An AI chief of staff for the first 20 minutes of the day
Designing a Microsoft 365 Copilot agent that replaces reactive inbox scanning with a prioritized daily briefing, then piloting it, using it daily, and teaching my team to build their own.
- The problem
- Mornings started with 20 to 30 minutes of reading to find 5 minutes of real action items.
- What I did
- I designed a Copilot agent that turns the last 24 hours of email, chat, and meetings into a six-part briefing, then piloted it for two weeks against five quality checks.
- The result
- The agent is in daily use, and a build-your-own handout turned one person's workflow into a team pattern.
Inbox tools organize information. They don't organize attention.
Most mornings start reactively: open email, scan threads, and try to reconstruct what actually needs you. The problem isn't volume. It's prioritization.
- 20 to 30 minutes of reading to find 5 minutes of real action items
- Follow-ups buried in long threads, then missed
- Walking into meetings without knowing what decision is needed
- Losing track of where you're blocked, or where someone is waiting on you
- Escalations surfacing late because nothing flagged them early
Six sections, in the order a busy person needs them
The agent runs on a schedule, reviews the previous 24 hours of email, Teams messages, and meeting activity, and delivers a structured briefing before the inbox opens.
- 01
Today at a glance
3 to 5 sentences: key meetings, top priorities, time-sensitive items.
- 02
Needs my attention
Responses, decisions, approvals, and reviews, ranked by urgency and stakeholder impact.
- 03
Decisions required
Open questions where I'm the decision-maker, and what's blocked until I act.
- 04
Risks and blockers
Escalations, overdue items, and missed follow-ups that could derail work in flight.
- 05
Meeting prep
For each meeting: the goal, what I need to know, and what I should be ready to decide.
- 06
Can wait
FYIs and low-priority threads I can safely deprioritize.
Not plug-and-play: four layers of configuration
Agent configuration
- Role, responsibilities, and recurring priorities written into the agent description
- Key stakeholders, major projects, and their acronyms named so ownership is attributed correctly
- Knowledge scoped to email, Teams chats, Teams meetings, and relevant SharePoint libraries
- Starter prompts for six use cases, including daily briefing, midweek review, and meeting prep
Prioritization logic
- Prioritizes by work state, not project label
- Needs my response, decision, approval, or review
- Someone is blocked or waiting on me
- Deadline is today, overdue, or approaching
- Leadership visibility, escalation, repeated follow-up, or a material change in scope or timing
Accuracy guardrails
- Don't infer ownership from participation: attending a meeting isn't owning the action
- Don't invent deadlines or decisions not evidenced in the source
- Don't summarize everything; surface only what needs attention or carries risk
- Consolidate duplicate signals across threads and channels
Scheduling and automation
- Runs Monday to Friday before inbox-open time
- Separate Wednesday follow-through review on the same agent
- Email notification so the briefing lands before the inbox is opened
- Validated with a Run Now test before relying on scheduled output
Short, instruction-dense, and explicit about what to leave out
"Generate my Morning Briefing. Review relevant activity from the previous 24 hours, plus unresolved or upcoming items that materially affect today. Prioritize actions requiring my response, decisions requiring my input, risks and dependencies, meaningful project changes, and meetings requiring preparation. Do not provide a general recap."
The last sentence does the most work. Without it, the agent defaults to summarizing everything, which is exactly the problem the agent exists to solve.
Tested before trusted
The agent stayed private during testing. After a validation run, I ran a structured two-week pilot against five quality checks, each with a defined fix if it fell short. The pilot is complete, and the agent is now part of my daily routine.
| Dimension | Quality check | If not working |
|---|---|---|
| Signal vs. noise | Did it surface useful, actionable items? | Add stronger prioritization filters |
| Accuracy | Did it correctly identify ownership and source? | Strengthen evidence and attribution rules |
| Prioritization | Were the top items truly the most important? | Refine the priority logic and signal hierarchy |
| Meeting prep | Was the prep specific and actionable? | Narrow the meeting look-ahead window |
| Wednesday value | Did the midweek review catch slipping items? | Keep, simplify, or remove if redundant |
From personal tool to team practice
- System
A system, not a prompt
Inputs, prioritization logic, guardrails, structured output, and a pilot feedback loop.
- Prompting
Prompt engineering
Explicit inclusions, explicit exclusions, and a fixed output structure, written for reliable repeat runs.
- Rollout
Build vs. deploy
Private testing, a validation run, and a structured pilot before relying on it.
- Enablement
Enablement
Presented as an AI segment at a team all-hands, with a handout so colleagues could build their own.
Where it landed
- Two-week pilot completed against five defined quality checks
- Agent in daily use, plus a separate midweek follow-through review
- Briefed the team on the agent and its capabilities at an all-hands
- Build-your-own handout turned one person's workflow into a repeatable team pattern
What this project required
Core
AI agent design · Prompt engineering · AI enablement · Pilot design and evaluation
Supporting
Knowledge source scoping · Accuracy guardrails · Microsoft 365 Copilot · Presenting AI to non-technical audiences