Can AI Cut Construction Rework?

The problem, in numbers

Rework is work done twice: a duct rerouted because it hit a beam, a slab cut open because the sleeve was missed, a wall rebuilt because the set-out was wrong. Everyone in construction knows it happens. Very few firms know what it costs them, and the published estimates disagree by an order of magnitude.

The best evidence comes from contractors' own books, which are rarely opened. In what its authors describe as the first longitudinal study of its kind, Peter Love and colleagues traced 19,605 rework events across 346 projects from one contractor. Where complete cost data existed, rework averaged 0.39% of contract value.^{Source 1} That sounds small until you set it against margin: over the same period the contractor's mean yearly profit fell by 28%, and just 88 events caused 34% of total rework costs.^{Source 1} A later study by Love found precompletion rework averaging 0.38% of contract value, and an estimated 0.76% once postcompletion corrections were included, and that actual costs had been underreported by 300%.^{Source 2} Earlier studies that combined change orders and quality issues put it as high as 12.4% of contract value.^{Source 2}

The Construction Industry Institute's benchmarking put mean field rework on heavy industrial projects above 3% of construction phase cost, and drew the conclusion that matters most here: field rework originates in design and project definition, so the effort to prevent it belongs early.^{Source 3} An industry-commissioned US survey adds a cause. It attributed 48% of rework to poor communication and poor project data, and FMI estimated those causes cost the US industry US$31.3 billion a year.^{Source 4}

What improves, and why

Two families of tools target rework, and only one of them is really new.

The first is model-based coordination. Federated design models are checked for clashes: pipes through beams, ducts through ducts, clearances that will not work. Geometric clash detection itself is mostly rules, not AI. What machine learning adds is triage. Clash runs produce thousands of hits, and a relatively high percentage are pseudo conflicts that a BIM manager must still inspect by hand.^{Source 7} Machine-learning classifiers can help sort the real problems from the noise, and preliminary tests found they performed comparably to a hand-built rule system at lower development cost.^{Source 7}

The second is reality capture. Cameras, laser scanners and 360-degree photos record what has actually been built, and computer vision compares it with the model to flag work that is missing, out of sequence or out of position. The value is timing. A sleeve missed in a slab is cheap to fix before the pour and expensive after it.

What the evidence actually says

The cost case for clash detection rests mostly on estimates. A peer-reviewed cost-benefit study of a multi-million-dollar infrastructure project calculated savings equal to 20% of contract value, but that figure is the estimated cost of clashes that would otherwise have reached site, not rework measured after the fact.^{Source 5} A US design-builder reported US$200,000 of coordination labour on a US$230 million project and US$2.22 million in estimated rework avoided, using typical cost values for about 932 of 2,000 clashes.^{Source 6} That is a credible practitioner account, but it is self-reported and its unit costs are not published.

Set those estimates beside the measured baseline and a tension appears. If one contractor's actual rework averaged well under 1% of contract value,^{Source 1,Source 2} a modelled saving of 20% on one project cannot be typical of the industry. Both can be true: one large infrastructure job may carry far more coordination risk than routine building, and measured figures may still be understated.^{Source 2} But the honest reading is that savings are real, uneven, and probably largest on complex jobs.

Reality capture is earlier on the curve. A Finnish university research programme, run with industry partners, concluded that using AI, point clouds and images to monitor progress is technically feasible, but found data collection was the bottleneck, and a quality-inspection trial lacked enough examples of poor work to train the system.^{Source 8} We found no controlled study that measures rework avoided by AI progress capture.

Where it falls short

• Savings are modelled, not counted. Almost every cost figure is an estimate of rework that did not happen. Few firms measure their rework baseline well enough to check.

• Triage can discard real clashes. A classifier that hides noise will sometimes hide a genuine conflict. The cost of one missed clash can erase the time saved on hundreds.

• Design is not the only cause. Client changes, defective materials and other changes also drive rework. Model coordination addresses only the part that starts in design.^{Source 2}

• Rare defects are hard to learn. Vision systems need examples of bad work to spot it, and good sites produce few.^{Source 8}

Where governance decides the outcome

Rework sits on the boundary between designer, contractor and trade, so the governance question is who owns a flag. When a model marks a clash as irrelevant, a named person should accept that decision and the record should show it. When reality capture finds work out of tolerance, the contract should say who pays and how fast it must be raised. Without that, early detection simply moves the argument earlier. Accountability for a dismissed clash cannot be delegated to the classifier.

How to think about adoption

Start by measuring your own rework, because the published range is too wide to borrow a baseline from. Track events, causes and cost on a handful of projects before buying anything.^{Source 1} Put coordination effort where the risk is: dense services, complex structure, tight ceiling voids. Use machine-learning triage to cut review time, but sample the clashes it discards. Treat reality capture as a pilot on repetitive work first, and judge every tool by one test: did someone act on the flag before the concrete went in?