The History of AI in Project Management

By Victor Hoang · Updated Oct 21, 2026

Picture a DuPont engineering department in late 1956. Projects had grown too big for any group of planners to hold in their heads. The company's Integrated Engineering Control Group started asking a question. Could an electronic computer help?

Computers began scheduling projects in 1957. That's when DuPont and Remington Rand ran critical path calculations on a UNIVAC I. Artificial intelligence proper only entered the story in the 1980s. Researchers at Carnegie Mellon and the Space Telescope Science Institute built constraint-based schedulers then. Generative AI reached everyday project managers after late 2022. The Project Management Institute (PMI) has tracked its adoption since 2019. The throughline: machines took over the arithmetic of planning long before anyone trusted them with judgment.

That distinction matters for what follows. Algorithmic scheduling such as CPM and PERT is a precursor, not AI. Its authors describe it as a planning and scheduling technique. The AI concept is a separate line of research that started at Dartmouth in 1955, and the two lines met slowly, in factories and observatories before they reached the ordinary project office.

Timeline at a Glance

The UNIVAC Schedules (1956-1959)

Computers learn to recompute a plan

Picture a DuPont office in 1957, with engineers drawing arrows between tasks and a UNIVAC I running in Newark, Delaware. This is the precursor, and it is worth being exact about what it was. In their 1959 paper, Kelley and Walker explain that traditional planning left each group scheduling on its own, and blame "the inability of unaided human beings to cope with sheer complexity". DuPont's Integrated Engineering Control Group began a survey in late 1956 that asked whether a computer could prepare a master schedule, revise it economically when conditions changed, and keep management informed. Remington Rand UNIVAC assigned James Kelley to help, and Morgan Walker represented duPont.

The answer came in stages. By May 1957 the two companies started a cooperative implementation, with Remington Rand supplying programs for the UNIVAC I, and the results were officially demonstrated in September 1957. A first live test, called Project A, produced the findings that made the case. When about 40 percent of the plan changed by March 1958, updating the schedule took only about 10 percent of the effort of the original. The computed results also showed only seven critical deliveries out of 156 items, three of which the normal scheduling group had missed. That is the critical path method in action.

The Navy's version ran in parallel. The 1959 paper by Malcolm, Roseboom, Clark and Fazar describes PERT for the Polaris missile program as a network of events with time estimates from responsible technical staff, expressed in probabilistic terms and run on the NORC computer. The method lives on in the PERT chart. The wider story of both is in the history of project management and the piece on who invented project management.

Keep precursor and concept apart. CPM and PERT computed an answer from rules a person supplied. They did not learn, reason or plan. What they did was establish the idea that a plan is data a machine can recompute, which every later era builds on.

What this means for teams today: The first value of software in a project was fast recalculation, not cleverness. If your team can't re-plan in an afternoon when scope shifts, fix that before you shop for AI.

The Dartmouth Summer (1955-1972)

AI gets a name while scheduling stays arithmetic

This era overlaps the last one, because two fields grew side by side and barely spoke. The summer research project that took place at Dartmouth in 1956 began with a proposal dated August 31, 1955. The proposal came from J. McCarthy of Dartmouth, M. L. Minsky of Harvard, N. Rochester of IBM and C. E. Shannon of Bell Telephone Laboratories, and it aimed to find how machines could "solve kinds of problems now reserved for humans". The AAAI's reprint of the proposal confirms that the 1956 summer project grew out of this document.

Through the 1960s, project management took up the network methods above and organized itself as a profession, while AI researchers worked on machines that could use language and solve problems. Sir James Lighthill's survey, dated July 1972, was written for the Science Research Council, and singled out one general cause for the disappointments: failure to recognise the "combinatorial explosion". Our history of AI and its AI timeline tell the winter that followed in full.

The combinatorial point matters for planners. Every extra task, resource and constraint multiplies the schedules that could exist, so a machine that tries them all stalls. The next era is about people who found ways around it.

What this means for teams today: Planning problems blow up fast as constraints pile on. If a tool promises an optimal plan for a messy portfolio, ask what it assumes and what it leaves out.

