AI won't manage your projects
AI won't manage your projects. It will write about them, sort them, summarize them and, if you let it, move their dates around. The hard part is a person saying no: dropping the feature that will not fit, telling a client the date has moved, closing work that is not paying off.
I run three products with AI agents doing a lot of the legwork, and I have not seen one do that part. I don't expect to.
The agent launches in project software this year promise something else: a new colleague who can run parts of the project.
What are the project tools promising this year?
The promise is that agents will join your team and run parts of the project for you. Each of the big vendors now pitches AI as a colleague rather than a feature, and two of them have named an agent after the job title itself.
- Asana's AI Teammates page opens with "Your team just got bigger". There are more than 30 prebuilt Teammates, and one of them is called Project Manager. Its listed skills include project scoping, timeline and dependency planning, and flagging blockers. Asana describes a Teammate as an agent "you assign work to like a colleague".
- Wrike announced Workmates at its Collaborate 2026 event: AI agents with their own identities in Wrike that you assign work to or @mention like a colleague. Wrike says it is building its own Workmates too, starting with a Project Manager one that suggests task owners from skills and capacity and assesses what threatens delivery. For now Workmates are in Wrike Labs, for Business plans and above, while they are tested.
- Atlassian's Rovo, with its chat and agents, is added to the paid cloud plans of Jira and Confluence, with no separate purchase.
- monday.com's agent page calls its ready-made and custom agents "your unlimited workforce".
The strongest case for agents is the admin they can take off your hands: turning a brief into a task list, chasing the status of twelve tasks, writing the same weekly summary for the third time. An agent that does that in a minute saves real hours. I use tools like them every day. Calling that agent a project manager is where I disagree.
What do AI teammates cost on top of the seats?
For three of the four, the agents are paid for by use, on top of what you pay per person. Here is what each vendor's own pages say:
- Asana's Starter plan is $10.99 per user a month billed annually and includes 5 AI requests per user a month, capped at 50 for the account. Extra requests cost $0.50 each prepaid or $0.60 on demand. Asana's help center counts a follow-up question as a request, and a request for a revision of a Teammate's work as another. The Asana request costs breakdown explains the charges.
- Atlassian switches Rovo to pay-as-you-go billing on December 3, 2026. Jira Standard includes 25 Rovo credits per user a month, pooled across the organization; beyond that, credits cost $0.01 each, or $10 per 1,000, and pay-as-you-go is on by default. The Rovo credit costs breakdown covers the allowances.
- monday.com accounts that joined the work platform from May 6, 2026 buy AI credits as part of their setup. monday.com's help center puts a simple agent task at about 10 to 50 credits and an extra complex one at 250 or more, and one agent run can include several tasks. The monday.com credit bundles set out the costs.
- Wrike's pricing page lists no separate price for Workmates while they are in testing.
None of these prices is outrageous. What bothers me is paying by the attempt. You pay a person for the job. You pay an agent for each attempt at it, including the attempt where you asked it to try again.
Why does the word "teammate" bother me?
A teammate can be held to a promise, and an agent cannot. Picture a teammate promising a page will be ready on Thursday. On Friday, they can be asked why it is not. They can say "I was wrong, and here is what it cost". If a client is waiting, they have to face that client, or sit next to the person who does.
An agent does none of that. It cannot be embarrassed, it does not lose anything when the estimate is wrong, and it does not remember Thursday on Friday unless someone wrote Thursday down for it. That is not a flaw to be fixed in the next version. It is what a tool is. Calling it a teammate borrows the accountability of a person for something that has none, and then bills you per attempt.
I would happily pay for "AI assistant" or "admin agent". The word "teammate" asks the buyer to treat software as someone who shares the responsibility.
What does a stalled project look like up close?
A stalled project can have clear tasks and current due dates while nobody makes the decision it needs. In October I had two content idea cards open in our own project account. The work on them had already been done, about 35 pages written from those ideas earlier in the year, and from July to September those pages brought in no signups.
