Manual road surveys vs AI video audits: an honest comparison

We sell one of these two things, so read this with the appropriate suspicion. What follows is the comparison we would want if we were on the buying side — including the cases where the walked inspection is simply the better instrument and we would say so in the room.

The two methods are not competing for the same job

Most comparisons in this category go wrong in the first paragraph, by treating a formal road safety audit and a road condition survey as one activity. They are not.

A safety audit is an engineering judgement: a qualified auditor examines a design or an in-service road and forms an opinion about crash risk — sightlines, junction geometry, conflict points, whether the layout invites the wrong behaviour. That is reasoning about causes, and no detection model performs it.

A condition survey is measurement: what is broken, how badly, where. It is the evidence base an audit and a maintenance programme both stand on, and it is almost entirely data collection.

Automation applies overwhelmingly to the second. The honest claim is not that AI replaces auditors; it is that it replaces the weeks of clipboard work that make good audits expensive and therefore rare.

Side by side, on the parts that overlap

On the condition-survey half of the job the two methods trade off cleanly: manual inspection wins on perception and adaptability, camera capture wins on speed, consistency and traceability. Neither dominates the other. Which is correct depends on whether the constraint you are actually fighting is attention or coverage.

Where manual inspection genuinely wins

Manual inspection remains the better instrument in four situations: anything below the road surface, design-stage audits where no road exists to drive, crash-site investigation that requires causal reasoning, and off-carriageway hazards a windscreen camera never points at. In these, automation adds nothing worth the cost.

• Anything below the surface. Subgrade failure, drainage voids, and structural capacity are not visual properties. A surveyor standing on a road after rain learns things no frame contains.
• Design-stage audits. There is no road to drive yet. The entire exercise is reading drawings and imagining behaviour.
• Crash-site investigation. Understanding why a particular junction keeps producing collisions is causal reasoning about human behaviour, not defect counting.
• Off-carriageway hazards. A dashcam is pointed at the road. Verges, embankment condition, drainage inlets and sightline obstructions behind vegetation are partly or entirely outside its view.

Where camera-based auditing wins outright

The advantage is not accuracy on any single defect. A careful inspector standing over a pothole will beat a model every time. The advantage is everything that follows from cost collapsing.

Frequency. When a survey costs a drive rather than an expedition, it stops being an annual event. That changes what the data is for: a once-a-year snapshot tells you the state of the network, while a monthly series tells you the rate of change — which is what actually predicts where next year’s failures will be.

Consistency. Two inspectors grade the same crack differently, and one inspector grades differently at 9am and 4pm. A model is wrong in the same way everywhere, which sounds worse and is in fact far more useful — systematic error can be corrected for, while inconsistent error cannot.

Defensibility. Every detection ties to a timestamped frame at a coordinate. When a claim arrives, or a regulator asks what was inspected and when, that trail exists without anyone reconstructing it from memory.

Coverage. Manual programmes ration attention to the corridors someone already suspects are bad. Automated ones can cover the whole network, which is the only way the unexpected problem gets found.

The cost comparison people get wrong

Per-kilometre survey cost is the wrong measure. It omits traffic management, which can exceed the inspection itself; it omits crew scheduling overhead; and it omits the largest cost of all — defects found too late, which turn patching into resurfacing. It also hides the automated method’s real expense: review time.

The usual framing is per-kilometre survey cost, and it understates the gap in both directions. It ignores traffic management, which can exceed the inspection itself. It ignores the coordination overhead of scheduling crews. And it ignores the largest cost of all, which never appears on any survey line item: the defects that were not found in time and became resurfacing instead of patching.

Set against that, the automated method has a real cost the brochures skip — review time. The output is a draft that a competent person still needs to look through, dismissing false positives and sanity-checking severity. Budget for it. A tool that pretends this step does not exist is telling you something about how much it trusts its own output.

How the decision usually resolves

In practice nobody picks one. The pattern that works is automated collection feeding human judgement: drive the network to find and grade what is there, then send engineers to the places the data says are worth an engineer’s time.

That inverts the current default, where the scarcest resource — qualified attention — is spent on data collection, and the corridors nobody suspected go unlooked-at for years. If you want to see the output format that hand-off produces, the public sample audit is unlocked, and the use-cases page sets out how different network owners tend to split the work.