Copilot & Foundry: The Microsoft AI Stack for Small Business · Lesson 8 of 10

Build It in Copilot Studio (Step-by-Step)

Time: ~35 min reading + ~30 min worksheet (hands-on) Goal: Walk through the actual build of the Northside HVAC helpdesk agent in Copilot Studio. By the end, if you're following along in a tenant, you'll have a working agent.

---

Before You Start

You need:

Access to a Microsoft 365 tenant with Copilot Studio available (M365 Copilot license, or trial)
Admin access to a SharePoint site where the agent's knowledge base will live
The content: Employee Handbook PDF, Benefits Guide PDF, IT AUP (can use placeholders for practice)

If you don't have a tenant, set up a Microsoft 365 Developer Program sandbox at developer.microsoft.com/microsoft-365/dev-program — it's free, includes E5 licenses, and is specifically designed for this kind of practice.

The screenshots-and-specifics of Copilot Studio change. The clicks may differ slightly by the time you read this. The concepts and the structural flow of the build will be stable. When in doubt, trust the concepts and navigate the current UI.

Step 1: Set Up the Environment (10 min)

1.1 Sign in to Copilot Studio

Go to copilotstudio.microsoft.com and sign in with your tenant admin account.

1.2 Choose or create an environment

Copilot Studio works within Power Platform environments. For a client project:

For a real production build: create a new environment specifically for this client (admin.powerplatform.microsoft.com → Environments → + New). Separating environments keeps governance, DLP policies, and access control clean per-client.
For practicing: use the default environment.

1.3 Verify licensing

Copilot Studio will warn you if licensing isn't in place. If you're on a developer sandbox or a fresh M365 Copilot trial, you're fine. In production with a real client, confirm:

M365 Copilot licenses deployed (for the users who'll interact with the agent)
Copilot Studio capacity assigned to the environment

1.4 (Production only) Apply DLP policies

In the Power Platform admin center, verify a DLP policy exists that:

Separates "business" connectors (SharePoint, Dataverse, internal APIs) from "non-business" connectors (social media, public APIs)
Blocks or restricts sensitive data from flowing to non-business connectors

This is one-time per environment. For Northside, we'd use their existing DLP policy or create a standard one.

Step 2: Create the Agent (5 min)

2.1 Create new

In Copilot Studio, click + Create → New agent.

2.2 Describe it

Copilot Studio will offer to help you create the agent from a description. Enter something like:

> "An internal HR and IT helpdesk agent for Northside HVAC Services, a 45-person HVAC company. It answers employee questions about PTO, benefits, expenses, IT policies, and other workplace policies. When it can't answer, it creates a Teams notification to Rachel, the office manager, with the user's question and context. Tone: friendly, helpful, practical. Never make up answers — if uncertain, always escalate to Rachel."

This generates starter topics and configuration. Review what it creates; adjust.

2.3 Name and describe

Name: Northside Helpdesk
Description: "Internal helpdesk for HR, benefits, IT, and office policy questions. Always escalates to Rachel when uncertain."
Icon: upload company logo if available

2.4 Initial greeting

Override the default greeting with something that sets expectations:

> "Hi! I'm the Northside helpdesk bot. I can answer questions about benefits, PTO, expenses, IT policies, and most HR topics. If I'm not sure about something, I'll send your question to Rachel and she'll follow up. What can I help you with?"

Clarity up front beats cleverness. Users will forgive a bot; they won't forgive a bot that pretends to know things it doesn't.

Step 3: Add Knowledge Sources (10 min)

In the agent's Knowledge section:

3.1 Add SharePoint source

Click + Add knowledge → SharePoint. Point at the curated SharePoint site or folder where the clean versions of the handbook, benefits guide, and IT AUP live.

Important: use a dedicated folder like `Sites/NorthsideTeam/AgentKnowledge/` rather than the root of the team site. This isolates what the agent can see and makes permission management cleaner.

3.2 Add document sources

For any PDFs or Word docs not in SharePoint (uncommon but happens), you can upload directly. For Northside, everything's in SharePoint, so skip.

3.3 Verify permissions

The agent searches as itself. Confirm the service principal / app it runs under has read access to the knowledge folder. If users have different access to different content, consider either:

Keeping the agent's knowledge scope narrow (everyone can see everything it knows)
Using Dataverse or a more sophisticated RAG pattern with user context (graduates toward Foundry — see Module 9)

For Northside v1, the simple approach works: everything in `AgentKnowledge/` is accessible to all employees, and the agent only knows what's in there.

