AI in Project Management: Separating the Evidence From the Forecast

Almost every article about artificial intelligence in project management opens with the same statistic: eighty percent of project management work will be eliminated by AI by 2030. It appears in vendor explainers, conference keynotes, and more than a few board decks. It is a Gartner prediction, and it was published in March 2019, several years before generative AI reached anyone’s desk.

That is not a reason to dismiss it. It is a reason to be careful about what kind of claim it is. A forecast made before the technology existed is a statement about direction, not a measurement of effect. The gap between those two things is where most organizations currently make their AI investment decisions, and it is worth closing before you commit a budget.

What has actually changed

The clearest measured change is adoption. A 2025 survey by the Association for Project Management found that seventy percent of project professionals reported their organization was using AI, up from thirty-six percent two years earlier. That is close to a doubling in two years, a genuinely fast diffusion curve for a discipline that tends to move slowly.

Adoption is not outcome, though. Knowing that seventy percent of organizations use AI tells you nothing about how deeply, on which tasks, or with what result. The productivity figures quoted alongside it are perception measures. Atlassian’s 2025 collaboration index reports that workers believe AI makes them roughly a third more productive. Belief about productivity and measured productivity are different variables, and in every technology wave so far they have diverged.

The honest position today is that adoption is real and rising fast, the capability is genuinely useful in specific places, and the outcome data is not yet in.

Why project management was always going to be a target

The underlying problem is old and expensive. Planview cites roughly thirty-five percent of projects succeeding. Research drawing on Project Management Institute data puts complete project failure at around twelve percent globally, with roughly thirty-five percent overrunning cost or experiencing scope changes. Different framings, same conclusion: a large share of project spend does not deliver what it promised.

Part of that failure has a mundane cause. For decades, project managers ran their work through slides, spreadsheets, and documents retrofitted to purposes they were never designed for, all requiring constant manual maintenance. The overhead of keeping the plan current competed directly with the work of thinking about the plan. When the two compete, maintenance wins, because maintenance has a deadline.

That is the gap AI is genuinely good at closing. Not judgment. Overhead.

A better question than “what can AI do for project managers”

The most useful framework I have found comes from a 2025 systematic literature review in Project Leadership and Society by Almeida, Fernandes, and Santos. Rather than listing capabilities, the authors sort project management knowledge areas by the type of knowledge each one depends on, then match tool types to that.

Their thematic analysis identifies integration, scope, communication, risk, and stakeholder management as the areas where AI holds the strongest potential. The interesting move is the second one: they distinguish work resting on formal knowledge, work resting on data-driven knowledge, and work resting on tacit knowledge, and argue the dominant knowledge type should determine which kind of AI tool fits.

Formal knowledge is codified: templates, procedures, documentation, the structure of a work breakdown. This is where generative AI performs best right now. It can draft a scope statement, produce a status report from scattered inputs, summarize a decision log, or turn a requirements document into a first-pass task list. The knowledge already exists in written form. The tool is rearranging and compressing it.

Data-driven knowledge lives in history and metrics. This is machine learning territory rather than generative AI territory: forecasting task durations from historical performance, matching people to work by skill and availability, scoring risk probability, modelling what happens if a vendor slips or a budget is cut. Value here scales directly with the quality of your historical data, a much harder constraint than most pilots assume.

Tacit knowledge is the part that never gets written down: reading a steering group, knowing which stakeholder objection is real and which is positioning, deciding what to escalate and when, holding a team together through a difficult month. Both Planview and IBM name this explicitly as the boundary. Project management involves emotional intelligence, stakeholder management, and creative problem-solving in novel situations that AI cannot replicate.

The practical value of this framing is that it gives you a test you can apply before buying anything. Take the task you want to automate and ask what kind of knowledge it depends on. If the answer is tacit, no tool on the market will do it, and a vendor telling you otherwise is describing a demo rather than your project.

Where the value is actually landing

Across the vendor guidance and the academic review, the same clusters recur.

Reporting and status synthesis, generating progress reports and stakeholder updates by pulling from tickets, chat, and project tools rather than by manual collection. This is the most consistently cited use case and the one with the clearest immediate return, because it removes work nobody wanted to do.

Planning and scheduling, analyzing historical data, team capacity, and task complexity to propose timelines that account for dependencies. The gain is realism rather than speed. A schedule built from what actually happened last time is harder to argue with than one built from optimism.

Predictive risk management, monitoring for patterns that signal trouble, such as tasks repeatedly overdue or dependencies that reliably cause delay, so mitigation happens early rather than at the next monthly review. Natural language processing can also scan project documentation to surface risks that were written down but never registered.

Resource and capacity management, matching skills to tasks and identifying portfolio bottlenecks before they bite.

Scenario modelling, simulating how staffing changes, vendor disruption, or budget adjustments affect delivery, which supports better trade-off conversations with sponsors.

Knowledge management and onboarding, summarizing meetings and decision logs into action items and surfacing historical context for people joining mid-project. This quietly addresses one of the most expensive failure modes in long programs, which is that the reasoning behind past decisions evaporates when people rotate off.

