SITELIVE 
Capture stopped being the bottleneck
A decade ago, laser scanning a floor plate was a specialist survey engagement priced accordingly. Today a terrestrial scanner such as the Leica RTC360 captures up to two million points per second with a stated 3D point accuracy of 1.9 mm at 10 m, 2.9 mm at 20 m and 5.3 mm at 40 m, and auto-registers setups in the field using a visual inertial system. A wearable such as the NavVis VLX 3 runs two 32-layer LiDAR heads at 1.28 million points per second each and states 5 mm local point-cloud accuracy, while its operator simply walks. Neither instrument is exotic and neither is slow. The technical constraint on as-built data is no longer the sensor.
Robots removed the last human step
The remaining friction was that somebody had to walk the scanner, every week, on a busy site, in addition to their day job — which is why capture cadence collapsed on most projects after the first enthusiastic month. Legged robots fixed that. Turner Construction now runs Spot with DroneDeploy's robotics platform on large data centre builds, executing mapped routes autonomously and repeatably overnight after the trades have left, uploading 360° walkthroughs that are mapped to the floor plan and pushed into the VDC pipeline for comparison against the BIM to catch anomalies before they become rework. Boston Dynamics documents the same pattern at Swinerton, where handheld scans fed AI comparison against the model to produce element-level progress and as-built deviation reports; the robot solved the part the software could not. Once capture is unattended and scheduled, weekly as-built truth becomes a background process instead of a project.
Read the accuracy asterisks, because they decide what the data is for
Instrument specifications are honest but easily misread. NavVis states its 5 mm figure as local accuracy, one sigma, measured in a dedicated 500 m² test environment, and says plainly that absolute accuracy depends on the size of the environment and must be controlled with control points. That caveat is the whole discipline of mobile mapping: SLAM systems accumulate drift over distance, so a point cloud that is beautifully self-consistent in one apartment can sit tens of millimetres out of position at the far end of a 200 m building unless it is constrained to surveyed control. Terrestrial scanners avoid the drift and inherit registration error instead. If the deliverable is a visual record, none of this matters much. If it is a payment claim, a setting-out check or a clash resolution against steel, control is the difference between evidence and decoration.
Comparing a scan to a model needs a tolerance policy
Automated scan-versus-BIM analysis reports deviation, and deviation is not the same thing as defect. Concrete, steel and blockwork are built to construction tolerances measured in millimetres to tens of millimetres; a design model is a geometric idealisation that usually contains none of them. Without an agreed tolerance per element type, an AI deviation report will confidently flag hundreds of elements that are entirely compliant, the team will lose faith in the second report and the tool dies. The projects that make this work write the tolerance policy first, agree who adjudicates a flagged element, and treat the deviation register as a formal quality record with a name against each disposition.
A point cloud is not a twin
The word "twin" is doing enormous unearned work in this industry. Useful contrast: Woodside Energy's Fuse platform, documented in a 2024 ISPRS paper, was first deployed at the Pluto LNG facility in north-western Australia, where it combines data from more than 200,000 industrial IoT sensors embedded in the plant with custom wireless sensors and enterprise systems, presented through 3D views built from CAD models, LiDAR scans and photogrammetry, plus 2D maps, mobile and augmented reality. Woodside describes Fuse as core technology inside its Pluto Remote Operations Centre, which operates the offshore and onshore assets from Perth. What makes that a twin is not the geometry — it is the live sensor feed, the workflows executed against it and the operating decisions taken from it. A handover point cloud has none of those three.
Twin decay is the honest limit
Every digital twin is true only as at its last update, and updating is a cost nobody budgets after practical completion. The asset changes: a valve is replaced, a wall is penetrated, a tenant fit-out reroutes services, and the model quietly stops describing the building while continuing to look authoritative — which is more dangerous than having no model at all. This is precisely what the ISO 19650 series addresses by treating information as something specified, exchanged, versioned and maintained across the whole asset lifecycle rather than delivered once in a box, and why the Gemini Principles argue for governance before technology. Ask any twin vendor the unglamorous question: who updates this in year three, on what trigger, funded by which budget line?
The economics, plainly
Per-scan cost is now trivial; the recurring costs are storage, processing, licences, and the human loop required to act on what the comparison finds. Value therefore accrues only where a scan is attached to a decision with money on it: verifying a progress claim before certification, catching a services clash before the ceiling closes, proving as-built position before a prefabricated element is fabricated, or planning a shutdown against measured rather than assumed geometry. Capture cadence with no decision attached is an expensive photo album — and it is the single most common way reality-capture budgets are wasted.
Where the public-sector twins fit
At the other end of the scale, jurisdictions are building twins as shared infrastructure rather than project deliverables. The NSW Spatial Digital Twin, led by DCS Spatial Services, is a cross-sector 4D environment — 3D plus time — that publishes and visualises government spatial data through a shared platform with a published policy framework governing how agencies contribute. It will not tell you whether your slab is level, and it is not meant to. It matters to contractors for a different reason: as owners and regulators normalise expectations of maintained, queryable spatial records, the standard for what a project must hand over at completion moves with them.
What it means on site
Set a capture cadence tied to the commercial calendar, not to enthusiasm — typically weekly, aligned to the claim cycle and the lookahead. Constrain every capture to surveyed control so the data can be used for more than pictures. Write the tolerance policy before the first deviation report. And treat the outputs as records with provenance — who captured it, when, against which model revision — because the value of as-built data is realised years later, in a dispute, a retrofit or a shutdown, by someone who was not there.