all research

Construction Management · 9 min read

Closing the loop between site and model: reality capture as a live data asset

Drone and 360 capture produces enormous datasets that mostly get archived. Agents turn that data into deviation detection, and deviation detection into design action.

8 days → 4 hrs

Capture to deviation report

23mm

Detectable deviation threshold

£1.9m

Modelled rework avoided

Most large sites now capture themselves weekly. Drone photogrammetry, 360 walkthroughs, laser scanning at key milestones. The data volume is substantial and the capture cost has fallen far enough that it is close to routine.

What happens to that data afterwards is the problem. It goes into a folder. Someone looks at it when there is a dispute. The comparison against the model that would have caught a problem early is a specialist task that costs money and takes a week, so it happens at milestones rather than continuously.

Continuous comparison changes what gets caught

An agent doing registration and comparison automatically shifts this from a milestone exercise to a weekly one. The technical work is not new. Point cloud registration and deviation analysis have existed for years. What is new is removing the human coordination cost that limited how often it ran.

Capture to actionable deviation report
hours
Data transfer and preparation80.594%
Registration to project datum120.794%
Model comparison201.294%
Deviation classification140.894%
Report and route to discipline100.892%
conventional agent-assisted

Eight days becomes about four hours. The interesting consequence is not the time saving. It is that a four hour turnaround means capture can happen weekly and still be acted on inside the same week, which puts deviation detection ahead of the work that would build on top of the deviation.

Modelled cost to correct a positional deviation, by weeks between occurrence and detection
£ thousand per instance
4Week 111Week 238Week 496Week 8178Week 12340At handover
observed range modelled projection

Treating engineering data as an asset

The wider shift underneath all of this is that structured data starts being managed as a primary output of engineering rather than a byproduct. Sensor readings, inspection records and as-built models are all describing the same physical thing at different moments, and connecting them enables a question that has historically been very hard to ask: is this element behaving the way we assumed it would.

The scan is not the deliverable. The difference between the scan and the model is the deliverable.

Practical constraints worth knowing early

Deviation thresholds have to be set against what matters rather than what is detectable. Modern capture will happily report every point that sits 5mm from where the model says it should be, and a report containing forty thousand deviations is functionally identical to no report. Thresholds should come from the tolerance in the specification, and classification should distinguish between a deviation that affects a following trade and one that does not.

Registration accuracy is the other constraint. If the point cloud is not tied properly to the project datum, everything downstream is measuring the registration error rather than the building. This is the single most common reason these deployments produce noise instead of insight.

Where to begin

Structural frame position on a single storey, compared weekly, with a threshold taken from the structural specification. It is narrow, the tolerance is unambiguous, the consequence of a missed deviation is well understood by everyone on the project, and it produces a result inside a month that either works or clearly does not.

Work with me

Run this model against your own project

I am Kanishk Kapoor, Technical Accounts Manager at AI Institute in Dublin. I build agentic AI systems with built-environment teams across Ireland and the UK. If any figure here looks wrong for your business, that is the useful conversation. Send me your assumptions and I will re-run it.

Continue reading