Can AI Keep Projects on Schedule?

The problem, in numbers

Late delivery is not the exception in public infrastructure. Victoria's Auditor-General reviewed the 110 major projects, each worth more than $100 million, in the state's 2025-26 capital program. Of the 81 with a completion date that could be compared with the one first published, 40 had pushed it back, five had brought it forward and 36 were unchanged.^{Source 1} Together the slippage added 369 quarters (about 92 years, on our arithmetic), and transport projects accounted for 161 of them.^{Source 1} In its 2023 review the same office found the same pattern: 40 of 89 projects had a different completion date from the one first announced.^{Source 2}

Time and money move together. Drawing on a sample of 258 rail, bridge, tunnel and road projects in 20 nations, in which nine in ten had cost escalation, Flyvbjerg and colleagues found that escalation grew with the length of the implementation phase: on average 4.64 percentage points for every year from the decision to build until operations began, or about US$46 million a year in construction costs on a US$1 billion project.^{Source 4} In Australia, the Grattan Institute calculated that governments spent $34 billion more on transport infrastructure over two decades than they first announced, with actual costs 21% above promised costs on projects of $20 million or more.^{Source 3}

What improves, and why

A conventional programme says where a project should be. Critical path software encodes the planner's logic and durations, and monthly updates show how far the work has drifted. By the time the drift shows on the critical path, much of the float that could have absorbed it may already be gone.

Predictive scheduling tools try to move that moment earlier in two ways. First, they learn from history. Instead of relying on one planner's estimate for each activity, a model trained on past programmes estimates the spread of likely durations, so optimism in the baseline becomes visible on day one. One vendor, nPlan, says its dataset holds more than 750,000 past project schedules.^{Source 8} Second, they read the live programme, scoring the probability that each milestone will be met as updates arrive.

That is the Faster case. The gain is not in how quickly concrete cures. It is in when the team learns that a milestone is at risk, and so how much time remains to resequence work, add crews or settle an interface before the delay locks in.

What the evidence actually says

The research base is wide and shallow. A 2026 systematic review coded 64 peer-reviewed studies of machine learning in construction project management published from 2015 to 2025. In a synthesis of ten comparative studies, it reported that hybrids with metaheuristic optimisation outperformed single algorithms in delay prediction, and it noted a significant gap between research advances and practical implementation.^{Source 7}

Headline accuracy figures need careful reading. One peer-reviewed study reported 91.67% accuracy for a random forest tuned with a genetic algorithm, but it identified its delay factors through a literature review and a questionnaire.^{Source 5} Another built a stacked ensemble of ten models on data collected by surveying experts, and its abstract reports no accuracy figure.^{Source 6} Studies like these show which risk factors are associated with delay. In our reading, they do not show that a model watching a live programme warns earlier, or more reliably, than a competent planner.

The vendor we reviewed claims a far larger dataset and discloses less. nPlan's home page refers to mean absolute error and a white paper but gives no error values in its text; its worked examples are its own reconstructions from public data, one forecasting a planned 2030 opening as most likely 347 days late with a 1% chance of meeting the date.^{Source 8} Forecasts of unfinished projects cannot yet be scored. We found no peer-reviewed study comparing finish dates on projects that used predictive scheduling with comparable projects that did not.

There is also reason to doubt that forecasting is the main bottleneck. Grattan found that 80% of cost overruns came from 14% of projects, and that the one third of projects announced before governments were ready to commit accounted for more than three quarters of overruns.^{Source 3} Victoria's auditor found that for 12 projects the budget papers explained a changed completion date only as aligning "with a revised project schedule".^{Source 1} In our reading, those are failures of decision and disclosure, not of prediction. And the cost of delay has been measurable at portfolio level since Flyvbjerg's study was published in 2004.^{Source 4}

Where it falls short

• Forecasting is not fixing. Victoria's auditor lists rising material costs, limited access to materials, skill shortages and rising labour costs as the main reasons major projects run over time and budget, with land acquisition delaying some.^{Source 2} A model may see these coming. It cannot remove them.

• Weakest where it matters most. In our assessment, a model trained on past programmes is least reliable on the unusual project: a new tunnelling method, a market at capacity, a first-of-its-kind asset. Those are the projects with the most at stake.

• Poor records in, poor forecasts out. Even public reporting is patchy: ten completed Victorian projects report a financial completion date rather than a practical one.^{Source 1} A model cannot learn from finish dates nobody published.

• Accuracy claims are not comparable. Classification accuracy on questionnaire data and a vendor's error on proprietary schedules measure different things.^{Source 5,Source 8}

Where governance decides the outcome

A probability of delay is uncomfortable information. It may contradict the date a minister announced or the programme a contractor priced. Decide before deployment who sees each forecast, who must respond to a high-risk flag and by when, and whether a forecast of slippage is reported to the client or, for public projects, disclosed. Grattan has argued that governments should continuously disclose to Parliament material changes to expected costs and benefits.^{Source 3} Agree in the contract whether model forecasts carry any weight in delay claims; otherwise treat them as management information only. Accountability for the programme stays with a named planner and project director, not with the model.

How to think about adoption

Start with your own record. Compare planned and actual completion on past jobs, because without a baseline no forecast can be judged. Pilot predictive scheduling on a live project alongside the existing programme, log every forecast with the date it was made, and score it when the milestone arrives. Ask vendors for error measured on completed projects like yours, not reconstructions. Give each high-risk flag a named owner and a recovery decision. For owners and governments, in our reading, the strongest lever in this evidence sits upstream: commit to a project, and announce its date, only once it is ready.