Separating the categories
Every construction product now claims AI, so the useful move is sorting deployments by evidence. Four categories have repeatable, operator-verified results on real projects: computer-vision progress tracking, predictive maintenance on plant, schedule risk analysis, and document intelligence. A fifth — generative AI drafting site records and contractual documents — is moving fast but its value depends entirely on grounding, which is the subject of the last section.
Vision-based progress tracking
Helmet- or tripod-mounted 360° cameras walk the site; models match imagery to the design model and score completion per element. Deployments on large commercial projects (Buildots being the most prominent) consistently surface the gap between reported and actual progress — the discrepancy the monthly claim argument is made of. Requirements for it to work: a current design model, regular capture discipline, and someone who acts on the discrepancies. It fails where the model is stale or capture is sporadic.
Predictive maintenance
The most mature category, imported from manufacturing. Models trained on vibration, temperature and telemetry flag failing components on crushers, conveyors, excavators and trucks days or weeks before failure; Siemens’ Senseye and the OEM platforms (Cat, Komatsu) publish sustained reductions in unplanned downtime. It needs sensor data flowing reliably and a maintenance team that trusts and acts on alerts — cultural adoption, not model quality, is the usual failure mode.
Schedule risk and document intelligence
Schedule-risk tools mine past programmes and live progress to flag activities likely to slip — genuinely useful as a challenge to optimistic lookaheads, dependent on honest progress data. Document intelligence — searching specs, contracts and drawings in natural language, extracting obligations and dates — is quietly one of the highest-ROI uses because the underlying language models are strong and the data (your own documents) already exists. Both categories reward the same thing: disciplined records.
The grounding rule
Generative AI writing diaries, claims, notices and reports is only as trustworthy as its grounding. The standard to hold any vendor to: every AI output should point back to the specific records that justify it — the photo, the diary entry, the delivery docket, the programme snapshot — and should say so explicitly when the data does not exist. An assistant that answers confidently without citing records is a liability generator on a construction contract, where the document trail is the asset.
How to buy AI without being burned
Three questions cut through most pitches. What data does it need, and do we actually produce that data reliably today? What is the operator-verified result on a project like ours — not the pilot press release? And can every output be traced to its evidence? If the answer to the first is "we’d need to change how the site works", budget for that change — the AI is the cheap part. The pattern across every category is identical: AI pays where the records are live and verifiable, and disappoints where they are not.