How I started building my AI project team—one agent at a time
In my last article, “Meet My AI Project Team,” I introduced some of the AI agents I’m building to help me manage projects.
But there is a big difference between saying:
“I should build an AI agent for that.”
and actually building one.
So let’s start at the beginning.
If I were hiring my first AI agent, which job would I give it?
For me, the answer is pretty obvious:
Email.
The Problem
Like most project managers, I spend a lot of time in Outlook.
Email is where customer requests arrive.
It’s where people ask questions.
It’s where decisions get documented.
It’s where action items appear.
It’s where deadlines get mentioned.
It’s where someone says:
“Just following up…”
And it’s also where things disappear.
Not intentionally.
They simply get buried.
A project manager can have dozens of emails arrive in a day across multiple projects.
The problem isn’t necessarily reading email.
The problem is determining:
What actually matters?
That’s the problem I want my first AI agent to solve.
Meet the Morning Briefing Agent
So let’s give my first AI employee a job title.
Morning Briefing Agent
Its mission:
Review my relevant new information and provide me with a concise, prioritized briefing of what requires my attention.
Notice what I didn’t say.
I didn’t say:
“Read my email and tell me what’s in it.”
That’s too broad.
Instead, I’m defining a specific business outcome.
Help me start my day knowing what matters.
That’s an important lesson when building AI agents.
Start with the problem, not the technology.
Step 1: Define the Job
Before building anything, I need to define the agent’s responsibilities.
Its primary responsibilities might include:
Identify
- Important new emails
- Customer requests
- Questions requiring a response
- Action items
- Follow-ups
- Deadlines
- Special requests
- Potential issues
- Potential blockers
Organize
Group the information into useful categories.
For example:
🔴 Needs Attention
Things requiring action or a decision.
🟡 Follow-Up
Things that need monitoring or a response.
🟢 Information
Things I should know but don’t necessarily need to act on.
🚨 Potential Issues
Items that may affect project delivery.
Suddenly, I’m not asking AI to “summarize my inbox.”
I’m asking it to perform a specific project-management function.
Step 2: Define the Inputs
Every good project manager knows:
Garbage in, garbage out.
AI is no different.
I need to define what information the agent is allowed to use.
For this workflow, that might include:
- New Outlook emails
- Email subject
- Sender
- Date/time
- Email body
- Relevant conversation history
- Attachments where appropriate
- Project information available to the workflow
Eventually, the agent might also use information from other systems.
For example:
Outlook + Teams + Salesforce + SharePoint + Confluence
Now the agent has a much bigger picture.
But I don’t necessarily want to connect everything on Day One.
That’s another lesson:
Start with a manageable scope.
Step 3: Give the Agent Instructions
Now comes the part that sounds a little strange.
I have to tell my AI employee how to do its job.
Not unlike onboarding a new team member.
I need to explain:
- What its role is
- What it should look for
- What it should ignore
- How it should prioritize
- What format I want
- What it should do when information is unclear
- What it is allowed to do
- What it is not allowed to do
For example:
Role:
You are an AI project-management assistant responsible for helping identify information requiring the project manager’s attention.
Objective:
Review relevant new information and identify actions, requests, deadlines, follow-ups, risks, issues, and important updates.
Important rule:
Do not invent information. If something is unclear, identify it as unclear.
That’s the beginning of the agent’s operating instructions.
Step 4: Teach It How I Think
This is where the project-manager part becomes really important.
Not every email that contains the word “urgent” is actually urgent.
Not every customer question is a project issue.
Not every request needs an immediate response.
Context matters.
For example:
“Can you send me the latest project timeline?”
That could simply be a routine request.
But:
“Can you send me the latest project timeline? We need to review the schedule because the implementation date may need to move.”
That’s different.
Now there may be a schedule risk.
The agent needs to recognize the difference.
This is where good instructions, examples, business rules, and context become important.
Step 5: Create the Output
I don’t want a 10-page AI-generated essay about my inbox.
I want something I can scan in a couple of minutes.
So I might define an output like this:
Morning Project Briefing
🔴 Requires My Attention
- Item
- Why it matters
- Recommended next step
📋 Action Items
- Action
- Project
- Due date
- Source
🔄 Follow-Ups
- Person
- Topic
- Last communication
- Suggested follow-up
🚦 Risks / Issues / Blockers
- Issue
- Project
- Impact
- Recommended action
📅 Upcoming
- Meetings
- Deadlines
- Milestones
ℹ️ FYI
- Important information requiring no immediate action
Now the AI has a job.
It has inputs.
It has rules.
And it has an expected output.
We’re getting somewhere.
Step 6: Add the Human in the Loop
This is where I draw a very important line.
My AI agent can identify an action.
It doesn’t automatically get to take the action.
It can say:
“Customer requested an updated implementation schedule.”
That’s useful.
But I’m still the person who decides:
Yes, I’ll respond.
Or:
No, that’s already been addressed.
Or:
This needs to go to the implementation team.
Or:
This is actually a project issue and needs escalation.
The agent assists.
I decide.
Step 7: Connect the Workflow
Now we get from AI assistant to AI workflow.
Imagine this happening every morning.
7:00 AM
The workflow starts.