The Knowledge-Based Schedulers (1980s-1990s)

Constraint reasoning meets the factory and the telescope

A job shop floor in the early 1980s, with orders, due dates and machines all pulling against each other. At Carnegie Mellon, Mark Fox and colleagues worked on ISIS, which a June 1983 technical report titles a constraint-directed reasoning approach to job shop scheduling. A year later, Fox and Stephen Smith published a paper in the journal Expert Systems. According to the abstract, scheduling depends on many constraints that often conflict, such as due dates, costs, production levels, machine capabilities and resource availability. ISIS represents them, uses them to guide a search for an acceptable schedule, and relaxes them when they can't all be met.

Notice the shift. CPM took a plan someone had drawn and computed its timing. ISIS tried to build the schedule while weighing trade-offs that humans had previously settled by negotiation. That is an early, well-documented case of AI methods aimed at the planning problem rather than at the clock arithmetic. It was a factory scheduler, not a project tool, so it belongs in this story as a relative, not a direct ancestor.

The best-documented project-like case is the Hubble Space Telescope. A 1991 paper by Glenn Miller and Mark Johnston presented at the Goddard Conference on Space Applications of Artificial Intelligence describes SPIKE, a long-range scheduling system built with AI techniques, and covers tools based on constraint satisfaction and neural networks. A 1990 paper by Jeffrey Sponsler and Mark Johnston says SPIKE is used to create long-term schedules for Hubble, and tackles what to reschedule when feedback invalidates earlier decisions. The telescope was launched in 1990, and by May 1998 another mission's science center had selected SPIKE as its long-range planning tool.

Expert systems as a category lost momentum in the late 1980s, partly because they were costly to maintain, as our history of AI explains. The scheduling work carried on, in places with a narrow problem and rich data. Telescope time is a far tidier planning problem than a product launch with shifting scope.

What this means for teams today: AI scheduling worked first where constraints could be written down. The tighter your definition of resources, dependencies and deadlines, the more any planning tool can do.

The Quiet Years (1990s-2010s)

Statistics learns from past projects, away from the spotlight

The documented AI line in project settings through these decades runs through research on predicting how much effort software projects will take. A 2023 review by Víctor Uc-Cetina calls a 1995 study by Srinivasan and colleagues "one of the earliest" to use machine learning for effort estimation, comparing regression trees and multilayer perceptrons with traditional parametric models. By 2012, Wen, Li, Lin, Hu and Huang were able to publish a systematic literature review of machine learning based software development effort estimation models.

Estimation was the right target, because it is where project managers feel uncertainty most. Our article on the cone of uncertainty shows why early estimates are wide, and learning models try to narrow them using data from past projects. The catch is built in: such a model is only as good as the history it learned from.

Meanwhile the profession's own story ran toward agile methods, portfolio management and cloud work tools, as the evolution of project management methodologies describes.

What this means for teams today: Data you collect now is the raw material for any future prediction. Record actual effort and duration against estimates, consistently, and you have the dataset that this era lacked.

The PMI Reckoning (2019-2022)

The profession starts counting AI

In June 2019, PMI published a Pulse of the Profession in-depth report, AI Innovators: Cracking the Code on Project Performance. The page reports that 81 percent of respondents said their organization was being impacted by AI technologies, and that project professionals expected the share of their projects using AI to rise from 23 percent to 37 percent over three years. It also describes five principles, called MELDS, that Accenture research identified in the organizations making the most visionary AI investments.

Consider what that report was: the profession's main body treating AI as something to measure. The numbers reflect expectations stated in a survey, not proven adoption, and the report was written before the tools most project managers now use existed.

The turn came at the end of the period. PMI's later report ties the profession's shift to the public release of ChatGPT in November 2022, which opened the next era. How AI moved from labs toward everyday work is also the subject of AI: from research labs to work.

What this means for teams today: Expect-versus-actual gaps are normal. When a survey says your peers will use AI on more projects, check what they actually changed in their weekly work.

