The Next Decade of AI in Construction: Five Shifts I Am Betting On

Construction Is Late, and That Is the Opening

Most of what gets called AI in construction today is really better record-keeping. Cameras that log what happened on the workfront, dashboards that report yesterday's progress, systems that document an incident after it has already occurred. That work matters, but it is the floor. The next decade is about climbing off it, moving from recording what already happened to predicting and preventing it while there is still room to act.

Construction has earned its reputation for moving slowly. McKinsey's productivity work ranks it among the least digitized sectors in the economy, and puts its labor productivity growth at about one percent a year over two decades, against 3.6 percent for manufacturing. The easy reading is a sector that resists technology. The more useful one is a sector sitting on twenty years of gains it has not captured, under rising pressure to finally capture them. Labor costs climb, schedules compress, and hyperscalers and industrial owners now demand tighter quality documentation and more predictable delivery than the old way can produce. AI is moving from pilot budget into work-package, inspection, and controls workflows faster than most teams planned for.

The shifts worth betting on all push the same direction. Each moves AI further upstream, from documenting field risk after the fact toward flagging it while the work can still be changed. And each runs into the same constraint, which is rarely the model and usually the operating loop around it: who gets the signal, what decision it drives, and whether anyone on the workface trusts it enough to act. That constraint is the fifth shift, and as anyone who has watched a good dashboard get ignored already knows, it largely decides whether the first four pay off.

Here is how I would place the five bets, by how ready each is to deploy today against how much leverage it carries on delivery once it lands.

1. Safety Moves From Lagging Metrics to Leading Signals

The industry standard for safety performance, total recordable incident rate and lost-time injury frequency, measures what already happened. By construction it is a rear-view mirror. The shift underway is to supplement it with leading signals pulled from continuous monitoring: posture-risk trends, access-violation frequency, PPE-compliance rates, behavioral exposure indices. The technology is no longer speculative. Intenseye reads existing CCTV feeds for more than fifty categories of unsafe act in real time, and its bottling-plant deployment at Swire Coca-Cola reported 60 percent fewer unsafe acts, a manufacturing result, but the same architecture a jobsite camera network feeds. On the construction side, Oracle bought Newmetrix in December 2022 for a safety-risk engine trained on roughly 17 million construction images that scores which projects are most likely to have an incident next week.

The model is the easy part. The operating change is the hard one: acting on an AI-generated signal before an incident report exists to justify it. A supervisor who pulls a crew off a task because an exposure index is trending the wrong way is making a call the old metrics would never have prompted, and defending it without the usual paperwork. The adoption numbers show how far that nerve still has to travel. Dodge Data and Analytics found that only about five percent of contractors use AI-based visual monitoring at all, and the most common objection, cited by 57 percent, is data accuracy, ahead of cost. Clear that trust hurdle and you intervene while the risk is still in the work method, before it becomes a number in the safety report.

2. Drone Inspection Goes From Piloted to Autonomous

Drone-based site inspection is already common. What is still early is full autonomy: UAV fleets that fly pre-planned surveys, process the imagery against the BIM model, and generate deviation reports without someone steering them or reviewing every frame. The regulatory wall came down in adjacent industries first. In November 2023 the FAA granted Percepto a nationwide waiver letting a single remote operator run up to thirty drone-in-a-box units beyond visual line of sight with nobody on site, the approval that had already put autonomous drones on Delek's Texas and Arkansas refineries. Skydio reported more than a thousand docks deployed and four million logged flights within a year of shipping its X10 dock, most of them in utility inspection and security.

None of those are construction megaprojects, and that is the opening. A refinery is a fixed asset; a jobsite changes every shift, which makes the coverage problem harder and the payoff larger. Brought onto a large program, the same dock-and-fleet architecture gives quality, progress, and structural monitoring a frequency that ground-based inspection cannot easily match. The bet here is on autonomy closing the gap between how often a site changes and how often anyone actually looks at all of it. On programs where coverage gaps carry real safety, quality, or progress risk, autonomous inspection is on its way from a pilot line item to standard scope.

