Copilot & Foundry: The Microsoft AI Stack for Small Business · Lesson 7 of 10
Project Kickoff: Northside HVAC Internal Helpdesk
Time: ~25 min reading + ~15 min worksheet Goal: Meet the fictional client, understand why this project is a good first agent build, make the architecture decision, and plan the data inventory.
---
The next four modules walk through one continuous project: designing and building an internal helpdesk agent for a fictional client, Northside HVAC. You'll see every phase, every decision, and every place we use each tier of the Microsoft stack.
We picked this project deliberately. It's:
- A realistic SMB problem
- Scoped small enough to finish
- Representative of ~60% of the Copilot Studio agent work you'll do in the real world
- Has a natural moment where Copilot Studio hits a wall — and we add Foundry (Module 9)
If you follow along with your own M365 tenant, by the end you'll have shipped a real working agent. If you don't have a tenant, the Microsoft Developer Program offers a free M365 E5 sandbox (90-day, renewable) — apply at developer.microsoft.com/microsoft-365/dev-program. It's what most consultants practice on.
Northside HVAC Services
Industry: Commercial and residential HVAC services Location: Mid-Atlantic region, one shop location Size: 45 employees Revenue: ~$7M/year Systems: M365 Business Premium across the company. ServiceTitan for dispatching and job management. QuickBooks Online for accounting.
The Team
- 30 field technicians (dispersed across company trucks, mix of smartphones and ruggedized tablets)
- 6 office staff (reception, dispatch, billing, inside sales)
- 4 managers (operations, service, sales, install)
- 3 admin (HR generalist, bookkeeper, office manager)
- 2 owners (Mike and Lisa)
Leadership
- Mike and Lisa Petrov — husband-and-wife owners, founded the business in 2008. Mike runs operations, Lisa runs the office. Both hands-on.
- Rachel, Office Manager — 15 years with the company. Knows everything. Answers everyone's questions. Has hit her limit.
The Pain
Rachel is the reason you're here. In a discovery conversation, she lays it out:
> "I'm getting pulled away from real work 30–40 times a day with the same questions. 'How do I expense a meal?' 'Do I have to turn my phone in when I'm on vacation?' 'Does my health insurance cover my wife's chiropractor?' 'Why is my 401(k) statement showing the old amount?' Most of these are answered in the handbook nobody reads, or the benefits PDF from the insurance broker, or a folder on SharePoint. But they come to me because I'm the one who always answers. I'm tired. And when I'm out sick, the whole office ships the questions to Lisa, who has less patience and less time than I do."
Lisa adds:
> "We've been saying for a year that we'd 'do something about it.' Write an FAQ, build an intranet, train the managers to answer more of these. Nothing's stuck. I heard about AI helpdesk bots and I want to know if that's real."
The Ask
Mike wants a working pilot in 90 days. Budget comfort: "somewhere around $20–$30K if the case is clear, more if we build something real." He's had vendors try to sell him $80K AI projects and isn't buying. He wants to see it work at modest scale first.
Before we architect it, let's articulate why this project is almost ideal for Copilot Studio:
1. Defined scope
A finite set of questions. Rachel can list the top 20 in 10 minutes.
2. Real, documented ROI
Rachel spends 1.5–2 hours/day on these questions. At her loaded rate, that's $15K+/year of recoverable time. Plus the effect on Lisa. Plus the effect on office morale when Rachel is actually available for real work.
3. Low stakes
Wrong answers are recoverable. Worst case: "check with Rachel if this doesn't sound right." Not a life-or-death or legal-liability situation.
4. Existing content
The answers already exist — they're just scattered. Handbook, PDFs, SharePoint, emails. We don't need to generate knowledge; we need to retrieve it.
5. Natural deployment channel
They're all in Teams all day. Deploying there is zero friction.
6. Scales to more agents later
If this works, the pattern repeats: a dispatcher helper, a sales assistant, a customer-service deflection bot. Land this one, earn the next.
7. The graduation story
As Rachel and Lisa add more content and the agent gets used more, we'll likely hit moments where Copilot Studio strains. That's our cue to add Foundry (Module 9). This project is designed to teach the graduation.
Apply the decision tree from Module 5.
- Single-user productivity task? No — multi-user.
- Is there existing SaaS? Not really. HR helpdesk bots exist but are either enterprise-scale or toys.
- Internal agent, moderate scope, sub-500 docs, M365-grounded? YES.
- (Don't need to keep going.)
Decision: Pattern 1 (Copilot Studio only) for the initial build.