The cards stayed open anyway. Their due dates moved from one week to the next, with the text unchanged and no comments.
There was nothing wrong with the tasks. Each card was clear and dated. The project had stalled on a decision that nobody had made: to stop. An agent that flags stale cards would have flagged them, and it would have been right. An agent that moves due dates would have moved them, which is exactly what was already happening. No assistant would have closed those cards, because closing them meant saying out loud that the work had not paid off.
Most stuck projects I have seen look like that. The board is tidy, the tasks are well written, and the real problem is a choice someone is avoiding. More updates and better summaries make that kind of project look healthier, not finish sooner. To understand repeated deadline slips, ask who owns the decision as well as the task.
Where has AI earned its place in my own work?
Almost everywhere except the decisions. Across the three products I run, AI agents do much of the legwork: drafting, coding, checking, pulling numbers out of reports. They are very good at it, and faster than I would be on most of it.
What they are bad at is knowing what a mistake costs. Two examples from our own work:
- An agent drafted the social posts for one of our release notes from roughly the first half of the note, and announced the wrong half of the feature. The section the release note itself pointed at first was in the part it never read.
- Another agent recommended merging one of our pages into another three times in a row. I said yes. The page it was merged into had been bringing in signups, and after the merge it fell from around position ten in Google to past fifty, and got no clicks for three weeks.
The agent was just as confident before the merge as after it. It did not lose anything. Picture asking it a fourth time. It would have made the same recommendation. I am the one who lost the traffic and the signups, because I am the one who said yes.
That is the job: carrying the cost when a recommendation is wrong. An agent can produce the recommendation. It cannot carry the cost.
Where should AI stop in a project?
At the decision. Everything before it is fair game, and a lot of what comes after it too. AI can:
- Summarize a long thread so the person who has to decide can do it in five minutes.
- Draft the status update from what actually changed this week.
- Flag the card nobody has touched in two weeks.
- Change a due date once someone has decided it moves, and update everything that hangs off it.
It should not decide priorities, decide that a deadline moves, or tell a client anything. All three decisions settle what we are not going to do, and who has to hear it. Pick what to drop and you have set the priority. Move the deadline and you have chosen what slips. Tell the client and you have made it real.
Even a brilliant agent that can decide your priorities cannot be the person the client calls when the priority turns out to be wrong. Ask who would apologise. That person makes the call.
What should a small team hand to AI tomorrow?
Hand it the boring writing. That is where the hours are, and it is where a wrong first draft costs almost nothing because someone reads it before it goes anywhere.
- Catching up on what changed while you were out, across tasks and comments.
- Turning meeting notes into tasks with owners and dates you can then correct.
- The weekly status summary, drafted from what actually moved rather than from memory.
- The first draft of a difficult client email.
Then a person reads it, decides, and sends it under their own name. Say a team needs to tell a client the launch will move two weeks. An AI draft helps that person prepare. The same words sent by an agent, with nobody's name on them, are a team avoiding a conversation.
Keep the tasks current before choosing an agent to read them. Every summary, draft and flag depends on what is on the board. Compare project tools with AI once the team does that reliably.
What does this view cost us?
It costs us speed and attention. Building AI this way means we ship it slower than the big vendors, and it is harder to put in a launch video. In Breeze, the AI built into the app is off until an admin turns it on, so some teams will never find it. A team that wants an agent with a job title will find several elsewhere, and some of them will be happy with that choice.
That is the right trade for the teams we build for: small teams where the person who decides is also the person the client calls. I accept the slower releases and the less impressive demo. A person still has to own the decision.
What will AI still not do in two years?
It will still not say no to a client on your behalf. The models will be better, the summaries sharper and the agents cheaper per attempt, and someone will still have to tell the client that the feature is out of this release, the date has moved, or the project should stop. That person needs to be someone the client can call back.
Use AI for everything around that moment. Just do not let a job title on a pricing page convince you that the moment itself has been automated. Saying no is most of project management.