3. RFIs Get Predicted Before They Block the Workfront

Requests for information are one of the most reliable sources of hidden schedule delay on complex work. The Navigant Construction Forum's study of 1.1 million RFIs across 1,362 projects put the average at 796 per project, roughly $1,080 to process each one, and a median reply time of 9.7 days. That is close to $860,000 of pure review cost per project before counting a single hour of crew time lost waiting on an answer at a critical workfront. Models trained on historical RFI logs, cross-referenced with design-package status, trade sequencing, and schedule data, can flag which scopes are most likely to generate RFIs before the work starts. The research is catching up to the idea: a 2025 study of 3,628 RFIs from a single airport program hit prediction accuracies between 59 and 92 percent depending on which features the models were given, and Autodesk's Construction IQ already sorts live RFIs by root cause and discipline to surface the risky ones. That lets a project engineer front-load coordination on the highest-risk interfaces, so the goal becomes fewer RFIs in the first place instead of faster handling of the same pile.

The wide accuracy spread in that airport study is the real lesson. The pattern recognition is within reach of current models; what separates 59 from 92 percent is clean history, RFIs tied to scope, discipline, owner response, schedule effect, and final resolution. Structured and labeled, that log becomes foresight instead of an audit trail read after the fact.

4. The Digital Twin Becomes a Live Safety and Quality Layer

Building information modeling gave construction a virtual model of what is being built. The digital twin extends that into the present tense: a live model that reflects the actual state of the work at any moment, fed by site cameras, wearable sensors, inspection records, and schedule updates. The application that matters most is automated conflict detection, the model flagging when field conditions diverge from plan before the divergence hardens into a defect or an incident. The first commercial version of this already runs on helmet cameras and 360-degree capture: Buildots compares walked imagery against the model across 27 Wates projects and more than 300,000 square meters, and Doxel claims an 11 percent average schedule acceleration on the programs it tracks, a vendor figure, but a plausible one for catching drift a week earlier.

That requires AI in the interpretation layer, reading sensor and image streams continuously and surfacing only the deviations worth a human's attention. It is the most data-hungry bet of the five, and the industry's own numbers say why it is still ahead of mainstream practice: in the same Dodge survey, only 26 percent of contractors rated their data quality as high. The useful work is already happening on programs disciplined enough to connect model state, field observation, and owner decision records. Where that discipline exists, the twin stops being a presentation object and starts catching problems while they are still cheap to fix.

5. The Human-in-the-Loop Is the Keystone

This is the keystone, the bet that decides how much the other four return. An AI safety alert no one acts on adds little, and a predictive schedule the project manager does not trust changes little. The gap is already visible in the adoption data: 86 percent of large contractors told Dodge they see AI as a competitive advantage, yet only about one in ten use predictive analytics, and 41 percent of small employers do not use their safety data at all. The belief is there, and the loop is missing. The return comes from closing the loop between AI output and field action: who receives the signal, what decision it supports, and how the result feeds back so the model earns trust over time.

The profile that makes this work is technical fluency joined to field judgment, someone who can configure and validate the systems, translate their output into a call a supervisor will actually make, and build the feedback loop that makes the next output more trustworthy. That capability belongs close to project engineering, safety, quality, and controls, with a field-fluent engineer who learns to run the loop, more than with a data-science role added at a distance from the work. Put that role in place early and the program gets to decide where AI belongs in the work, and where it does not.

Where I Am Placing the Bets

Put the five together and the pattern is hard to miss: each moves construction from recording the past toward predicting and preventing it, and each depends on the operating loop that turns a signal into a decision on the workface. The capability bets, safety signals, drones, RFI prediction, the live twin, will keep maturing whether or not any single firm invests in them; the vendors and the FAA are seeing to that. The keystone bet matures when a team deliberately puts field-fluent people in charge of the loop. That is where I am putting my attention over the next decade, and it is the cheapest high-return move most programs have not yet made.