How AI helps councils decide which roads need work

A dashboard camera can record the same road an inspector sees. The useful AI question is what happens afterwards: can that footage become a dependable record that helps someone arrange maintenance? West Sussex County Council describes such a trial in an announcement updated on 5 October 2026. It plans to equip eight inspection vehicles with cameras to identify potholes, cracks, surface deterioration and faded markings.[1]

The proposed workflow sends observations into the council’s Confirm asset-management system for inspectors to review and validate. That matters because the existing process includes stopping, photographing and manually recording defects. The trial aims to assess practical improvements; the announcement does not yet report measured savings or safety results.[1]

Turn the picture into a usable observation

Vaisala’s published account of its RoadAI approach provides one example of the mechanism. Imagery collected from a moving vehicle is analysed for different defect types, and the results contribute to condition information for road sections. The camera supplies visual evidence; computer vision converts parts of that evidence into categories a maintenance system can use.[2]

A useful output therefore needs more than a box drawn around something that resembles a pothole. It needs an interpretable defect description and a way to relate observations to a stretch of road. Vaisala’s account illustrates this progression from images to section-level analysis. It is a developer’s description of a particular approach, without independent confirmation here of its advertised productivity benefits.[2]

Automated road measurement already has a history

The Department for Transport’s technical guidance describes established surveys that combine downward-facing images for crack detection with laser measurements of road shape and texture. The crack software can produce maps of individual cracks; other measurements describe features such as ruts or surface irregularity. Different sensors reveal different properties of the same road.[3]

This gives newer camera-based AI a meaningful comparison. Its value depends on the information it supplies and how consistently it supplies it. A visible-crack map, a physical profile measurement and a maintenance condition category should not be treated as interchangeable outputs simply because all three can be collected automatically.[3]

Make different surveys speak the same language

England’s PAS 2161 framework addresses comparability by specifying five condition categories while allowing different approved monitoring technologies. Its national reporting transition concerns classified local roads in England, including local-authority motorways and A, B and C roads. It is not an instruction for every road authority in the UK to adopt AI.[4]

Authorities collect the new-format data during the 2026–2027 financial year for reporting in spring 2027, with transition provisions for existing contracts. The categories relate to the maintenance that a road section may need. The Department for Transport warns that the new five-category results cannot be compared directly with the older red, amber and green series.[4]

That last point matters when reading claims that roads have suddenly improved or deteriorated. A changed measuring system can move the reported numbers even when the physical road has not changed. The statistical question is whether the observations and definitions are comparable before interpreting the movement as a change on the ground.[4]

Test consistency, including the awkward cases

TRL’s report on the 2025 approval trials describes how survey outputs were compared with condition assessments assembled from engineers. The work also examined repeatability: a useful monitoring system should give reasonably consistent results when it surveys the same road again. Neither test is the same thing as counting correctly detected potholes in photographs.[5]

The details prevent an overly simple accuracy claim. Analysis used 100-metre sections partly to reduce location-alignment problems. Providers received calibration information, and Essex was excluded from the final performance evaluation after more of its benchmark data were shared. Sections repaired between surveys were also removed from comparisons. Those choices matter when interpreting how independent and comparable the test results are.[5]

An earlier 2024 TRL network study in Surrey had already identified the underlying tension. Common calibration helped reduce differences between systems, while tuning too closely to a particular test network could make that apparent consistency misleading. The report called for representative networks and robust engineer benchmarks. A condition scale becomes useful through this work on agreement, not merely by naming five categories.[6]

Choose the intervention that fits the road

Blackpool’s Project Amber illustrates why better observation can lead to an initially surprising maintenance choice. The council describes using detailed road imagery to inform surface preservation. Some selected roads can look fairly sound: the purpose is to seal fine cracks before water penetrates and damage spreads. Its account describes a maintenance programme, without establishing a measured AI contribution to the outcome.[7]

Hampshire County Council explains a range of treatments, from patching individual defects and sealing cracks to surface dressing and more extensive resurfacing. A road with repeated pothole problems may need work beyond another local patch. Selecting the treatment therefore requires an understanding of the road’s wider condition, alongside the immediate visible defect.[8]

Consider an invented comparison. One 100-metre section has a single deep pothole; another has widespread fine cracking. A photograph count could rank the first as one defect and the second as many. That count alone does not determine which needs an urgent visit, which treatment is suitable or which job should come first. The example shows why a useful record preserves the defect’s nature and context, rather than only a total.

For a resident, the most useful result would be a clearer connection between what was observed and what the council did next. For an inspector, it could be less time transcribing observations and more time judging them. AI’s contribution can be assessed along that route: a usable observation, a consistent condition record and a maintenance decision someone can explain.[8]

Report an error or suggest a correction