Why start here
AI assistants such as ChatGPT, Gemini, Claude and Microsoft 365 Copilot can save a project manager hours each week. They can also put a confident, wrong sentence into a sponsor report. This level gives you a practical map. Some tasks AI drafts well. Some it can help with while you make the decision. Some must stay with you. You will sort real tasks, learn the three ways AI output most often goes wrong, and look at dated evidence of how the profession is using AI.
Think of an AI assistant as a capable new project coordinator on their first day. They write quickly and fluently. They know nothing about your project, your sponsor or last month's steering group, and they will not tell you when they are guessing.
Task 1: Sort your tasks into three categories
Use these three categories:
- AI drafts it: AI writes the first version from material you give it. You check and edit it. The risk is low because you review everything before anyone sees it.
- AI assists, human decides: AI suggests options, gaps or challenges. The judgement and the final answer are yours.
- Human only: The task carries accountability, authority or a relationship that cannot be handed over. AI may help you prepare, but the act itself is yours.
Worked example: 10 everyday tasks
| # | Task | Category | Reason |
|---|---|---|---|
| 1 | Turning rough meeting notes into tidy minutes | AI drafts it | Restructuring text you supply is a strength. Check every action and owner. |
| 2 | First draft of a status report narrative from your notes | AI drafts it | Saves writing time. The facts must come from you, not the model. |
| 3 | Rewording a stakeholder email into plain English | AI drafts it | Tone and clarity editing is low risk when you check the result. |
| 4 | Brainstorming risks for a pre-mortem | AI drafts it | A broad first list helps. The team then filters and owns it. |
| 5 | Suggesting a work breakdown structure from a scope statement | AI assists, human decides | AI does not know your organisation's deliverables or exclusions. |
| 6 | Sanity-checking task estimates | AI assists, human decides | It can flag estimates that look optimistic. It has no data on your team's actual speed. |
| 7 | Proposing a RAG (red, amber, green) status | AI assists, human decides | The rating is a judgement you must be willing to defend. |
| 8 | Approving or rejecting a change request | Human only | This is a decision made under delegated authority. Accountability cannot move to a tool. |
| 9 | Accepting a risk on the sponsor's behalf | Human only | Accepting a risk commits the organisation. It needs a named person. |
| 10 | Having a difficult conversation with an under-performing supplier | Human only | Trust, context and the relationship depend on you. |
Do this now
- List 10 tasks from your own last two weeks of work.
- For each one, ask: Do I supply all the facts? If yes, it is a candidate for 'AI drafts it'.
- Ask: Does it need knowledge of my project that the tool does not have? If yes, move it to 'AI assists, human decides'.
- Ask: Does it commit money, scope, risk or a relationship under my authority? If yes, mark it 'Human only'.
- Write a one-line reason beside each task. If you cannot give a reason, you have not finished sorting it.
Task 2: Recognise the three failure modes
AI assistants produce fluent text by predicting likely words. That causes three typical failures.
- Made-up facts (hallucination). The model invents a figure, date, name or reference that sounds plausible.
- Where it harms: a status report that says '85% of testing complete' when your notes gave no such figure.
- Missing context. The model knows only what you told it, so it misses constraints that are obvious to you.
- Where it harms: a draft schedule that ignores a 12-week supplier lead time or a change freeze over year-end.
- Overconfident tone. Wrong or uncertain output reads exactly as certain as correct output.
- Where it harms: a stakeholder communication or business case that says 'the project will deliver on time' when the real position is 'on time if the design is approved by Friday'.
Do this now
- Take one AI draft you have produced. Highlight every number, date and name. Trace each one back to a source you trust.
- Look for sentences that sound certain. Add the conditions or assumptions that the model left out.
Task 3: Read the evidence and date it
Published surveys show how quickly the profession is adopting AI:
- APM and Censuswide, published September 2025: of 1,000 UK project professionals, 70% said their organisation uses AI and another 29% planned to adopt it. In APM's 2023 survey the figure was 36%. In the same 2025 survey, 41% of respondents named inaccurate or untrustworthy AI output as a challenge.
- APM, published 31 March 2026: in a further survey of 1,000 project professionals, 27% said AI is fully embedded in their workflows.
- PMI, Pushing the Limits, published September 2024: of 500 project professionals, roughly two in five used generative AI in more than half of their project work.
- Gartner, March 2019: predicted that 80% of today's project management work would be eliminated by 2030. This is a forecast, not a measurement.
Why these numbers may already be out of date
- Tools and features change monthly. A 2024 survey describes a different set of tools from today's.
- Surveys collect answers months before they are published.
- Samples differ. The APM figures are UK-only, and PMI's are global.
- The questions differ. 'My organisation uses AI' is not the same as 'I use AI on most projects'.
- The answers are self-reported, so they measure what people say they do, not what they actually do.
Do this now
Whenever you quote a statistic, give the source, the publication date and the sample. For example: 'APM, September 2025, 1,000 UK project professionals: 70% of organisations use AI.'
Key takeaways
- Sort each task into AI drafts it, AI assists, human decides or human only, and give a reason for each.
- AI is strongest when you provide the facts and it provides the wording.
- Watch for made-up facts, missing context and overconfident tone, especially in status reports, schedules and stakeholder messages.
- Decisions that commit scope, money, risk or relationships stay with a named person.
- Always quote a statistic with its source, date and sample, and treat it as a snapshot rather than the current position.