We'll likely graduate to Pattern 2 (add Foundry) at some point — that's explicitly part of the plan. But we start in Copilot Studio because that's where the decision tree lands, and it's the cheapest fastest path to proving value.
What will the agent know? Based on a quick walkthrough with Rachel, here's the data landscape:
Primary sources (clean, structured)
- Employee Handbook (PDF) — 34 pages, up to date, covers PTO, benefits summary, conduct, general policies
- Benefits Guide 2026 (PDF) — 22 pages, from insurance broker, covers health/dental/vision/401(k)/life
- IT Acceptable Use Policy (Word doc) — 8 pages
Secondary sources (messier)
- SharePoint site: "Northside Team" — various folders with SOPs, forms, announcements. Some old, some current.
- SharePoint folder: "Forms" — blank forms for PTO requests, expense reports, equipment requests
Tertiary sources (leave for later)
- Email chains where Rachel has answered specific questions before (not going to index these — too messy)
- Rachel's head (document the important stuff during discovery)
The knowledge curation task
Before we build anything, someone needs to spend ~4 hours doing cleanup:
- Confirm the handbook and benefits guide are current
- Archive obsolete SharePoint content or mark it clearly
- Apply sensitivity labels to anything sensitive (payroll-specific content stays out of the agent)
- Create a simple SharePoint site specifically for "things the agent can learn from" and put clean copies there
This is part of the engagement. Budget time for it. Clients want to skip this step; don't let them — it's the difference between an agent that works and one that embarrasses you.
Agree on what "working" means before you build. For Northside, the metrics we'll track:
Primary metrics
- Volume handled: questions per week routed to the agent
- Deflection rate: % of questions answered without Rachel's involvement
- Rachel's time reclaimed: self-reported + estimated hours/week
Secondary metrics
- User satisfaction: simple 👍/👎 thumbs-up in the agent interface
- Top unanswered questions: what the agent failed on (for content gap analysis)
- Escalation rate: % of questions routed to human
Targets by end of 90-day pilot
- 20+ questions per week handled by agent
- 70%+ satisfaction rate
- Rachel reports saving 5+ hours/week
These are realistic, defensible targets. Err on the conservative side — you want to *exceed*, not apologize for missing.
Before committing:
| Dimension | Score 1–5 | Notes | |---|---|---| | Data readiness | 4 | Content exists, scattered but clean-able | | Technology readiness | 5 | M365 BP, SharePoint in good shape | | Workflow readiness | 4 | Q&A pattern is well-defined | | People readiness | 3 | Field techs are mixed on tech; office team is fine | | Leadership readiness | 5 | Mike + Lisa aligned and enthusiastic | | Total | 21/25 | GO |
Scoring 21/25 is "go fast" territory. Good project.
Here's how we'll structure the 90-day project:
Weeks 1–2: Discovery & Prep
- Detailed session with Rachel to capture top 20 questions
- Content cleanup and curation
- SharePoint site setup for agent's knowledge base
- Governance review: sensitivity labels, Purview setup check
Weeks 3–4: Build v1 in Copilot Studio (Module 8)
- Create the agent
- Configure topics for the top 10 questions
- Add knowledge sources
- Configure escalation-to-Rachel action
- Publish to Teams, test with Mike/Lisa/Rachel
Weeks 5–6: Pilot with 8 Users
- Enable for 8 users across departments
- Weekly check-ins
- Iterate based on real use
Weeks 7–9: Full Rollout + Measure
- Enable for all 45 users
- Office hours for adoption support
- Start collecting metrics
Weeks 10–13: Evaluate + Iterate
- Gap analysis: what questions did it fail on?
- Content additions, topic tuning
- Decision point: stay in Copilot Studio or start planning Foundry upgrade? (Module 9 scenario)
Total engagement fee: $22K (sweet spot for Mike's budget) Microsoft costs (annual recurring): ~$17K (45 Copilot licenses) + minimal Azure Ongoing managed service after 90 days: $1K/month retainer
---
- Northside HVAC helpdesk agent is a textbook first-agent project — defined scope, real ROI, low stakes, existing content.
- Decision: Pattern 1 (Copilot Studio only) based on Module 5 decision tree.
- Data inventory matters — budget time for content cleanup before you build.
- Agree on success metrics up front. Track them. Report them.
- Do a formal readiness check (Foundations Module 9 style).
- Structure the 90-day engagement explicitly — Discovery → Build → Pilot → Rollout → Measure.