3.4 Configure generative answers

In Copilot Studio's settings, enable Generative answers and point at your knowledge sources. Set:

Answer tone: friendly, professional
Authoritative sources: SharePoint site
Citations: enabled (agents should cite which document the answer came from)

Step 4: Configure Topics for the Top 10 (15 min)

Now the craft. Topics are where you author specific handling for high-value questions — the ones you know recur. For everything else, generative answers on the knowledge base handles it.

For Northside, author explicit topics for:

Topic 1: Request PTO

Trigger phrases:

"request time off"
"PTO request"
"take vacation"
"request PTO"
"how do I request time off"

Flow:

Acknowledge
Direct user to the PTO request form in SharePoint (link)
Remind them of the 2-week advance notice policy (from handbook)
Offer to create a Teams calendar hold (optional action)
End gracefully

Topic 2: Check PTO balance

Trigger phrases:

"how much PTO do I have"
"PTO balance"
"vacation days left"

Flow:

Explain that PTO balance is tracked in [payroll system, e.g., ADP]
Provide direct link to the employee self-service portal
Offer to escalate to Rachel if they can't access it

Note: we're NOT trying to look up balances directly. That's a connector project and likely Foundry territory for a real integration. The agent points users to the source.

Topic 3: Expense a meal / purchase

Trigger phrases:

"expense a meal"
"submit expense"
"how do I expense"
"expense report"

Flow:

Explain expense policy highlights (what qualifies, max amounts)
Link to expense report template
Explain approval process (submit to Lisa by end of month)
Link to the full expense policy in SharePoint

Topic 4: Health insurance questions

Trigger phrases:

"health insurance"
"insurance covers"
"medical plan"
"dental"
"vision"

Flow:

Explain this is a general info topic — specific coverage questions go to insurance broker
Provide broker contact info
Link to the benefits guide
Mention open enrollment dates if relevant time of year

Topic 5: 401(k) questions

Trigger phrases:

"401k"
"retirement"
"401(k) statement"

Flow:

Point to Principal/Fidelity portal (or whichever plan administrator)
Provide contact for plan administrator
Link to generic plan doc

Topic 6: Report a tech problem

Trigger phrases:

"my computer is broken"
"IT issue"
"can't log in"
"tech problem"

Flow:

Ask what the specific issue is
If password reset → point to self-service Entra reset
If general IT → create a Teams ticket with context for [IT partner or internal IT contact]

Topic 7: Remote work policy

Trigger phrases:

"work from home"
"remote work"
"WFH"

Flow:

Summarize policy: office staff, case-by-case basis, manager approval
Link to full policy
Direct to manager for specific requests

Topic 8: Calling in sick

Trigger phrases:

"call in sick"
"I'm sick"
"sick day"

Flow:

Remind them of the notification process (text/call dispatch by 6am)
Field techs call dispatch directly
Office staff notify Lisa
Link to sick leave policy

Topic 9: Update personal info

Trigger phrases:

"update my address"
"change my W-4"
"update beneficiary"
"change my phone number"

Flow:

Direct to self-service portal
For things not in self-service (beneficiaries, etc.), route to Rachel
Link to relevant forms

Topic 10: Escalate to Rachel (fallback)

Trigger: no other topic matched, AND generative answers returned low confidence Flow:

Tell the user: "I'm not sure about that — let me get Rachel on it."
Capture the user's question
Action: Post a Teams message to Rachel with user name, question, and timestamp
Confirm to user: "I've sent this to Rachel. She'll follow up."

Step 5: Configure the Escalation Action (5 min)

The escalation action is the most important thing the agent does — it's the safety net.