Sentiment as an early warning signal, analyzing the tone of comments and status text over time to detect whether work is genuinely on track, independent of what the RAG status says. More interesting than it first appears, because the gap between reported status and felt status is where most project failures hide.

Notice what these have in common. Almost all attack the administrative and data-processing layer rather than the decision layer. That is consistent with the strongest version of the case for AI in project management, which is not that it makes better decisions but that it buys back the time in which humans can make them.

Four gates that decide whether any of it works

Data quality. AI systems improve as they process more information, and accuracy depends on a steady flow of current, consistent project data. The inverse is the part organizations underestimate. A duration forecast trained on plans never updated after week three is not a forecast. It is a confident restatement of an old assumption.

Integration. The tool has to read from the systems where work already lives, which means mapping how it interacts with your existing delivery, finance, and communication platforms before deployment rather than after. An AI tool on an island produces insight nobody can act on, and adds a place to check rather than removing one.

Adoption. Training needs to cover both operating the tool and interpreting what it produces. Resistance to change, integration friction, cost, and anxiety about employment are all named as common obstacles. That last one deserves more candour than it usually gets. If your rollout narrative is that AI will eliminate eighty percent of project management work, do not be surprised when project managers decline to help you implement it.

Oversight. Every source consulted lands in the same place: AI recommendations need critical review against organizational policy and goals, and too much automation without human oversight increases rather than reduces risk. In practice that means naming a person accountable for each class of output, not writing a principle into a policy document. A generated status report reads as finished work, which is what makes an unreviewed one dangerous.

The first two gates are engineering problems with known solutions. The second two are management problems, and they are the ones that quietly kill rollouts. A common pattern is that a pilot succeeds because a motivated team compensated for thin data and did careful review out of professional pride, and the wider rollout then fails across teams that had no such compensation and were never asked to provide it.

What happens to the role

The optimistic version, which the sources broadly agree on, is a shift away from administration toward coaching, stakeholder engagement, and strategic thinking. That is a genuinely attractive description of the job, and the one most project managers thought they were signing up for.

Two cautions belong alongside it.

The administrative work was doing something. Manually assembling a status report forced someone to look at every workstream once a week. It was tedious, and the tedium was performing accidental quality control. Automate the assembly without replacing the review, and you get a report describing a project nobody has actually examined, produced on a reliable weekly cadence, with no friction to slow it down.

And the skills the optimistic version promotes, namely stakeholder judgment, coaching, and strategic framing, are the tacit ones. They are the hardest to teach, the slowest to develop, and traditionally they were developed by doing the administrative work and absorbing context along the way. If junior project managers no longer build plans by hand or write status reports from scratch, the profession needs a deliberate answer for how they acquire the judgment senior roles require. Nobody has published a good one yet.

A reasonable way to start

Audit where your project managers actually spend their time, then sort those activities by knowledge type. Formal and data-driven work is your candidate list. Tacit work is not.
Start with reporting and synthesis. Clearest return, lowest risk, and it builds credibility for the harder use cases.
Before buying anything predictive, assess your historical project data honestly. If it is incomplete or inconsistent, fix that first.
Define the review step before you deploy the capability, and name who owns each type of output.
Train on interpretation, not just operation. Knowing when a recommendation is wrong is the skill that matters.
Measure something real. Cycle time on reporting, forecast accuracy against actuals, or rework volume. Adoption rates and satisfaction scores will look good regardless of whether anything improved.
Be straight with the team about what this is for. If the goal is removing overhead so people can do the work they were hired for, say that. It is both more accurate and more persuasive than the elimination narrative.

The honest summary

AI is doing something real in project management, and it is mostly not what the headline predictions describe. It is removing a maintenance burden that has weighed on the discipline for thirty years, and it is making certain kinds of forecasting cheap enough to do continuously rather than quarterly. Those are meaningful gains.

The academic review is candid about the limits of what we know. Tools evolve fast enough that specific conclusions date quickly, and effects vary considerably across business contexts and project types. The literature does not yet offer a unified way for a project manager to determine where and how AI tools are most effectively applied.

Which means the most valuable thing an organization can do right now is not to pick the right tool. It is to get its project data in order, decide who reviews what, and build the internal evidence base the industry does not yet have. The organizations that do that will be able to answer the outcome question for themselves in two years. Everyone else will still be quoting a prediction from 2019.

References

Almeida, P. M., Fernandes, G., and Santos, J. M. R. C. A. (2025). Artificial intelligence tools for project management: A knowledge-based perspective. Project Leadership and Society, Volume 6, Article 100196. DOI: 10.1016/j.plas.2025.100196
McGrath, A. and Downie, A. (2025). What is AI in project management? IBM Think.
Grace, F. How to use AI for project management: A complete guide. Atlassian Work Management.
Appelbaum, B. Using Artificial Intelligence for Project Management. Planview.
Gartner (2019). Gartner Says 80 Percent of Today’s Project Management Tasks Will Be Eliminated by 2030. Press release, March 2019. Later portfolio predictions from Gartner, Build a Generative AI Roadmap for Your Portfolio Management Life Cycle, July 2024, ID G00795336.
Association for Project Management (2025). AI use in project management nearly doubles in just two years, APM survey finds. September 2025.