Copilot Studio: Building a Custom HR Agent in 4 Weeks

Copilot Studio: what is it, really?

Microsoft 365 Copilot answers questions using your data in SharePoint, OneDrive, Teams, and Outlook. But when an employee asks something specific - "how many vacation days do I have left?", "what's our mileage reimbursement policy?", "how do I report a workplace injury?" - Copilot 365 can't answer, because the answer lives in an external HR system or in an internal policy PDF.

That's exactly what Copilot Studio is for: building custom conversational agents wired into your specific data sources (Dataverse, SharePoint, SQL, business APIs), accessible from Teams, from a website, or embedded directly into Microsoft 365 Copilot as an additional "skill."

This isn't a 2018-era chatbot. It's a generative AI layer that orchestrates searches, API calls, and actions - all through a low-code canvas.

Real-world case: an HR assistant agent at an 800-employee manufacturer

To make it concrete, here's the case of an HR agent deployed at a Canadian manufacturer in early 2025. The context: 800 employees across 4 sites, and a 6-person HR team swamped by repetitive questions (40% of their time spent answering questions whose answers are already in the policies).

The goal: build a Teams agent available to every employee, able to answer the 80% of common questions 24/7 and to intelligently route the other 20% - the ones that need a human - to the HR team.

Result at 6 months: 4,200 conversations handled by the agent, a 76% self-resolution rate, a 30% reduction in the HR team's workload, and user satisfaction of 4.3/5.

The agent's architecture

The agent combines four data sources and three actions:

• Source 1 - SharePoint: HR policies, the employee handbook, FAQs. Indexed automatically by Copilot Studio (Generative Answers).
• Source 2 - Dataverse: a table of vacation/sick-day balances per employee, synced nightly with the HRIS.
• Source 3 - HRIS API (BambooHR in this case): real-time retrieval of personal employee information.
• Source 4 - Power Automate flow: automatic creation of HR tickets in the ticketing system whenever the agent can't answer.
• Action 1 - Submit a time-off request (manager approval via an approval workflow).
• Action 2 - Report a workplace injury (creates a record in the OHS system with automatic notification to the health-and-safety committee).
• Action 3 - Request an employment verification letter (PDF generation plus email delivery).

Four weeks, week by week

The project was delivered in 4 weeks - a pace we recommend for this kind of focused agent:

• Week 1 - Scoping: 6 employee interviews and 2 workshops with the HR team to identify the 30 most frequent use cases. Defining the scope. Deciding "what the agent does / what it doesn't do."
• Week 2 - Build: setting up Copilot Studio, ingesting SharePoint, designing Dataverse, writing the first manual topics for the critical cases, and testing with 5 users.
• Week 3 - Integrations: HRIS API, Power Automate flows, Azure AD SSO authentication, and Purview sensitivity labels on employee data.
• Week 4 - Pilot and hardening: testing with 50 employee volunteers, tuning the system prompts, setting up the human fallback, and governance (data retention, audit logs).
• Broad launch in week 5 with internal communications, manager training, and weekly oversight for the first 4 weeks.

Governance and security

An HR agent handles sensitive personal data. The non-negotiable controls:

• SSO authentication is mandatory: the agent knows who's talking and reveals only that employee's data.
• Purview sensitivity labels on every source: an HR document labeled "Confidential - Internal" is restricted from being shared by the agent.
• Power Platform audit logs enabled: every conversation is traceable and exportable in the event of a Law 25 audit.
• Power Platform DLP policies: this agent is barred from connecting to consumer connectors (Gmail, Twitter, and the like).
• Retention policy: conversations are deleted automatically after 90 days, except in the case of an active dispute.
• Law 25, Section 12.1: employees are explicitly informed that the agent is an AI and have the right to request a human response.

Usage measurement and ROI

Without measurement, an agent dies within 6 months. The metrics we track:

Volume - number of conversations, by topic, by site, by day of the week. This surfaces peaks and emerging topics.

Self-resolution rate - the percentage of conversations that end without human escalation. Target: > 70%.

User satisfaction - a 1-5 rating requested at the end of each conversation. Target: > 4/5.

Time saved - extrapolated: (number of resolved conversations × average HR time saved per request). For this client: ~120 hours/month.

Total cost - Copilot Studio plus Dataverse licensing plus maintenance. Here: $2,400/month in production. Net positive ROI by month 3.

When Copilot Studio is the right fit

Copilot Studio shines in cases like these: an HR assistant (vacation, policies, onboarding), a Tier 1 IT assistant (passwords, access, ServiceNow FAQs), an internal sales assistant (catalog, configurator, order status), and a public-sector assistant (citizen information, online services).

Conversely, Copilot Studio is not the right answer for: generating creative content (use Copilot 365), analyzing complex documents (use Copilot 365 with SharePoint access), or doing quantitative prediction (use Power BI plus ML).

The classic trap: trying to build one mega-agent that does everything. You're far better off with 4 specialized agents that each do one thing well.