To use AI for project management, give it the chasing. Have it check each project for client requests nobody answered, meeting times that ran out and tasks stuck on one person, and tell you every morning what to do. At VantaSoft, our project agent, Reid, does this for our client projects. People still send every client message and decide every date.
Reid started out reading client email and nothing else. A missed customer request in October showed us that the useful work is noticing what never arrived, so we added a morning check for it. Here is his week.
| When (Pacific time) | What he does | What reaches an owner |
|---|---|---|
| Every 30 minutes | Reads new email from each client with a live project and updates that project's board | A note on anything an owner should look at |
| Every 30 minutes | Works the project tasks assigned to him, and skips any task that is waiting on a person | A task handed back to whoever has to decide, with what he needs from them |
| Weekdays, 8:05 a.m. | Checks every client project for slipping dates, unanswered requests, expired meeting times and stuck tasks | One short report with the owner's moves first, or nothing on a quiet day |
| When a customer pays the setup fee | Opens that customer's project and its first tasks | One line in his report |
| Every morning, 6:15 | Archives tasks that were finished more than a week ago, so the active list stays short | Nothing |
From October 30 to November 10, 2026, his client email check finished 490 times. The morning check ran on all 20 weekdays from October 14 to November 10. On one of those mornings nothing needed a person, so he posted nothing.
How do you use AI for project management?
Start with the chasing. Write down the date that matters on each project, what counts as at risk, and what the AI may change by itself. Then let it read the places where waiting hides, which are your sent email, your project board and your calendar. Have it check its facts before it tells you anything.
Here is the order we would follow again.
- Write down the date that matters for each project: a launch, a delivery, or a meeting the work has to be ready for. Ours often sat in a comment or in the project notes, not in a date field, so Reid reads both.
- Write down what "at risk" means, in numbers. Our five rules are below. Without them an AI flags everything or nothing.
- Point it at what has not happened yet. An email tool sees mail that arrives, but most delays are mail that never comes, like a request the client has not answered or meeting times nobody picked. Reid reads our sent email for requests with no reply after two business days.
- Make it check before it tells you. Before Reid lists a request as unanswered, he looks for the answer somewhere else: another email thread, a calendar reply, a note on the task or a signed document.
- Decide what it may change by itself. Ours updates tasks, adds notes and closes a task once its written finish line is met. Client messages, meetings, prices, scope and launch dates stay with an owner.
- Keep the report short. Each item is one line of about 25 words, and the owner's moves come first, nearest deadline first. When nothing needs a person, he posts nothing.
What does an AI project manager do on its own?
Ours keeps the project board current, collects status and decides what is at risk. He does not talk to clients. Anything that commits the company, such as a reply, a meeting, a price or a launch date, goes to an owner as a one-line ask: who, what, how long it has waited and what it holds up.
| The job | Who decides |
|---|---|
| Update a task, and add a note the first time it starts waiting on someone | Reid |
| Open a new customer's project and its first tasks after the setup fee is paid | Reid |
| Close a task once its written finish line is met, with a note on what proved it | Reid |
| Mark a project complete once the customer is live | Reid, on proof, and he tells an owner |
| Say whether a date still holds | Reid says it plainly, and an owner decides what to do |
| Write or send a reminder to a client | An owner |
| Book or move a client meeting | An owner |
| Agree to a price, a change in scope or a launch date | An owner |
| Send finished work to quality testing | An owner approves each one |
| Drop an item, or pause it until something happens | An owner, and Reid records it and stays quiet about it until then |
Marking a project complete is the only change to a project's status he makes by himself, and only when there is proof the customer is using what we delivered. Other work starts from that status, so it has to be true.
When we build an AI project manager for another business, the role page lists what that job covers and where it stops. Across the whole company, our chief of staff agent does the same kind of follow-through for leadership, as described in what our AI chief of staff does for three owners.
What counts as "at risk"?
We gave Reid five written rules. Anything that matches one goes in the morning report, and nothing else does. The numbers are ours. Change them to fit how fast your clients usually reply and how long your projects run.
| Rule | Our threshold |
|---|---|
| A request to the client with no answer | 2 business days |
| A target date with open work that has not started, is stuck on a person or waits on an unanswered request | Date within 14 days |
| A task stuck on the same person | More than 2 business days |
| An open task with no update while its project has a date | 5 business days |
| A due date that has passed | Any |
Meeting times we offered that have now passed count too, since the request has to go out again with new times. Reid works back from each date, lists the steps between now and then in order, and says whether the date still holds. One line in his instructions reads: "Do not soften a date that will not hold."
He does not repeat himself. An item that has not changed comes back after two business days, or sooner when it gets worse: the date moves inside seven days, a new blocker shows up, or the offered times run out. If the project is software you are paying someone to build, our guide on how to tell if a software project is off track covers the warning signs from the buyer's side.
What went wrong when an agent started watching our projects?
Here are five problems from his first months on this work, and the rule each one left us with. Our record of what went wrong with our AI agents has failures from other jobs.
The email check could not see silence
On October 8 we asked a new customer who would handle three setup steps, and offered three times for an intro call. Five days later neither email had an answer. Their soft launch was nine days away, and nothing had flagged it, because Reid's only job then was reading mail that arrived. On October 13 we added the weekday morning check, which looks for requests with no reply and meeting times that have run out.