---
Next up: Module 8 — Build It in Copilot Studio. Step-by-step.
---
Time: ~15 min Goal: Put the planning rigor from Modules 1–6 into practice before we build anything.
---
Imagine you're Rachel. Brainstorm the top 15 questions you get asked repeatedly — the kind a first-pass agent should handle. Think HR, benefits, IT basics, office logistics.
- ______________________________________________________
- ______________________________________________________
- ______________________________________________________
- ______________________________________________________
- ______________________________________________________
- ______________________________________________________
- ______________________________________________________
- ______________________________________________________
- ______________________________________________________
- ______________________________________________________
- ______________________________________________________
- ______________________________________________________
- ______________________________________________________
- ______________________________________________________
- ______________________________________________________
(We'll use this list as input to Module 8 when we build topics.)
For each type of content, decide: Include in agent's knowledge, Exclude, or Include with sensitivity label.
- Employee Handbook (general policies): ______________
- Benefits Summary PDF from insurance broker: ______________
- Individual employee's specific benefits enrollment: ______________
- Company-wide IT Acceptable Use Policy: ______________
- Current list of employee names and job titles: ______________
- Salary ranges for each role: ______________
- Individual salary records: ______________
- Expense reimbursement policy: ______________
- Past expense reports (all employees): ______________
- Company holiday schedule: ______________
- 401(k) plan document (generic): ______________
- Employee reviews / performance data: ______________
Review the success metrics proposed for Northside:
- 20+ questions/week handled
- 70%+ satisfaction
- Rachel reports saving 5+ hours/week
Answer:
Are any of these too aggressive for 90 days? Which one(s)? ______________________________________________________
What metric did we leave out that we should add? ______________________________________________________
What would you commit to in writing to the client, vs. what would you leave as aspirational? ______________________________________________________
---
Exercise 1 — realistic top-15 helpdesk questions:
- How do I request PTO?
- How much PTO do I have left?
- What holidays does the company observe?
- How do I expense a meal / mileage / purchase?
- Where do I submit an expense report?
- What does my health insurance cover?
- When does open enrollment start?
- How do I add a dependent to my health plan?
- How do I access my 401(k) account?
- What's the policy on remote work?
- How do I request a new piece of equipment (phone, tool, laptop)?
- What's the procedure for calling in sick?
- Who do I talk to about a payroll issue?
- What's the bereavement leave policy?
- How do I update my contact info / beneficiaries / W-4?
This is roughly the pattern of 60–70% of the questions Rachel actually gets.
Exercise 2 — sensitivity calls:
- Include — general policy, no sensitive data
- Include — generic benefits summary is fine
- Exclude — individual records. Agent should never have access.
- Include — general policy
- Include with sensitivity label — org chart is typically fine but mark as Internal
- Include with sensitivity label — if there's a published range (common at modern companies), Internal label. If not published, exclude.
- Exclude — hard no. Individual compensation data never goes in the agent.
- Include — general policy
- Exclude — individual expense records are personal data
- Include — non-sensitive
- Include — generic plan doc is fine. Individual statements would be excluded.
- Exclude — sensitive HR data
Rule of thumb: policies and generic documents = yes. Individual-specific records = no, even if Rachel has access. The agent should know the *rules*, not the *data*.
Exercise 3 — metrics critique:
Too aggressive: 20+ questions/week in first pilot group of 8 users is plausible; hitting it across 45 users is easier. But depending on how "sticky" the initial deployment is, you could fall short. More realistic: 15+ questions/week during pilot phase, 30+/week post-full-rollout.
Missing metric: content gap rate — what % of questions is the agent failing to answer at all (vs. answering correctly vs. escalating). This is what drives your iteration plan. Also: Rachel's workload reduction independently verified (maybe track a week of her inbox / interruptions before and after).
Commit in writing: I'd commit to:
- Working agent deployed to all users by day 60
- Weekly metrics report during 90-day period
- Content gap analysis and remediation plan by day 90
I'd leave as aspirational:
- Specific Rachel-time-saved number (it'll depend on adoption)
- Specific user satisfaction %
- Any ROI dollar figure beyond "we expect"
Under-promise, over-deliver. That's how you earn the second engagement.
---
Check yourself: Did Exercise 1 come easy? If you struggled, re-interview an imaginary Rachel with more specific questions — Modules 8–9 need this list as input. Ready to build in Module 8.
Previous: Pricing, Licensing, and TCO · Next: Build It in Copilot Studio (Step-by-Step)