5.1 Create a Power Automate flow

In Power Automate, create a flow triggered by Copilot Studio:

Trigger: When Copilot Studio agent invokes this flow
Input parameters: `user_name`, `user_question`, `conversation_snippet`
Action: Post message to a Teams channel or chat with Rachel
Format: something like:

``` 🔔 Helpdesk agent escalation

From: {user_name} Question: "{user_question}" Context: {conversation_snippet} Timestamp: {current time} ```

5.2 Connect to the agent

In Copilot Studio, in the Actions tab, register this flow as an action called `EscalateToRachel`.

5.3 Wire it to Topic 10

Edit Topic 10 to invoke this action when triggered.

5.4 Test it

Trigger Topic 10 manually; confirm Rachel (or the test user playing Rachel) gets the Teams message.

Step 6: Test and Iterate (10 min)

6.1 Use the Test pane

Copilot Studio has a built-in Test pane on the right side of the builder. Use it.

Test scenarios:

"How do I request PTO?" → expect Topic 1
"I forgot my password" → expect Topic 6 (or Entra SSPR flow)
"What's my vacation balance?" → expect Topic 2
"How do I expense lunch?" → expect Topic 3
"Can I bring my dog to work?" → expect knowledge lookup or escalation
"Um, weird question..." → expect gentle escalation

For each test:

Did the right topic fire?
Was the answer correct?
Was the tone right?
Did citations appear where relevant?
If it escalated, did the Teams message arrive?

6.2 Iterate topics

If trigger phrases aren't firing correctly, add variants. If the generative answer is hallucinating, narrow the knowledge source or add explicit topic. If tone is off, adjust the agent's description and system guidance.

Step 7: Publish to Teams (5 min)

7.1 Publish

In Copilot Studio, click Publish. Name the version. Add release notes.

7.2 Deploy to Teams

In the Channels tab, enable Microsoft Teams. Copilot Studio will package the agent as a Teams app.

7.3 Upload to Teams admin center

Download the app manifest, upload via Teams Admin Center, and assign to:

Pilot group first — 8 users in a dedicated group
Full tenant later after pilot

7.4 Communicate

Send a short email to the pilot group with:

What the agent is
Where to find it (Teams left nav or Apps search)
Example questions to try
How to give feedback (👍/👎 in the agent, or email Rachel)

What to Show Mike and Lisa in the First Demo

When Mike and Lisa see v1:

Open Teams, find the agent
Ask "How do I request PTO?" — watch it answer
Ask "What does my health insurance cover?" — watch it guide to the broker
Ask "Can I bring my dog to work?" — watch it escalate to Rachel
Show Rachel getting the Teams notification
Ask "How much PTO do I have left?" — watch it point to the self-service portal

Then frame it: > "This is v1. It handles 10 topics explicitly, and anything else gets searched against the handbook and benefits guide. Over the next 6 weeks, the pilot group will surface questions it doesn't handle well — we'll add topics and tune. By day 90, it should be handling 70%+ of what comes to Rachel today."

That demo closes the pilot phase. Mike and Lisa are now invested.

---

Key Takeaways

Set up environment first — proper separation, DLP, licensing before anything else.
Agent creation is description-first — Copilot Studio generates scaffolding from a well-written description.
Knowledge sources live in a dedicated SharePoint folder — keeps scope clean, permissions manageable.
Author explicit topics for the top 10 — the rest covered by generative answers.
The escalation topic is the safety net. Build it first, make it reliable.
Test systematically before publishing, iterate on trigger phrases and tone.
Deploy to a pilot group first, not everyone. Always.
The first demo closes the pilot phase. Show Mike and Lisa something they can touch.

---

Next up: Module 9 — When to Bring in Foundry. Northside's agent starts hitting limits; we graduate to Pattern 2.

---

Time: ~30 min (hands-on) Goal: Actually do (or simulate doing) the build. Nothing replaces hands-on reps.

---

Exercise 1: The Build, For Real (20 min)

If you have a tenant (your own or a Developer Program sandbox), actually do it. Create a Copilot Studio agent. Add at least two knowledge sources (upload PDFs or point at a SharePoint folder with a few sample docs). Configure at least three topics. Test in the Test pane.

If you can't do it live right now, plan it out in writing:

Agent name: ______________

Agent description (2 paragraphs, suitable for the description prompt): ```

```

First three topics:

Topic A: ______________ Trigger phrases (4+): ______________ Flow (2–3 steps): ______________

Topic B: ______________ Trigger phrases (4+): ______________ Flow: ______________

Topic C: ______________ Trigger phrases (4+): ______________ Flow: ______________

Knowledge sources planned:

______________
______________

Escalation action design (one paragraph): ```

```

Exercise 2: Tough Test Cases (5 min)

Write 6 test questions that would stress-test the Northside agent — the ones likely to expose weaknesses. For each, predict what *should* happen.

Test question: ______________