Two inspectors, one car, two reports
An inspection business sells consistency; a remarketing platform sells a listing the buyer will not dispute. Both are judged on the same thing — two qualified people looking at one vehicle should write the same report — and the variation between them is rarely about competence. One inspector calls a panel scuffed and the next calls it scratched; one shoots the wing at an angle that hides the crease; one takes the trim from the seller's description and the other from the car.
By the time a buyer challenges the listing, the evidence that would settle it was never captured to a standard that can be compared.
What changes, workflow by workflow
Identity and specification. CarScope reads the plate, the chassis number and windscreen markings from the photographs themselves, and profiles body type, trim, model-year range and equipment. Tyre size, tread depth and the odometer come from the close-ups. The listing spec sheet is assembled from the car rather than typed from the seller's account of it.
Capture quality. ImageScope states, per photograph, the axes reviewers otherwise argue about afterwards — lighting, blur, surface reflectivity, viewing angle, camera distance, and whether the frame shows the exterior or the cabin. Reflectivity is the one that decides whether a highlight on a panel gets read as damage.
Damage to one standard. DamageScope masks each damage area at pixel level, classifies it against a closed set of types, and attributes it to a named exterior part from a taxonomy of more than 30 parts. Severity is derived from the type and the measured share of the panel affected, not from an opinion, so the same photograph produces the same band. Each finding carries a repair-or-replace decision, a suggested method and estimated labour hours.
Paperwork read to one standard. ArchDoc grades whether a document image is readable before reading it, then extracts the fields that matter for that document type, with the source image attached to the value. An identifier that cannot be found in the text printed on the page is flagged for confirmation rather than passed on as certain.
The second reading. ExpertArch reads the finished file — photographs, damage findings, vehicle profile and document fields together — and separates what is verified from what is uncertain and what is missing. The report leaves with its gaps already named, which is cheaper than a buyer naming them.
The artefact is the proposition
A condition report, a listing spec sheet, a dispute file: none of them improve because a model is accurate in the abstract. They improve because every line is structured, comparable and attached to the image it came from — so the report stops being what one inspector wrote and becomes a record the next inspector, the buyer and the adjudicator read the same way.
Where the person stays, and what is refused
The platform returns nothing rather than guess a plate it cannot read cleanly, and deep reads need the close-up: no chassis-number shot, no chassis number; no tyre shot, no tread depth. "Nothing found" and "could not be assessed" stay separate outcomes, and the unusable images are named, so silence is never mistaken for a clean vehicle.
No grade is issued and nothing is auto-approved. Every value carries a confidence score, which is what lets each firm set where its own automation stops — but the grade, the price and the signature stay with the inspector.
How it starts
The workflows are independently callable, so a first deployment need not be the whole pipeline. Most operations start where the record is weakest — capture grading and vehicle identification, on the photographs their people already take — then extend into damage and documents once the output has been compared against their own completed files.