Every project-management tool I use has grown an AI assistant this year, and the pitch is always the same breathless promise: it'll spot the slipping dependency before you do, draft the status update nobody wants to write, chase the action items, forecast the finish date. Some of that is genuinely useful. Some of it is exactly the kind of confident automation that gets a project into trouble. Having run delivery for two decades and actually used these tools in anger, I want to give the honest split — not the vendor version, and not the reflexive-cynic version either.

1. The demo versus the Tuesday

AI project tools demo beautifully, because a demo is a clean dataset and a happy path. Real delivery is a Tuesday where the status field says "on track" and you know from the tone of a stand-up that it absolutely isn't. The gap between those two is the whole job. So I judge these tools not on the demo but on whether they help on the Tuesday — and the answer, it turns out, depends entirely on which part of the job you point them at. Aim them at the administration and they're a genuine gift. Aim them at the judgement and they'll hand you a confident, well-formatted wrong answer.

2. What it genuinely takes off your plate

Let me be generous first, because it's earned. The current tooling is legitimately good at the high-volume, low-judgement work that eats a project manager's week: reconciling status across half a dozen tools that never agreed, summarising a long thread that ran overnight while you slept, drafting the update or the meeting notes so you're editing instead of starting from a blank page, and flagging that a dependency's dates have quietly moved. This is exactly the industry pattern — AI strengthening the support work rather than replacing the leadership. On a distributed program, where the administrative overhead is brutal, reclaiming those hours is not a small thing. It's the difference between spending your day on the project versus on the paperwork about the project.

3. What it can't do, and shouldn't be asked to

Now the other half, because this is where projects get hurt. The AI does not know that the sponsor's cheerful email is actually a soft warning. It can't read the silence from a team member who's stuck and too proud to say so. It won't tell you which of two "green" reports is quietly lying, because both look green in the data. Prioritisation under genuine ambiguity, the conflict conversation that has to happen, the political read of who really needs to be in the room — these are the core of the job and they're precisely what the tool can't touch. The failure mode isn't that the AI refuses these; it's that it'll cheerfully generate a plausible answer to them, and a plausible answer to a judgement question is more dangerous than no answer at all.

The framing I've settled on: let AI run the administration of the project so you can spend more of the day on the judgement of it. Reverse that split and you've automated the easy 20% and abdicated the essential 80%.

4. Agentic PM and the trust problem

The newer promise is agentic — tooling that doesn't just answer when asked but acts on its own: browsing the project's context, flagging deviations, initiating a nudge when something slips, even self-correcting against a baseline. The capability is real and improving fast. The problem is trust, and it's a governance problem before it's a technology one. An agent that can act inside your project — reassign, message stakeholders, change plans — needs the same discipline I'd apply to any autonomous account: a named owner, a bounded scope, visibility into what it did and why, and a human in the loop for anything consequential. An agent that quietly chases a vendor with the wrong information, or reshuffles a plan on a bad inference, does damage at machine speed. Use the autonomy for the low-stakes chasing and reconciling; keep the consequential moves behind a human decision.

5. The data-literacy gap decides who benefits

Here's the quieter point that I think separates the teams who'll gain from this from the ones who'll be misled by it: you have to be able to challenge the AI's output. When the tool forecasts a finish date or flags a risk, the valuable PM is the one who can look at it and say "that's wrong, and here's why — it's not accounting for the acceptance testing the utility does in the field." The dangerous one takes the confident forecast at face value because it came from the system. The whole industry is waking up to this — the emphasis now is on building enough data literacy in the core team to interpret and challenge AI-generated forecasts, not just consume them. An AI project tool makes a strong PM stronger and a weak one more confidently wrong. It's an amplifier, and what it amplifies is your existing judgement.

6. How I actually use it

So, concretely: I let it handle the reconciliation, the summarising, the first drafts and the dependency-watching, and I treat every one of those outputs as a good junior's work — useful, time-saving, and requiring a review before it goes anywhere that matters. I keep the forecasting on a short leash, always asking what assumption it's making that I know to be wrong. And I keep the judgement, the difficult conversations and the political reads firmly with the human, because that's the part I'm actually paid for and the part no model has earned the right to do. Used that way, AI has genuinely given me back time on delivery. Used the other way — as a reason to think less — it's a fast route to a confidently-managed failure. Pick the split deliberately. If you're introducing AI tooling into a delivery team and want to get that split right, let's talk.