The Generative Turn (2022-present)

Assistants arrive in the weekly status report

Today's project manager opens a status draft written by a language model and decides what is true. The change since 2022 is that AI can handle text, the medium most coordination work is made of. PMI's October 2023 report, Shaping the Future of Project Management With AI, lists survey figures: 21 percent of respondents said they use AI, 82 percent of senior leaders said AI will affect projects, and 91 percent of respondents believed it will affect the profession. It also cites unpublished PMI research in which about 20 percent of project managers had extensive or good practical experience with AI tools and 49 percent had little or none.

Adoption then moved quickly. At its Global Summit on September 19, 2024, PMI announced that high adopters of generative AI, meaning people using it in at least half of their projects, had grown by 86 percent since January 2024, and that roughly two in five project professionals surveyed were in that group. In the same release it said it had acquired Cognilytica, whose CPMAI program trains people in managing AI and machine learning projects, and upgraded PMI Infinity, its AI assistant.

The organizational picture lags the individual one. A PMI report on organizational support, based on the third wave of its survey (July-September 2024) of 2,000 project professionals in 18 industries and 12 countries, says only 1 percent of organizations believe they have reached generative AI maturity. In organizations with little or no support, 8 percent of respondents reported advanced usage; in organizations with strong support, 41 percent did. The report also names the heaviest users "Trailblazers" and says they apply generative AI to complex tasks like risk management and planning.

What the new tools do is the same job as the old ones, one level up. CPM recomputed the schedule when the plan changed. A status agent reads the plan, notices slippage and drafts the update, which is the pattern described in our AI project status agent blueprint. PMI's companion report describes professionals applying generative AI to everything from meeting summaries to complex planning and risk management.

What this means for teams today: Whether AI helps depends more on your support, training and clean plan data than on which tool you pick. Name an owner for AI use before the first status report goes out under their name.

Key People

Key Facts

Key Facts: AI in Project Management History

• Kelley and Walker's paper records that Remington Rand and duPont began a cooperative CPM effort in May 1957, using duPont's UNIVAC I in Newark, Delaware.
• The PERT paper says it was applied to the Polaris program and run on the NORC computer.
• SPIKE, a scheduler built with AI techniques, was developed to plan Hubble Space Telescope observations.
• PMI's 2019 report found that 81 percent of respondents said their organization was impacted by AI technologies.
• PMI's 2024 release said roughly two in five project professionals surveyed use generative AI in more than half of their project work.
• A PMI report says only 1 percent of organizations believe they have reached generative AI maturity.

The Pattern Behind the History

We call this the Arithmetic First pattern. It is our synthesis of the timeline, not a research finding.

1. Arithmetic before judgment. CPM recomputed schedules in 1957; assistants drafting updates came 65 years later.
2. High-stakes programs first. Polaris, Hubble and a factory floor used the new methods before ordinary teams did.
3. Data sets the ceiling. Constraint schedulers needed written-down constraints, and learning models needed records of past projects.
4. Expectation runs ahead of practice. PMI's 2019 survey measured what people expected, and the 2025 report puts organizational maturity at 1 percent.
5. The human stays accountable. Every era kept a named person responsible for the plan.

For readers who use the ACE vocabulary, the capabilities arrived in this order: Analyze (CPM), then Predict (estimation models), then Generate (status drafts), then Execute (agents that act on the plan).

What It Teaches Teams Today

• Treat recalculation as the baseline. A team that can't re-plan quickly gains little from AI on top.
• Write constraints down. The 1980s schedulers worked because resources, dates and dependencies were explicit.
• Keep estimate-versus-actual records. They are the training data every prediction method needs.
• Support beats tools. PMI's organizational data shows 41 percent advanced usage with strong support and 8 percent without it.
• Name an owner. Accountability for a project's outcome has always sat with a person, and it still does.

To go further, read the project management timeline and the story of the critical chain method, another planning idea that took decades to land.

The history of AI in project management is the history of handing machines the arithmetic of a plan, then the drafting of it, while the judgment stays human. The teams that gain most are the ones whose plans were clear enough for a machine to read in the first place.