AI in Project Management: What It Can and Can't Do

Generative AI in project management is good at the paperwork around a project and unreliable at the judgment inside it. Today’s tools draft plans, summarize meetings, turn notes into status updates and list risks hidden in long documents. They also invent facts, estimate poorly on unfamiliar work and carry no responsibility for what they produce. The best controlled evidence comes from AI at work in general: in a randomized experiment at Boston Consulting Group published in 2026, the same AI tool made some work better and some worse, on tasks that looked about equally hard.¹

For a project manager, the working rule follows from that. Give AI the jobs you can check against a source document in a few minutes, check them, and keep every decision someone will hold you to. Readers new to the discipline will find the basics of managing a project a better first stop.

How is AI used in project management? Four jobs it does well

AI in project management mostly means generative AI working on text: drafting documents, summarizing conversations, pulling risks out of long files and analyzing past project data. In a 2023 report, the Project Management Institute (PMI) sorts these jobs by complexity and calls for more human checking as tasks get harder.²

PMI’s framework has three levels. Simple, low-stakes work such as summarizing meeting notes or generating routine reports can be largely automated. Middle-level work, such as a first draft of a schedule or a risk analysis, needs moderate review by an experienced person. Complex work, such as a business case or a decision with many dependencies, stays with the project manager, who uses AI for ideas and analysis.² PMI is an association with an interest in promoting AI skills training, and the report is guidance built on its members’ experience rather than a study.

Drafting has the best evidence. In a 2023 randomized experiment published in Science, 453 college-educated professionals, including managers and consultants, did short writing tasks such as reports and delicate emails; those given ChatGPT finished in about 40 percent less time, and graders rated their work higher.³ A weekly status update is the same kind of task, so a similar saving is plausible there, though nobody we could find has measured it on project documents.

Meetings are where workers used AI most, and they show the limits of individual adoption. In a 2025 field experiment led by Microsoft researchers across 66 firms, workers given Microsoft’s own AI assistant used it most often inside Teams, Microsoft’s meeting app, where it can summarize meetings, yet their time spent in Teams meetings did not detectably change over six months.⁴ The authors suggest that changing how meetings run needs colleagues to agree new norms, a theme covered in adding AI to shared team tasks.

Risk spotting sits between the two. A model can work through a long contract or a year of change requests and list clauses or patterns that look risky, a job PMI counts as assistance that needs an experienced person’s review.² We found no controlled study measuring how many real risks such lists catch or how many false alarms they raise, so treat the list as a prompt for your own review. Risks that no document mentions still have to come from the team, for example in a pre-mortem.

Give AI the first draft of anything that is mostly assembly: the minutes, the status summary, the skeleton of a plan. Keep for yourself anything that is mostly a judgment.

The same tool helped on some tasks and hurt on one

Generative AI’s abilities are uneven in ways people do not see coming, according to a preregistered experiment with Boston Consulting Group consultants, which appeared in the journal Organization Science in 2026. Tasks that look equally hard can fall on either side of what the tool can do, which the authors call a jagged technological frontier.¹

The study

Moderate evidence

Inside the frontier and beyond it: a field experiment with 758 BCG consultants

Each consultant was assigned at random to work with no AI, with GPT-4, or with GPT-4 plus a brief guide to prompting. Across 18 realistic consulting tasks, those with AI finished 12.2 percent more tasks and worked 25.1 percent faster, and their output was rated higher. A separate task was built to fall beyond the tool’s reach. Its spreadsheet seemed to hold everything, but the decisive details sat in interview notes, and there the AI users were 19 percentage points less likely to get the recommendation right than those working without it.¹

The wrong answers did not look wrong. Graders scored the AI users’ recommendations as higher quality whether or not the underlying answer was correct.¹ The evidence comes from one firm and consultants below manager level using a 2023 model, and only one task sat beyond the frontier.

Project work is full of tasks shaped like that business case. A schedule spreadsheet can look complete while the real constraint sits in a hallway conversation: the supplier who hinted at a delay, the sponsor who will not approve overtime. An AI summary built from the spreadsheet will read confidently and miss it. Before trusting an AI answer, ask whether the deciding fact lives in the documents you gave it or in context only your team holds.

What are the risks of using AI in project management?

The main risks are invented facts, misplaced trust when nobody owns the output, and confidential project data sent to public tools. The US National Institute of Standards and Technology (NIST), in its 2024 guidance on generative AI, names confidently stated false content, which it calls confabulation and others call hallucination, as one of the technology’s core risks.⁵

NIST explains why it happens. Generative models write by predicting likely next words from what they absorbed in training, so the same process that yields accurate answers can also yield fluent, false ones, especially in long answers and specialist subjects.⁵ In a project, that might be a risk register entry citing a contract clause that does not exist, or a status summary saying a dependency was resolved when it was only discussed.

