AI agents are joining project teams. They write code, draft documents, open pull requests and post updates, and they do it around the clock. In McKinsey's 2025 State of AI survey, 62% of organizations said they are at least experimenting with agents. Gartner predicts that by 2028, at least 15% of day-to-day work decisions will be made autonomously by agentic AI, up from 0% in 2024.
There's a popular story that this makes the project manager less relevant. I think it's the opposite, and it's the case I made last week to the Master Projects User Group (MPUG), a vibrant community of project management professionals: agents multiply the work, and the project manager is the one who holds it together.
And when I say "project manager", I mean it broadly. That includes the people with the title, but also the engineers, marketers and ops leads who get handed a project on top of their day job. If you're the one keeping the work moving, this post is for you.
Before we add agents, it's worth being honest about where we're starting. Oxford professor Bent Flyvbjerg and his colleagues studied some 16,000 major projects. Only 8.5% were delivered on time and on budget. Just 0.5% did that and also produced the benefits they promised.
When you look at why projects miss, the same three causes come up again and again. None of them is the quality of the plan or the people. They come down to information and coordination.
Problems surface too late. The Project Management Institute (PMI) found that 56% of project dollars are at risk because of ineffective communication. Status drifts green until the review. Most of us have seen a "watermelon" scorecard or two: the ones that look green on the outside but hide the red underneath. The problem was usually visible to someone, just not to the people who could act on it.
Decisions stall. Decisions get made in meetings, everyone nods, and then nobody writes down who owns the next step. Projects slip in the gaps between meetings.
Coordination can't keep up. Asana's Anatomy of Work research surveyed more than 13,000 knowledge workers. It found that about 60% of their time goes to "work about work": chasing status, hunting for the latest version, reconciling spreadsheets.
One consumer brand we work with tracked the same milestone dates in Airtable and in Smartsheet, then re-keyed them into SAP by hand. Eventually they gave up and told every team to track its own dates. Nobody was doing anything wrong. There was simply more to keep in sync than any one person could manage.
So what happens when you add agents to that mix? In my experience, and the early research backs this up, agents don't fix these problems. Left unmanaged, they amplify all three.
Signal gets buried. With people, the problem was often that nobody told you. With agents, everything tells you, all the time. The one flag that matters gets lost in the volume. And a wrong answer from an agent is formatted just as nicely as a right one.
Accountability blurs. Agents can do the work, but they can't own the outcome. As far as I know, nobody has fired Claude for a bad pull request. When an agent hits a decision it can't make, it waits, or worse, it guesses. Either way, the decision bottleneck gets tighter.
The plan drifts at machine speed. Agents only know what they're told. If the plan is stale, they act on old context with complete confidence, and they start pulling in a different direction from the people they work alongside.
Strip away the frameworks and certifications, and I think anyone running a project does three things. Agents don't change those jobs. They make each one harder.
1. Connect and listen. Stay on top of what's happening across meetings, email, chat, documents and trackers. This is already where most of the time goes. With agents posting around the clock, reading everything stops being humanly possible.
2. Create and maintain the plan. Creating the plan was always the easier part. Keeping it current is where things slip. With agents changing code, documents and scope as they go, that gets much harder. It also matters more, because the plan is now the context agents work from. A stale plan doesn't just mislead people. It misdirects agents.
3. Communicate and act. Report status, flag drift, chase follow-ups, escalate risks. Now there are two audiences. People need to know what the agents did and what needs human judgment. Agents need clear goals, constraints and current context. If there's too much noise and not enough signal, both audiences tune out.
Whether agents are already on your team or still on the horizon, three habits go a long way:
I'll be candid: these habits are easy to agree with and hard to keep up. I often liken project management to dieting and exercise. We all know it's important, but… Each of the three jobs is getting a lot bigger, and I don't believe project managers scale by working harder.
The answer, I believe, is to use AI on the project management work itself. Let it take the toil (the listening, assembling and reconciling) so people can focus on judgment.
This is the problem I spent years on. I worked on Microsoft Project for 17 years, then ran product for Atlassian's Server and Data Center business. We built good tools. But the dashboards were only as good as the data people remembered to type in. Garbage in, garbage out. AI finally changes that, because it can sit in three meetings at once and read every Slack or Teams thread without getting tired.
That's how we run our own team at Empwr. A teammate recently raised a feature request in our Slack feedback channel. Nobody asked anyone to log it. Empwr picked up the conversation, suggested a task and linked it back to the thread it came from. From there, someone could accept it or dismiss it. That's a scope change that would normally slip through the cracks, caught while it was still a conversation.
Empwr also builds and maintains the project timeline and goals, and flags risks as it sees them, such as capacity constraints or problems with tasks on the critical path. When it flags a risk, Empwr notifies the right person in Slack with the context they need, so the project keeps moving forward.
Two things matter in how this works. First, every risk, task and decision links to the meeting, ticket or message it came from, so you can check it rather than take it on faith. Second, people make the calls. Empwr takes the assembling, not the judgment. Customers tell us one of the things they value most is that it's neutral. Instead of everyone leaving a meeting with a different version of what was agreed, there's one factual record anyone can go back to.
That plan also becomes something your AI assistants can work from. Copilot, Claude and ChatGPT are very good at producing polished reports. The question is whether what's in them is true. A plan built from what actually happened is what makes their answers true.
All in all, I'm optimistic. Agents add real capacity, and teams that get coordination right will get a lot out of them. But someone has to direct the agents, own the outcomes and keep people and agents working from the same plan. That's the definition of the job. It's getting bigger, not going away.
We're still early, and six months from now this will have moved again. I'd love to hear how your teams are handling it. Who's accountable for an agent's work on your projects? And how do you keep agents from flooding people with updates?
If you want to see how this works on one of your own projects, you can start for free or book a demo.
See the results for yourself.
