The model can rank the cracks, but an inspector still signs for the road

Every few months a supplier shows me a video of a vehicle driving along a motorway while coloured boxes appear around cracks, potholes and faded markings. The demonstration is usually impressive and the accuracy figure is usually above ninety per cent. I have learned to ask three questions before the lights come back on: ninety per cent of what, measured against whose judgement, and what happened to the defects the model did not box?

My position is that image-based defect detection is worth having on a road network, but its proper role is to decide the order in which qualified people look at things, not to replace their judgement. Contractual and safety responsibility for a pavement or a structure does not move to a model because the model is usually right. If anything, the better a model performs on average, the more carefully an owner needs to understand the cases where it fails.

What we tested and what it found

In 2025 we mounted cameras on two of our routine patrol vehicles and ran a defect-detection model over 310 kilometres of carriageway on our interior Portuguese concession for six months. The patrol vehicles already drive the network daily, so the images cost little to collect. The model divided the pavement into 20-metre segments and flagged those it judged to contain cracking, potholes, edge deterioration or worn markings, with a severity score for each.

Over the trial it flagged about 4,100 segments. Our pavement inspectors reviewed a sample and agreed that the flagged defect was real and correctly classified in 71 per cent of cases. Most disagreements concerned severity rather than existence: the model was more alarmed by surface cracking than our engineers were, and less alarmed by edge deterioration near drainage, which matters more to the structure of the road.

That figure was useful, but it was not the one I cared about most. A model that flags too much wastes inspectors' time; a model that misses a severe defect may leave a hazard on a carriageway. To measure misses, we asked inspectors to walk a blind sample of 600 segments the model had not flagged, without knowing which segments they were. They found severe defects in 14 of them. Across the combined sample, the model had detected 93 per cent of the severe defects present. That sounds reassuring until one remembers that the remaining seven per cent are exactly the defects an owner most needs to find.

Two of the missed severe defects were under standing water after rain. Three were in a section where a camera had been knocked out of alignment and nobody had noticed for nine days. Neither was a failure of the algorithm in the narrow sense. Both were failures of the system the algorithm sits in.

That distinction matters for how an owner assigns responsibility. A camera bracket is a maintenance item, and so is the routine that checks image quality before the model runs. We now reject any patrol run in which a sample of frames fails a simple alignment and exposure check, and the patrol supervisor, not the data team, owns that check. It is unglamorous work, and it prevented two further blind spots in the following quarter.

Measure the misses, not the headline accuracy

The accuracy figures in supplier presentations are often calculated over every segment, most of which contain no defect at all. On a well-maintained road, a model that declared every segment healthy would score highly on that basis and be useless. The measures that matter to an owner are narrower: of the defects that are really there, how many does it find, broken down by severity; and of the segments it flags, how many deserve a visit. Both need measuring against our own inspectors on our own roads, because a model trained on one country's surfaces, markings and weather does not automatically transfer to another's.

They also need measuring again. A camera replacement, a new surfacing material, a change of season or a different patrol speed can each shift performance without anyone touching the model. We now repeat the blind-sample exercise every quarter and treat a fall in the detection of severe defects as an incident, with the same seriousness as a failed sensor in a tunnel.

There is a commercial point as well. Several suppliers proposed pricing linked to the number of defects detected. I declined. An incentive to find more defects is not the same as an incentive to find the right ones, and it would have pushed the model towards exactly the over-flagging our inspectors were already filtering out.

Where the signature stays

After the pilot we changed our inspection practice in one specific way. Flagged segments are reviewed within a period set by their severity score, and the model's output is used to order the inspectors' routes. Routine visual inspections on the planned schedule continue unchanged, and the principal inspections of bridges and tunnels are not touched by the model at all. The engineer who records a defect, decides a repair priority or certifies a condition still signs for it, and the record shows whether the model was involved.

Some colleagues saw this as excessive caution. The business case for the pilot had assumed that inspection hours could be cut by about a third. In practice we have used the time differently: inspectors spend less of it driving past healthy pavement and more of it on the segments that need judgement. Total hours have fallen by about twelve per cent, and I do not expect much more while we are measuring misses properly.

The handback of our coastal concession in 2031 makes the boundary more important, not less. The grantor will assess condition against the contract, through its own inspectors. If our maintenance plan has been shaped by a model whose blind spots we did not understand, we will discover them at the most expensive possible moment.

I expect these tools to improve, and I expect to use them more widely. The cameras are cheap, the patrol vehicles drive the road every day, and earlier detection of deterioration is valuable. But the question a board should ask about artificial intelligence in maintenance is not how accurate the model is. It is who signs for the asset, what evidence they rely on, and how the organisation would know if the model began to miss what matters.