Why Your Day Gets Hijacked Before 10am
Microsoft's Work Trend Index studied how office work actually happens, and the numbers are worse than most people assume. The average worker receives 117 emails a day and is interrupted roughly every two minutes - around 275 times in a working day, before you count anything on your phone.
The cost is not the minutes. It is that your attention gets cut into pieces too small to do senior work in. Strategy, judgement, difficult writing - none of it fits into a four-minute gap between Teams pings. Most productivity advice answers this by asking you to be more disciplined. A coordination agent answers it differently: it reads the inbox, checks the calendar, and hands you a single view of the day before the day starts deciding for you.
What a coordination agent actually is. An assistant you have connected to your email and calendar, given a standing brief, and told exactly where its authority stops. It reads, sorts, drafts and flags. You decide.
The Five-Part Prompt Behind an AI Agent for Email and Calendar
The prompt is the whole build. Not the tool, not the settings - the instruction you write. Every reliable one has the same five parts, and you will reuse this shape for everything you build afterwards.
The boundary is the part people leave out, and it is the part that matters. An agent with no stated limit does the most complete version of the task it can imagine, which on a Monday morning might mean replying to a client in a tone you would never use.
The categories matter more than they look. Three buckets force a judgement call on every message. Ask for a summary instead and you get a shorter inbox, not a decision about what deserves you.
Start With Email. Add the Calendar Later.
Run the email brief for a week or two. Notice where it misjudges urgency and adjust the wording until it stops. If you would rather have rules doing the sorting before an agent ever sees the inbox, the guide to automating your email in an hour covers that route.
Only then connect the calendar. This is where it stops being a summariser, because the two sources start talking to each other.
What comes back is different from a to-do list. It might tell you the 2pm steering group has an unanswered question from the finance director sitting in your inbox since Thursday, and that both hours before it are already booked. You would otherwise have found that out at 1:55pm.
For the background on what separates this from a chatbot, the post on what actually differs between ChatGPT, Copilot and AI agents covers it.
Promote It the Way You Would Promote a Person
Nobody hands a new joiner the client relationship on day one. Same logic here - the failure mode of AI at work is trusting it too early, not too late. Four stages, and skipping one is where people get burned.
1. Visible
It sorts and reports. You read every line.
Move on when its urgency calls match yours2. Efficient
It sorts and drafts. You edit before sending.
Move on when you are editing lightly, not rewriting3. Automatic
It runs at 7am. You review it with coffee.
Move on after a full week with no corrections4. Delegated
It acts on defined, low-risk items. You spot-check weekly.
Reversible actions onlyBe deliberate about stage four. Anything touching sensitive data, money, or a client relationship stays at stage two permanently. That is not caution for its own sake - it is where the judgement you are paid for actually lives.
Start This Today
- Connect your inbox, and read the permissions screen properly. Grant read and draft. Withhold send.
- Write the brief with all five parts, and save it somewhere you can paste from.
- Run it manually each morning for two weeks. Correct it out loud - the corrections are the training.
- Add the calendar and ask for conflicts, preparation needed, and what can wait.
- Once a full week passes with nothing to correct, schedule it for 7am.
When you want this running without you
Once the morning brief is reliable, the next step is connecting it to everything else - the CRM, the project tracker, the shared drive. Make.com is the no-code platform I recommend for that, because you build the workflow visually and it connects to over 1,500 apps. The free tier is enough to test the idea properly.
Try Make.com free →This post contains an affiliate link. I only recommend tools I use or would use myself.
Frequently Asked Questions
No. You are writing instructions in plain English and clicking a connect button. The skill involved is being precise about what you want, not writing code. If you can brief a new assistant, you can brief an agent.
Read the permissions screen properly before you accept it, rather than clicking through. Start with read and draft access only, so the agent can prepare replies but cannot send them. If you feel uneasy, that instinct is useful. Narrow the permissions until you do not.
Two to three weeks of daily use is a reasonable guide. You are looking for a stretch where you read the output and change nothing. Until that happens, run it manually. Scheduling an agent you do not yet trust just automates a mistake.
Many do, and that is a reasonable position for a regulated employer. You can still use the same method by pasting a morning export of subject lines and senders into the assistant, or by building it against your personal inbox first. The prompt structure works either way.
Where This Leaves You
The agent is the easy part. Five sentences and a connect button takes an afternoon. The harder question is what you are willing to hand over and what stays with you, and that does not get easier as the tools get better.
Start with the inbox, because it is the thing currently making the decision for you. If you are unsure where this sits on the wider curve, the level-by-level guide to getting good at AI places it in context.
One practical AI tip, straight to your inbox.
Practical AI agent ideas for professionals - no jargon, no hype. Join the Linda & AI newsletter.