AI in project management: what actually works
Most published advice on AI for project managers is either breathless or dismissive. Neither is much use when you are running a programme on Monday.
Here is the distinction that has held up in practice.
What AI is genuinely good at
Compression. Many inputs, one summary. Forty status updates into a draft report. A two-hour workshop recording into decisions and actions. Nine months of a risk register into themes. This is the strongest use and the one to start with.
Drafting. A first version of anything — a report, a business case section, a stakeholder email, a set of interview questions. The draft is rarely what you send. It is reliably better than a blank page, and it is faster to edit than to write.
Breadth. Generating candidates you might not think of. What risks might we have missed? What questions should I ask about this architecture? What could go wrong with this cutover? AI is good at the long tail precisely because it has no stake in the answer.
What it is bad at
Judgment. It cannot tell you which of five risks actually matters here, because that depends on your organisation, your sponsor, and what happened last time.
Accountability. It cannot hold any. A report is credible because someone put their name to it and can be asked about it in the room.
Knowing what is politically true. Every programme has things everyone knows and nobody writes down. AI has no access to that, and it will produce plausible text that ignores it — confidently.
Numbers it has no basis for. Ask for probability and impact scores and you will get them, formatted beautifully, invented entirely. This is the most dangerous failure mode because the output looks exactly like analysis.
The rule
AI drafts. A named human decides and signs.
Every failure mode in this series is a version of crossing that line. It is not a productivity constraint — reviewing a good draft is still far faster than writing from scratch. It is what keeps the output trustworthy.
The corollary matters as much: if you would not have written it, do not send it. A report you did not read is not your report.
Classify your data before you start
Project work spans a wider range of data sensitivity than people assume.
| Data | Sensitivity | Sensible handling |
|---|---|---|
| Plans, milestones, delivery status | Commercial | Ordinary care |
| Architecture, technical documentation | Commercial, sometimes higher | Check IP and NDA terms |
| Risks and issues | Commercial, sometimes reputational | Care with anything about vendors |
| Vendor and contract material | Contractual | Check confidentiality terms first |
| Anything naming individuals | Personal data | EU-hosted tooling, lawful basis, abstract where possible |
| Performance, conduct, conflict | Personal, potentially special category | See the coaching guide before doing anything |
The last two rows are where organisations get into genuine trouble, and where most AI-for-PM advice is silent. Both are covered in their own guides in this series.
For anything in those rows, an EU-hosted model is the baseline, not a nice-to-have — see Mistral instead of Claude and ChatGPT for how to route by workload rather than standardising on one provider.
The trap nobody warns you about
AI makes producing documents nearly free. The instinct is to produce more of them.
That is precisely backwards. Governance fails from too much reporting and too few decisions, and cheap generation makes the existing problem worse at speed. The reporting pack grows, the signal thins, and the steering group reads even less of it than before.
Use the time saved to shorten the report, not to lengthen it. If AI halves the time it takes to produce your status pack, the correct response is a two-page pack produced in a quarter of the time — not a forty-page pack produced in half.
This series
- AI for status reports — the best starting point
- AI for risks and issues — as a challenge function, not an oracle
- AI for stakeholder analysis — where personal data begins
- AI for team coaching and conflict — the highest-risk use, handled carefully
- AI for technical insight — for non-technical PMs on technical programmes
- Staying relevant as a project manager — the honest career question
Start with status reports. It is the highest-volume, lowest-risk task in the discipline, and getting it right teaches the review discipline everything else depends on.
Frequently asked questions
- Will AI replace project managers?
- It will replace a substantial part of what many project managers spend their time doing — compiling status, writing minutes, updating plans, drafting documents. It does not replace the part that was always the job: making decisions happen, holding accountability, and navigating an organisation. Which of those describes your role determines whether this is a threat.
- What is the single biggest mistake?
- Sending AI output onward without reviewing it. A status report that reaches a steering group unread by its supposed author is worse than no report, because the group acts on it believing a human stood behind it. The signature is the product; the draft is not.
- Is it safe to put project data into an AI tool?
- It depends entirely on the data. Plans, architecture and delivery status are usually commercially sensitive but not personal data. Anything naming individuals and describing their behaviour, performance or relationships is personal data under GDPR, and needs an EU-hosted tool and a lawful basis at minimum. Classify before you paste.
- Where should we start?
- Status reporting. It is the highest-volume, lowest-risk, most obviously compressible task in the whole discipline, and it is where the time saving is immediate and measurable. Get that working before touching anything involving people.