The first morning check flagged work that was already done
Its first run, on October 13, listed four items as waiting that had already happened, two of them app installs that were already finished. An owner had to ask for a cleanup. Now Reid looks for the answer in other places before he calls a request unanswered. When there is no way to check, he writes "not verified" instead of "no reply."
Five lines could have been checked first
On November 6, five lines in the morning report could have been checked before they went out. A fix it listed had been live for two weeks. A data connection it flagged had been working every hour. One task's results were already recorded, another task's finish line had been met at that day's meeting, and a task the report called done was still open. An owner asked that anything checkable be checked before it is raised. Now every item is checked on the morning it is reported, repeats included, and a task whose finish line is met gets closed with a note saying what proved it.
It wrote emails nobody asked for
In August, one run of the client email check wrote out full reply copy that nobody had asked for. Since August 28 the check has a narrower job: keep the board current and tell an owner what to look at. The morning report names the reminder an owner should send in one line and leaves the writing to people.
A customer paid and no project opened
On October 22 a customer paid the setup fee, and no project existed until an owner asked why. Now the payment notice itself is the trigger. When one arrives, Reid opens the customer's project and its first tasks, then reads the project back to confirm it is right before calling it done.
It still slips now and then. On November 10 he posted a note that two of a customer's agents had not been installed, then checked and found both had gone in on November 4. He corrected the note in the same run.
What should you write down before AI watches your projects?
Write down what a new project coordinator would need on day one. The trouble we hit came from the things nobody had written down, like where the real date lived and what counted as an answer. Here is the sheet we would fill in first, with our own answers as examples.
| Write down | Our answer, as an example |
|---|---|
| The date that matters for each project | The launch date, the delivery date, or a meeting the work has to be ready for |
| Where that date lives | The project notes and recent comments, not only the date field |
| What counts as at risk | The five rules above |
| Where to look for silence | Our sent email to each client, the project board and the calendar |
| Where to look before calling something unanswered | Other email threads, calendar replies, task notes and signed documents |
| What it may change by itself | Task notes and status, and closing a task once its finish line is met |
| What only a person does | Client messages, meetings, prices, scope, launch dates and sending work to testing |
| How it reports | One line per item, owner's moves first, and silence when nothing needs a person |
| How you tell it to stop raising something | A waiver with who, when and why, or a pause with the condition that ends it |
Keep the sheet with your project notes, and add to it each time the AI asks you something new. Our guide to keeping a human in the loop with AI agents covers how to set approval lines like these for other kinds of work.
What are your options for AI in project management?
There are three common routes. You can use the AI features inside your project tool, paste notes into a chat assistant, or have an agent set up to work across your email, board and calendar on a schedule. They differ most in how much of the chasing they do and how much of your project they can see.
Prices are as each company published them on November 10, 2026.
| Option | What it does | Price |
|---|---|---|
| AI in your project tool, such as Asana | AI Teammates on paid plans; on Advanced, AI Meetings suggest action items and who should own them | Starter $10.99 per user a month billed yearly; Advanced $24.99 |
| A chat assistant you paste notes into | Drafts status updates, agendas and meeting summaries from what you give it | Included in a chat plan you may already have |
| Your own agent, which we build | Works across your email, board and calendar on a schedule, under your written rules | $1,000 to build, then $200 a month |
Asana's monthly billing is $13.49 per user for Starter and $30.49 for Advanced, and both plans include 5 AI Teammates requests per user a month, up to 50 per account. Our prices are starting prices. We sell the third option, so weigh our view with that in mind. If your projects live in one tool and there is little back-and-forth with clients, the AI in that tool may be enough. An agent is worth more when the waiting happens in email and calendars, outside the board, as it did in our October miss.
Frequently asked questions
Will AI replace project managers?
In 2019, Gartner predicted that 80% of today's project management tasks would be eliminated by 2030 as AI takes over (Gartner press release, March 20, 2019). In our company, the tasks that moved to an agent were collecting status and chasing answers. Decisions about dates, scope and what to tell a client stayed with people.
How are project managers using AI today?
Mostly for writing and summing up. In the Project Management Institute's spring 2024 survey of 500 project professionals who already use generative AI, 43% said they use it for more than half of their project task execution. PMI found it used heavily to summarize project documents and reports, compile executive summaries and condense meeting discussions (Pushing the Limits, PMI).
Can ChatGPT be my project manager?
A chat assistant is good at drafting an update from notes you paste in. Noticing that a client never replied takes something that reads your sent email and your board every day, under rules you wrote, and that is the part we handed to an agent.
How much does an AI project manager cost?
It depends on the route. Asana's AI comes with its paid plans, starting at $10.99 per user a month billed yearly. A custom project agent from us starts at $1,000 to build plus $200 a month for hosting and maintenance (see pricing). Plan for your own time as well, since someone still has to act on the morning report.
What should an AI project manager never do on its own?
Ours never emails a client, books or moves a meeting, or agrees to a price, a change in scope or a launch date. It also never sends work to testing without an owner's say. Start any agent with read access to one project, and widen what it can change one job at a time.
Where to start
Pick one project with a date coming up. Fill in the sheet above for it, give the AI read access to that project's board and your sent email to that client, and compare its morning report with your own notes for two weeks. Our 13 real AI agent examples show where a project agent fits on a full team, and our guide to hiring an AI agent has a job description you can adapt. When you are ready, Start Automating and we will walk through your projects with you.
