A custom Copilot agent that does the categorizing, so people don't have to

The organization introduced a requirement for the M&C team to log weekly hours against 10 official activity categories in a project tracking platform. There was no standard mapping between people's actual work and those categories, and no tooling to help.

• No clear mapping between daily work and the 10 activity categories
• Weekly reporting depended on manual recall, which invites inaccuracy and inconsistency
• Work varied widely across individuals, functions, and ownership levels, so one-size-fits-all would fail
• A colleague's existing agent, built for a CRM context, was a useful baseline but couldn't be applied as-is

I started from a colleague's existing agent: its system prompt structure, categorization logic, and setup method. I mapped what could be adapted versus what had to be designed new, and documented that line throughout the build. The agent was designed, built, and packaged for the team within a single extended working session.

The agent runs on a six-step weekly cadence:

1. 01

   Profile check

   Confirm the user profile is current before starting.

2. 02

   Scan

   Review the past week's signals using batched windowing and deduplication.

3. 03

   Draft

   Produce a categorized time report in a structured format.

4. 04

   Interpret and validate

   Surface interpretation calls and flag uncertainty before review.

5. 05

   Review

   Present the draft with a confidence column; the user confirms or corrects.

6. 06

   Finalize

   Produce submission-ready output.

I designed a multi-layer user profile so the agent can categorize work the way the person actually experiences it.

Work hierarchy

Role, portfolio, programs, projects, and recurring activities

Ownership level

Own, Lead, Contribute, Advise, or Monitor, set per program and project

Stakeholder tiers

Core, Strategic, and Emerging contacts

Invisible work

Effort that never shows up on a calendar

Known distinctions and exclusions

Negative knowledge: what something is not, so the agent stops miscategorizing look-alikes

Profile change log

A dated record of what changed and why

1. 01

   Scan first, not interview first

   I rejected a five-question onboarding interview. The agent scans available signals, drafts a profile, and the user corrects it. People should do less work up front, not more. That's the difference between a tool people try once and one they keep using.

2. 02

   Negative knowledge as a first-class field

   Miscategorization usually comes from things that look like one category but belong to another. A dedicated "known distinctions and exclusions" field fixes that once, instead of making users re-explain it every week.

3. 03

   Specific update suggestions, not generic prompts

   After each weekly draft, the agent proposes concrete profile updates based on what it observed. "Anything to update?" gets a shrug. "You spent 6 hours on X, which isn't in your profile. Add it?" gets an answer.

4. 04

   Ownership hierarchy over flat lists

   Owning a program and advising on it are different kinds of hours. A flat list of programs loses that signal; an ownership level tells the agent how to weight the work.

• Agent deployed and available to the team, built and packaged in a single extended session
• Six versioned deliverables, each written for a specific audience
• Calendar automation plus setup guide, so adoption doesn't depend on remembering
• Collaborator feedback gathered and follow-on improvements scoped; design decisions documented for handoff

Quality control was part of the build, not an afterthought: I caught and corrected inaccurate time-increment language across every deliverable before release, and rebuilt one structurally flawed document rather than patching it.

Download the one-page summary (PDF)

Read the essay: Build agents that do the work first