↓
Check relevant new Outlook information.
↓
Identify important messages.
↓
Extract potential actions.
↓
Identify follow-ups.
↓
Look for deadlines.
↓
Identify potential risks, issues, and blockers.
↓
Group information by project.
↓
Prioritize.
↓
Create Morning Project Briefing.
↓
Deliver briefing to me.
Now I don’t have to remember to ask:
“Hey AI, what happened while I was away?”
The workflow is designed to do the work as part of my normal process.
That’s the difference between using AI and building AI into the workflow.
Step 8: Test It
This is where my project-management brain kicks in.
I wouldn’t release a new software feature to a customer without testing it.
I’m not going to do that with an AI agent either.
I’ll give it real-world examples.
What happens when:
- An email contains multiple action items?
- A deadline isn’t explicitly stated?
- Someone says “ASAP”?
- A customer asks a question that isn’t actually my responsibility?
- An email contains conflicting information?
- An action has already been completed?
- An email is simply FYI?
- A customer is expressing frustration?
- An issue is implied rather than explicitly stated?
The goal isn’t to make the agent perfect.
The goal is to make it reliable enough to be useful.
Step 9: Measure the Results
This is another area where I think project managers have an advantage.
We measure things.
So I’m going to measure this.
Before the agent:
How much time do I spend reviewing and organizing morning email?
After the agent:
How much time do I spend reviewing the briefing?
Then I compare.
Maybe I save 10 minutes.
Maybe 20.
Maybe more.
But time saved isn’t the only metric.
I can also measure:
- Actions identified
- Follow-ups identified
- Missed items
- False positives
- Accuracy
- Time saved
- Items requiring human correction
Because the goal isn’t:
“I built an AI agent.”
The goal is:
“The AI agent improved the way I work.”
Step 10: Improve It
The first version won’t be perfect.
And that’s okay.
In fact, I expect it not to be.
Maybe it flags too many emails.
Adjust the rules.
Maybe it misses certain types of customer requests.
Add examples.
Maybe the output is too long.
Change the format.
Maybe it doesn’t understand a particular project term.
Give it more context.
This is exactly how we improve processes.
Build → Test → Learn → Adjust → Repeat.
Agile thinking works surprisingly well here.
Then Hire the Next Agent
Once the Morning Briefing Agent is working, I don’t stop.
I look for the next repetitive task.
Maybe it’s:
Status reports.
So I build a Status Report Agent.
Then:
Handover notes.
Build a Handover Agent.
Then:
Customer emails.
Build a Communication Agent.
Then:
Requirements.
Build a Requirements Agent.
Then:
Issues and blockers.
Build an Issue Agent.
And eventually…
These agents can start working together.
From Individual Agents to an AI Team
This is where the idea gets really interesting.
The Morning Briefing Agent might identify:
“Customer is waiting for an updated implementation timeline.”
The Action Agent can turn that into:
Action: Update implementation timeline.
The Reminder Agent can identify:
Follow up Friday.
The Status Report Agent can recognize:
Potential schedule concern.
And the Project Analyst can eventually tell me:
“The timeline request may indicate an emerging schedule risk. Consider reviewing the implementation milestones before responding.”
Now the agents aren’t just performing isolated tasks.
They’re contributing to a larger workflow.
That’s the AI team I’m building.
The New Definition of Delegation
For years, project managers have been taught to delegate.
We delegate tasks to our teams.
We assign responsibilities.
We establish deadlines.
We monitor progress.
We review results.
I think we’re entering a world where we can also delegate certain digital tasks to AI agents.
But the same principles apply.
You wouldn’t give a new team member a critical task with no instructions, no context, no boundaries, and no way to measure success.
So why would we do that with AI?
AI agents need good project management too.
I’m Not Hiring AI to Manage My Projects
I’m hiring AI to help me manage my time.
That’s the distinction.
The projects still need people.
Customers still need conversations.
Teams still need leadership.
Problems still need judgment.
Decisions still need accountability.
But if AI can take care of some of the repetitive information processing happening around those activities…
I get time back.
And that’s the real value.
One Agent. One Problem. One Workflow.
That’s my approach.
I’m not trying to create a giant autonomous AI machine that takes over my job.
I’m starting much smaller.
Find a problem.
Define the job.
Give the agent the right information.
Set clear instructions.
Define the output.
Add human oversight.
Test it.
Measure it.
Improve it.
Then hire the next agent.
Eventually, those agents become a team.
And that team becomes part of how I manage projects.
The Future Project Team
I can already imagine where this could go.
I start my morning.
My AI team has already been working.
The Morning Briefing Agent has reviewed new information.
The Action Hunter has identified commitments.
The Follow-Up Detective has found outstanding responses.
The Issue Watchdog has identified potential concerns.
The Information Librarian has gathered supporting information.
The Project Analyst has helped organize everything.
And I sit down with one question:
“What needs me today?”
That’s the future I’m working toward.
Not replacing the project manager.
Not removing the human from the process.
But creating an intelligent team around the human.
Because maybe the biggest productivity opportunity isn’t getting AI to do more work.
Maybe it’s getting AI to do the work that allows me to do better work.
And that’s why I’m building my AI Project Team.
One agent at a time.
