Industries / Construction and property / Commercial and residential construction
Scheduling
Schedulers put every task in order, estimate how long each will take, and track the job against that plan so that trades, materials and inspections arrive when they are needed.
4
Milestones achieved, from 1931 to 1994.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1931
A tower in 410 days
Construction of the Empire State Building in New York began on 17 March 1930 and finished in one year and 45 days, with the frame rising four and a half stories a week, by the building's own history.
Credit
builders Starrett Bros. & Eken; architects Shreve, Lamb & Harmon
1959
The critical path
DuPont wanted to use its UNIVAC I computer to plan plant maintenance. Its engineer Morgan Walker and James Kelley of Remington Rand found a way to compute the chain of tasks that sets a project's finish date, ran it first on 24 July 1957, and published it in 1959 as the critical path method (CPM). John Fondahl of Stanford spread it through construction from 1961. Its activity-on-node form, the precedence diagramming method, has become "the prevalent technique of our time", by Miklós Hajdu and Shabtai Isaac's 2016 history.
Credit
Morgan R. Walker (DuPont); James E. Kelley Jr. (Remington Rand); John Fondahl; Bernard Roy in France, who developed a similar method
1983
Schedules on a desktop computer
Primavera, founded in 1983 in Bala Cynwyd, Pennsylvania, sold critical path scheduling software for personal computers. The Philadelphia Inquirer called it a construction project software maker when Oracle bought it in October 2008. Oracle still sells it as Primavera P6, and in 2026 newer AI scheduling tools still take P6 schedules as their input.
Credit
Primavera Systems
1994
The last planner commits
Under the Last Planner System of Glenn Ballard and Greg Howell, the foremen who will do the work, the "last planners", commit each week only to tasks that are ready, and the share of commitments kept is tracked. Before it, research cited by the Lean Construction Institute found only 54% of the week's assignments made by foremen were completed. Ballard and Howell first used it to rescue a troubled project in 1994 and founded the Institute in 1997. A 2024 review of 112 studies by Agrawal and colleagues calls it widely adopted and finds that much of it is still done by hand.
Credit
Glenn Ballard and Greg Howell
Watch for
AI-generated schedules
McKinsey and ALICE Technologies say ALICE's software, which treats labor, equipment, space and sequence as variables and simulates the outcomes from a building model or a Primavera P6 schedule, shortened early adopters' schedules by up to 20%, and one data center's construction schedule by 40%.
Achieved whenAn owner or an independent party publishes the planned and actual finish of a building delivered on such a schedule, against the schedule it replaced.
Claimed by
McKinsey and ALICE Technologies, April 2026
Open
01
Work sampling by camera
A tool that reads footage from a fixed site camera and, at random moments, records whether each worker in view is doing direct work, preparing, waiting or walking. This is work sampling, the method planners use to measure how fast crews really work, and those rates set the durations in a schedule. Publish a test of the tool against a trained human observer watching the same footage.
The briefCredit · unclaimed02
Why weekly plans fail
A model that reads the free-text reasons recorded each week for Last Planner tasks that were not finished and sorts each into a standard category, such as missing information, late materials, earlier work not done, or crew moved, and a public test of it against two planners' coding.
The briefCredit · unclaimed
Open 01 · the brief
Work sampling by camera
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A tool that reads footage from a fixed site camera and, at random moments, records whether each worker in view is doing direct work, preparing, waiting or walking. This is work sampling, the method planners use to measure how fast crews really work, and those rates set the durations in a schedule. Publish a test of the tool against a trained human observer watching the same footage.
Why it is still open
A 2026 conference paper by Ahmad Muhtarom and Hung-Ming Chen warns that reading camera detections frame by frame as productivity breaks the rules of work sampling and turns them into work sampling tables instead; its summary reports no comparison with human observers.
Built with what exists
A time-lapse or network camera; an open person detector and tracker, such as a YOLO-family model with ByteTrack; a vision-language model to label activity; a script that draws random sampling moments.
First step
Ask a builder for permission, and the crew for consent, to film one framing or drywall crew for two days, then label one hour by hand.
Done looks like
At least 20 hours of footage covering three trades, each sampled at least 1,000 times by both the tool and a trained observer, with agreement for each activity class published as Cohen's kappa, a standard measure of agreement between two raters.
Open 02 · the brief
Why weekly plans fail
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A model that reads the free-text reasons recorded each week for Last Planner tasks that were not finished and sorts each into a standard category, such as missing information, late materials, earlier work not done, or crew moved, and a public test of it against two planners' coding.
Why it is still open
A 2024 review of 112 studies finds that this knowledge sits in unstructured text documents that need manual review, and lists its automatic extraction as future work.
Built with what exists
Weekly work plans, with names removed, from a builder that uses the Last Planner System; a small open language model fine-tuned on one graphics chip (see the Note "One GPU to fine-tune") or a general AI model; a spreadsheet for the human coding.
First step
Ask a builder that runs the Last Planner System for six months of weekly plans from one project, with names removed.
Done looks like
At least 500 missed tasks from at least three projects, coded by two planners and the model, with the model's agreement with each planner published next to the planners' agreement with each other.