Deploy Agents on Microsoft: Copilot Studio, Fabric & Power Platform

20 slides. The simple, step-by-step path to shipping a working AI agent inside a Microsoft-centric shop — Copilot Studio for the agent, Fabric for the data, Power Platform for the actions, and governance so IT says yes.

20 lessons · ~4 min · free, no signup.

Lessons

Full course text

1. The Microsoft agent stack in one slide

Copilot Studio builds and publishes the agent. Microsoft Fabric holds the data it reasons over (OneLake). Power Platform (Power Automate + Power Apps) gives it hands to take action. Entra ID and Purview keep it governed. Four pieces, one license family.

Source: Microsoft Learn — Copilot Studio overview

2. Agent vs. Copilot vs. flow

Microsoft 365 Copilot is the assistant Microsoft ships. An agent is your version: scoped instructions, your knowledge, your actions. A Power Automate flow is deterministic automation an agent calls. Most real deployments use all three together.

Source: Microsoft Learn — Agents in Copilot Studio

3. Step 0 — pick a workflow worth automating

Choose one high-volume, low-ambiguity task with a written process: policy Q&A, RFP first drafts, ticket triage, order status. Write today's baseline (volume, minutes per case, error rate). No baseline, no proof, no budget next quarter.

Source: McKinsey — The State of AI

4. Step 1 — create the agent in Copilot Studio

Go to copilotstudio.microsoft.com → Create → New agent. Describe it in plain language, name it, set the tone, and paste your instructions (role, scope, what to refuse, when to hand off to a human). Fifteen minutes gets you a testable agent.

Source: Microsoft Learn — Create and deploy an agent

5. Step 2 — add knowledge

Attach SharePoint sites, OneDrive folders, Dataverse tables, public websites, or uploaded files. Point it at the smallest authoritative set — curated beats comprehensive. Turn on 'only answer from knowledge' when accuracy matters more than coverage.

Source: Microsoft Learn — Add knowledge to an agent

6. Step 3 — give it actions

Actions are how an agent does something: a Power Automate cloud flow, a connector (Outlook, Teams, SQL, ServiceNow), a Dataverse write, or a REST API. Describe each action in plain language — the model reads that description to decide when to call it.

Source: Microsoft Learn — Add actions

7. Step 4 — test in the built-in pane

Use the test panel with real questions, including the ugly ones. Turn on the activity map / generative orchestration trace to see which knowledge and actions fired. Fix the instructions first — most bad answers are scope problems, not model problems.

Source: Microsoft Learn — Test your agent

8. Step 5 — publish where people already are

One click publishes to Microsoft Teams, the Microsoft 365 Copilot chat, a SharePoint site, your website, or Dynamics 365 channels. Rule of adoption: never make anyone open a new app. Ship into Teams first.

Source: Microsoft Learn — Publish and channels

9. Fabric & OneLake, plainly

Microsoft Fabric is one SaaS analytics platform (data factory, warehouse, lakehouse, real-time, Power BI) sitting on OneLake — a single tenant-wide data lake. One copy of the data, many engines. It's the enterprise memory your agent should read from.

Source: Microsoft Learn — What is Fabric?

10. Step 6 — land data in Fabric

Ingest with Data Factory pipelines or Dataflows Gen2 into a Lakehouse. Model it as bronze (raw) → silver (cleaned) → gold (business-ready tables). Agents should only ever query gold. Clean semantics in, trustworthy answers out.

Source: Microsoft Learn — Lakehouse medallion architecture

11. Shortcuts & Mirroring beat copying

OneLake shortcuts reference data where it lives (ADLS, S3, Dataverse); Mirroring continuously replicates Azure SQL, Cosmos DB, or Snowflake into Fabric for free. Fewer copies means fewer stale answers and lower cost.

Source: Microsoft Learn — OneLake shortcuts

12. Step 7 — connect Fabric to the agent

Two proven paths: (a) build a Power BI semantic model and let the agent answer over it, or (b) expose gold tables through Fabric SQL endpoints / Dataverse and call them as an action. Ask questions in business language; if the model can't name the metric, neither can the agent.

Source: Microsoft Learn — Fabric data agent

13. Step 8 — Power Automate does the doing

Build the flow first and test it standalone: trigger → get data → decide → write/notify. Then attach it as an agent action. Keep flows small and single-purpose; an agent orchestrating five clean flows beats one 40-step monolith nobody can debug.

Source: Microsoft Learn — Power Automate cloud flows

14. Autonomous (triggered) agents

Copilot Studio agents can run without a human prompt — triggered by an email arriving, a Dataverse row changing, a Teams message, or a schedule. That's the jump from chat assistant to digital coworker. Start human-in-the-loop, then remove the click once accuracy holds.

Source: Microsoft Learn — Triggers for autonomous agents

15. Step 9 — governance so IT says yes

Use a dedicated Power Platform environment with a managed Dataverse, Entra ID groups for access, DLP connector policies to block risky data paths, and Purview for sensitivity labels and audit. Agents inherit the user's permissions — get SharePoint permissions right first.

Source: Microsoft Learn — Power Platform governance

16. What it costs

Copilot Studio bills on messages — prepaid capacity packs or pay-as-you-go via Azure, with different message rates for classic answers, generative answers, and autonomous actions. Fabric bills on capacity units (F SKUs) you can pause. Model the message volume before you launch company-wide.

Source: Microsoft — Copilot Studio pricing

17. Step 10 — measure and iterate weekly

Watch the Copilot Studio analytics: engaged sessions, resolution rate, escalation rate, abandoned topics. Pair it with your baseline from Step 0. Each week: fix the top failure, tighten instructions, add one missing action. That loop is the whole job.

Source: Microsoft Learn — Analyze agent performance

18. Myth: you need developers to start

Copilot Studio, Power Automate, and Fabric are all low-code first — a business analyst can ship v1. Bring developers in for custom connectors, complex APIs, Azure AI Foundry models, and CI/CD via solutions and pipelines. Start no-code, graduate deliberately.

Source: Microsoft Learn — Copilot Studio for makers

19. Your 5-day Microsoft agent sprint

Day 1: pick the workflow, write the baseline. Day 2: create the agent + attach SharePoint knowledge. Day 3: build and attach one Power Automate action. Day 4: point it at a Fabric gold table for the numbers. Day 5: publish to Teams for 10 pilot users and start the weekly loop.

Source: Robauto — agent deployment methodology

20. Agents built. Now let them transact.

You can stand up a governed, data-connected agent on the stack your company already pays for. Final course: how agents actually buy and sell — and how to be buyable.

Source: Continue → Agentic Commerce 101

All courses · Robauto