.
Key concept · Five-step design framework
- Frame the job to be done — the recurring outcome, not the tool steps.
- Map inputs, knowledge, and actions — sources, tools/connectors, required write operations.
- Define guardrails — who can invoke it, on what data, when it asks for approval, how actions are logged/reversed.
- Prototype the critical path — a narrow happy-path slice tested on real artifacts.
- Operationalize — telemetry, quality checks, cost controls; pilot → production via rollout rings.
A readiness checklist gates any proposal: business value (owner, users, measurable outcome, definition of done), identity & access (least-privilege runs-as model), data scope (enumerated, labeled corpus), actions & tools (failure paths, approval points), security & compliance (DLP, auditability, logging), change control (versioning, rollout rings, rollback), measurement (adoption, quality, latency, cost KPIs), and support (ownership, runbook, ethics review). Ready-to-pilot agents include an Executive Briefing Pack Generator, Portfolio Risk Insights Analyzer, Compliance-Aware Content Redactor, and a Telemetry-to-Insights Report Generator.
Key concept · RACI & starter KPIs
Assign every capability with a RACI matrix — Responsible (does the work), Accountable (one owner answerable), Consulted (two-way input), Informed (one-way updates). Pilot KPI targets: adoption ≥ 30% of the cohort, first-output acceptance ≥ 70%, median time saved ≥ 25%, policy violations ≤ 1 per 1,000 runs, and cost within baseline ± 10%.
Architecture-first configuration
Copilot for Sales and for Service both accelerate work through context retrieval, summarization, content generation, and workflow execution — but only after the data estate is ready. The recurring discipline is validate data → connect systems → map fields → set least-privilege access → enable actions → automate with Power Automate.
| Aspect | Copilot for Sales | Copilot for Service |
|---|---|---|
| Grounds on | CRM opportunity/account/contact/activity data, sales docs | Cases, transcripts, knowledge articles, SLAs, agent notes |
| Core loop | Email → summary → CRM update → proposal → seller review | Case intake → summary → knowledge retrieval → suggested actions → agent review |
| Sample KPIs | Less prep time; faster lead-to-opportunity; proposal consistency | Lower handle time; higher first-contact resolution; better article retrieval |
AI Builder augments both with classification, information extraction, predictive analytics, and multi-step reasoning; connectors handle cross-system approvals, notifications, and escalations. Guardrails are non-negotiable: sensitivity labels, DLP, audit trails, data residency, human review before external content, and telemetry monitoring for adoption and drift.
Choosing the right AI feature
Power Platform exposes four AI building blocks. The AI Hub is the central, governed workspace that discovers, manages, and orchestrates them across Power Apps, Power Automate, and Copilot Studio.
- AI Hub
- Central orchestration for enterprise AI — governed management of models, connectors, and copilots at scale.
- Copilot
- Natural-language creation of apps, flows, tables, pages, and logic across Power Platform.
- AI Builder
- Prebuilt + custom models: classification, prediction, extraction, detection, document understanding.
- Copilot Studio
- Low-code conversational agents, extensible into Microsoft 365 and external systems.
| Scenario | Copilot | AI Builder | AI Hub | Copilot Studio |
|---|---|---|---|---|
| Text generation | ✔ | — | ✔ | ✔ |
| Prediction | — | ✔ | ✔ | — |
| Document extraction | — | ✔ | ✔ | — |
| Conversational agent | — | — | — | ✔ |
| Automation creation | ✔ | — | ✔ | — |
Key concept · The custom-model lifecycle
AI Builder custom models follow a plan the architect owns: data acquisition & labeling → training & validation → production deployment → monitor drift & accuracy → retrain/retire triggers. Across all of this, the architect balances governance & security, data strategy, extensibility & integration, and monitoring & optimization.
Module 7 quiz
Fourteen questions on the three Dynamics 365 experience models, agents-as-products, Copilot for Sales/Service, and Power Platform AI / the AI Hub. Answer first, then expand Show answer.
Part A · Multiple choice
Which Dynamics 365 experience model places Copilot capabilities directly inside workspace pages and operational views?
- A Sidecar
- B Embedded
- C Outside
- D Offline
Show answer
B — Embedded. Sidecar sits alongside for chat; Outside is external orchestration beyond the UI.
The module's core principle for Microsoft 365 agents is:
- A "Ship fast, govern later"
- B "Treat an agent like a product, not a prompt"
- C "Bigger models are always better"
- D "Avoid human review to save time"
Show answer
B. Agents ship with a backlog, guardrails, and metrics — mission, grounding, guardrails, and telemetry.
In a RACI matrix, how many roles should be Accountable for a given task?
- A As many as possible
- B Exactly one
- C At least three
- D Zero — it's optional
Show answer
B — Exactly one. Only one role is ultimately answerable; multiple can be Responsible, Consulted, or Informed.
Which is the central, governed workspace for discovering and orchestrating AI across Power Platform?
- A AI Builder
- B Copilot Studio
- C AI Hub
- D Power BI
Show answer
C — AI Hub. It centralizes models, connectors, and copilots for consistent, secure, monitored use at scale.
For building a low-code conversational agent, which Power Platform feature is the right fit?
- A AI Builder
- B Copilot Studio
- C Copilot in Power Apps
- D AI Hub only
Show answer
B — Copilot Studio. Per the feature selection matrix, conversational agents map to Copilot Studio.
What must happen before enabling Copilot for Sales or Service, per the architecture-first approach?
- A Nothing — enable it and iterate
- B Validate/standardize/label data and configure least-privilege access
- C Delete the existing CRM
- D Turn off all audit logging
Show answer
B. Data readiness and role-based access come first, then connect systems, map fields, and enable actions.
Which document-processing task is a natural fit for AI Builder rather than a conversational agent?
- A Multi-turn chat with a customer
- B Extracting structured fields from invoices/emails
- C Voice IVR routing
- D Drafting a persona
Show answer
B. Document extraction and prediction map to AI Builder in the feature matrix.
Part B · True or false
The "Outside" experience model lets external agents orchestrate across apps and tasks beyond the Dynamics 365 UI.
Show answer
True. It uses Dataverse or custom APIs for governed access and must honor data-residency and approval constraints.
A well-designed Microsoft 365 agent needs no telemetry as long as it has good instructions.
Show answer
False. Telemetered outcomes (adoption, quality, latency, cost, business impact) are part of what separates an agent from an ad-hoc prompt.
Human review before sending externally generated content is a required operational guardrail.
Show answer
True. Alongside sensitivity labels, DLP, audit trails, and data residency.
AI Builder offers only prebuilt models and cannot be trained on organization-specific data.
Show answer
False. It provides both prebuilt models and custom models trained on your data (prediction, classification, extraction).
Part C · Short answer
Name the three Dynamics 365 F&SC AI experience models.
Show answer
Sidecar, Embedded, and Outside (external orchestration).
List the five steps of the Microsoft 365 agent design framework.
Show answer
Frame the job to be done → map inputs, knowledge & actions → define guardrails → prototype the critical path → operationalize.
What do the four letters of RACI stand for?
Show answer
Responsible (does the work), Accountable (one owner answerable), Consulted (two-way input), Informed (one-way updates).