Accountability does not move to the tool. In a well-known US case in 2023, lawyers filed a court brief citing judicial opinions that ChatGPT had made up, and a federal judge in New York fined them and their firm $5,000.⁶ The judge’s objection was to the unchecked citations, and his opinion says the lawyers who sign a filing carry the duty to make sure it is accurate.⁶ A project manager is in the same position. Whoever signs a plan, a status report or a forecast answers for it.

Whoever signs a plan, a status report or a forecast answers for it.

Confidential data is the third risk. The OAIC, Australia’s privacy regulator, advises organizations, as a matter of best practice, to keep personal information, and sensitive information above all, away from generative AI tools that anyone can use, citing the privacy risks.⁷ Project files often hold supplier prices, staff names and client details. Other countries set their own data-protection rules, so check with your national regulator and follow your organization’s policy; what a workplace AI policy should cover sets out the rules worth writing down.

Can AI estimate project timelines accurately?

AI can estimate project timelines only as well as the past data behind the estimate, and it is weakest on work unlike anything it has seen. A 2007 systematic review of software cost estimation by Barbara Kitchenham and colleagues found that models built on other companies’ project data were often worse, and never clearly better, than models built on a company’s own history.⁸

Of the seven independent comparisons in the review, four found cross-company models significantly worse and three found no clear difference, so the authors called the results inconclusive.⁸ The review predates generative AI, but the problem carries over. A chatbot asked “how long will this migration take?” has no record of your team’s speed, your approval queues or the last migration that slipped. It produces a plausible-sounding typical answer.

Novelty makes it worse. In a 2025 study of AI assistance for 5,172 customer-support agents, the gains were largest on moderately rare problems, where agents had less experience but the system still had enough training data to draw on, the authors report.⁹ Brand-new work has no such data. A first-of-its-kind project, a new regulation, or a team that has never worked together sits where AI estimates are least trustworthy, and where the planning fallacyplanning fallacy: The tendency to underestimate how long your own tasks will take, even when you know similar tasks ran late before. Estimating from your track record on past tasks, rather than from the plan, tends to correct it.Full entry in the glossary already pulls human estimates too low.

Ask the AI for its assumptions, then compare its estimate with the actual durations of two or three of your own past projects that most resemble this one. Where you have none, say so in the plan and add a range.

Will AI replace project managers?

No evidence so far shows AI replacing project managers. The forecast most often quoted is a March 2019 Gartner press release predicting that by 2030 AI would eliminate 80 percent of the work of today’s project management discipline, taking over data collection, tracking and reporting.¹⁰

That figure is a prediction from an analyst firm, issued alongside a report for paying clients and a Gartner conference, with no method or data released with it.¹⁰ It also came out more than three years before ChatGPT’s public release in November 2022.² Repeating it as a finding gives it more weight than it has.

Myth

Research shows AI will take over 80% of project management tasks by 2030.

Fact

That number is a 2019 Gartner forecast from a press release, with no published method, made more than three years before ChatGPT appeared.

A 2025 index of generative AI exposure from the International Labour Organization (ILO), built from the nearly 30,000 tasks listed in Poland’s occupational classification, expects most jobs to be reshaped rather than removed, since most occupations are made up of tasks that still need a person.¹¹ That fits what project managers spend their days on. A project manager’s working day is mostly talking, negotiating and deciding, the work AI drafting tools touch least.

PMI’s own 2023 survey work points the same way: 21 percent of respondents to its annual global survey said they used AI always or often in managing projects, a self-reported figure from the association’s own survey.²

21%of respondents to PMI's 2023 annual global survey said they used AI always or often in managing projects (self-reported)Source: PMI, 2023

Adoption is real but partial, and the reporting work Gartner expected to vanish is where the drafting and summary gains described above apply. Let AI take the reporting chores it handles well, and put the hours saved into the conversations and decisions it cannot handle.

A hand-over test for any project task

The safest routine is to hand AI the work that is easy to check, verify it against the source, and keep the decision yourself. PMI’s framework asks project managers to ensure adequate oversight of AI output, with more human review as tasks grow more complex.²

A project manager writing a monthly steering update could ask AI to turn three weeks of meeting notes into a first draft, then spend ten minutes checking each date, figure and named owner against the notes, removing anything confidential and adding the one judgment the draft cannot make: whether the project is on track.

Before you hand a project task to AI

• Is the answer mostly assembly from documents you can give it?
• Can you check the output against a source in a few minutes?
• Does the deciding fact live in the documents, or only in your team?
• Have you removed personal and confidential data, or used an approved tool?
• For an estimate, have you compared it with your own past projects?
• Will a named person review and sign the output before it is shared?

The bottom line

AI in project management earns its place on drafts, summaries and first passes at risks, and it fails quietly on judgment, novel estimates and facts it cannot see. Start with one recurring document, such as the weekly status update, let AI write the first draft, and check every date, number and owner against your notes before you send it.