42 HR processes scored, and a roadmap agreed on in one 90-minute review

The challenge

HR at the university runs on experienced people and fragmented systems: shared mailboxes, hand-maintained spreadsheets, and the same templated emails sent over and over. Information about one employee lives in half a dozen places.

Interest in AI was real. One team had already built Copilot agents on its own. What was missing was a plan the whole department could stand behind, resting on written rules for who may use AI, on what data, and with whose approval.

What we did

1. 01Discover, then go deep

   We mapped the work with HR leadership, then ran deep-dives with the teams that do it, writing down what they said in their own words.

2. 02One scorecard

   Each process was scored on eight equally weighted criteria: frequency, time saved, pain, data readiness, risk, sensitivity, build complexity, and reusability. Risk, sensitivity, and complexity count against a process.

3. 03Shared patterns instead of one-off builds

   The same job showed up in different teams. Nearly all scored processes fit one of these: mailbox triage and draft response, grounded Q&A, templated communication, dataset reconciliation, and form-field extraction.

4. 04Saying no, and saying fix the tool

   Employment eligibility verification and formal investigations scored lowest, because the regulatory exposure and the cost of an error outweigh what AI would save. Some tedious tasks existed only because a system produced an unusable report, so the recommendation was a report fix, not an AI layer.

5. 05Governance before scale

   Two things have to be in place before any of it scales: a sensitivity label that lets Copilot work with internal HR material, and one governed SharePoint library for each team’s reference documents.

How the scored processes fell

Results

• One plan, with owners

  The review ended with the pilot area confirmed, a documentation owner named per team, and governance decisions assigned.

• The first agent already in build

  The mailbox triage agent was started before the review, so the department saw a working demonstration, not a slide.

• Clear limits on AI

  High-risk processes are documented as out of scope, with reasons.

• Evidence behind each recommendation

  Each roadmap item traces to scored processes in a matrix the department was still using as of September 2026.

Read about the service: AI Strategy & Roadmap, AI Governance, Microsoft 365 Copilot Adoption and Agents, Power Platform

What came next: an agent for the HR mailbox

Demonstrated, pilot pending, September 2026. No outcome figures are published until the pilot is measured.

The assessment found that most of the six top-ranked HR processes were the same job: read a shared mailbox, classify each email, and answer from material the team already has. Employee Relations ranked first, so the first build is an agent for that mailbox, made so the same agent can serve Payroll, Benefits, and Recruiting without starting over.

1. 01It reads

   A Power Automate flow watches the shared mailbox and triggers the agent, which works out what the sender is asking for.

2. 02It matches

   The agent answers only from the team’s documented responses in a governed SharePoint library. If it can’t match with confidence, it leaves the email for normal handling.

3. 03It drafts

   It writes a reply in the team’s voice, stages it in the drafts folder, and logs the request and draft to a SharePoint tracking list.

4. 04A person sends

   A team member reviews, edits if needed, and sends. No reply goes out on its own. The team chose that design during the assessment.

Today

Demonstrated with Employee Relations, from an inbound email to a grounded draft staged for review, with the request logged.

What’s left

Response formatting and testing against real mail once the team’s response inventory comes back.

How it will be measured

Drafts generated, drafts accepted, and time recovered, from the first week of the pilot.