AI in project management: what actually works, and what is theatre

Most AI-in-project-management conversations are about the wrong thing. The question is not what a model can do. It is which of your senior hours it gives back, and what it costs you in supervision to get them.

Where it genuinely works

Four use cases survive contact with a real programme. They share a shape: high volume, low judgement, and a human who can spot a wrong answer in seconds.

• Status drafting. A weekly report assembled from the tool, the risk register and last week’s actions. The manager edits rather than writes. This is the single largest time saving available.
• Meeting capture. Transcription plus extraction of decisions, actions and owners into the log. The value is not the transcript; it is that the decision log stops decaying.
• Document and correspondence search. On a programme with thousands of emails and drawings, being able to ask what was agreed about a delivery date and get the three relevant documents is worth more than any dashboard.
• Signal surfacing. Delay and risk patterns in data the team already produces: tasks slipping in the same workstream, actions repeatedly re-dated, the same risk reworded three times.

Where it does not

Two uses look attractive and consistently disappoint.

• Estimation. A model will produce a confident duration from insufficient history. The confidence is the problem: it is indistinguishable from a good estimate until the milestone arrives.
• Autonomous decisions. Anything that changes a plan, a commitment or a supplier relationship without a human. The supervision cost exceeds the saving, and the accountability does not transfer.

The governance nobody sets up

The failure mode in practice is not a wrong answer. It is an unreviewed one that entered the record and then got quoted back six weeks later as fact.

• Define which outputs require human review before they leave the team, and make that visible in the tool rather than in a policy document
• Keep client and commercial data out of tools you have not cleared. The fastest way to lose an AI programme is one confidentiality incident
• Log what was AI-drafted, so a future reader knows what was written and what was generated
• Measure the time actually saved. If nobody can state it after a quarter, it was theatre

How we use it on engagements

We build AI into the workflow where routine work is consuming senior hours, and leave it out where it would only add a supervision burden. The judgement and the signature stay human, because that is the part a client is paying for. The speed and the cost advantage are real, but they are not the differentiator.