Industries / Transport and vehicles / Automotive / Collision repair
Estimating
Estimating is writing the list of parts, labor hours and materials a crash repair needs, which the shop and the payer then agree on or argue over.
5
Milestones achieved, from 1963 to 2025.
2
To watch for: announced, not yet confirmed outside.
3
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1963
Repair times reviewed in public
In 1938, when Duke Norman started in the body shop of Robertson Buick in Chicago, the only reference for repair times was carmakers' factory bulletins, which timed the removal and replacement of undamaged parts on undamaged cars, so shop managers and insurance adjusters bargained over each job. Norman timed each operation ten times and took the average, became an editor at a repair-data publisher, National, in 1950, and in January 1958 joined Glenn Mitchell's company, whose estimating books were then sold regionally. From 1963 to 1968 he held meetings across the United States at which shop owners, technicians, dealer associations and insurers reviewed the books' times and operations.
Credit
Duke Norman and Glenn Mitchell, by the account of Gary Ledoux's history of the trade, YesterWreck
1993
Times measured with a stopwatch
Thatcham Research, a not-for-profit research centre in England funded mainly by motor insurers, began researching repair times in the early 1970s because insurers wanted a reference "rather than mere 'opinion'" for agreeing the cost of a repair. Its staff measured a "core time" for each job with a stopwatch under work-study rules, added allowances such as fetching parts from stores, and converted the result to the British Standard 100 rating, the average pace of a trained technician. The Thatcham Time System, released in 1993, gave times for replacing parts and panels and is offered inside most estimating systems.
Credit
Thatcham Research, by the account of the trade magazine Paint & Panel (2010)
2008
Shops report database errors
Estimating systems price labor from databases holding a time for each operation. MOTOR, one of the data providers, says its times are "based on individual operations using new, undamaged OEM replacement parts installed on new, undamaged vehicles". The Database Enhancement Gateway (DEG), set up by shop associations after industry meetings in 2007 and launched in 2008 by Repairer Driven News's account, passes a user's report of a wrong or missing time to the providers behind CCC, Mitchell and Audatex and keeps every inquiry and answer in a public database. Its inquiry 40734, on the 2022–2026 Mercedes-Benz C300 (W206), led MOTOR to raise the time for replacing the rear body panel from 10.2 to 17.5 hours, with the update due in March 2026.
LessonGuide times are measured on undamaged cars, and they stay right only if the people doing the work can send corrections back.
Credit
Society of Collision Repair Specialists (SCRS) and Alliance of Automotive Service Providers (AASP), which fund it; MOTOR for the correction
2016
Work left off estimates
SCRS published a Guide to Complete Repair Planning, in a 2016 edition revised that November, listing panel by panel the operations shops perform that "go unrecognized in the estimate development and final billing processes". CRASH Network and the consultancy Collision Advice survey shops each quarter on which of these "not-included" operations they bill and whether the eight largest US insurers pay. In 2016, 63% of shops said they had never put a seatbelt inspection on a final invoice; in the July 2025 survey of 603 shops, 20% said so, and 49% of shops that charged for it were paid always or most of the time.
LessonWork the carmaker requires is not paid by default; someone has to find it in the procedure and write it into the estimate.
Credit
SCRS; Mike Anderson (Collision Advice) and CRASH Network
2025
Calibrations on more estimates
Driver-assistance cameras and radars often have to be recalibrated, that is re-aimed and tested, after a repair, and a scan reads the car's computers for fault codes. CCC Intelligent Solutions, a claims and estimating software company, counted a calibration on 28.3% of repairable claims in its systems in 2025, up from 21.8% in 2024, and a scan on more than 85% of estimates written by direct-repair shops in 2024. Over 90% of those scans were on the first estimate, but most calibrations were added later on supplements, while the car was already being repaired. In CRASH Network and Collision Advice's October 2024 survey, 90% of shops that billed for a pre-repair scan said the largest insurers paid for it always or most of the time.
Credit
CCC for the counts; CRASH Network and Collision Advice for the survey
Watch for
Photo estimates in minutes
Solera says its Qapter AI estimating platform turns photos of a damaged car into an estimate in about three minutes, which an estimator then reviews, cutting estimate writing from as much as several days to under 20 minutes. It says shops using it have seen repair cycle times improve by 25% and backlogs fall by up to 70%.
Achieved whenan outside party compares a sample of these estimates with the final repair invoices and publishes the gap.
Claimed by
Solera, December 2025
Missed work in 90 seconds
VaraFix says its AI estimate review platform scans a finished estimate for commonly missed operations, materials and procedures and returns supporting documentation in about 90 seconds. It gives no accuracy figures.
Achieved whena published test on real estimates reports how many missed operations it finds and how many it flags wrongly.
Claimed by
VaraFix, August 2026
Open
01
Test set for missed work
A public set of at least 200 collision estimates, with owners' names, VINs and addresses removed and the shops' and owners' consent, each paired with its final invoice, including supplements, and the carmaker procedure pages for the repair. Two trained estimators mark, each on their own, every operation the first estimate missed. Any estimate-review tool, or a general AI model, can then be scored on what it finds and what it flags wrongly.
The briefCredit · unclaimed02
Guide times on wrecked cars
A filmed, timed study of bolt-on replacement jobs (front bumper cover, front fender, front door shell, headlamp, liftgate) done by students or technicians on damaged cars, compared with the times the estimating systems give for the same model, with a DEG inquiry filed wherever a measured time differs by more than 20%. Only outer bolt-on parts, with no welding, airbag parts, high-voltage parts or structural work, and no electric or hybrid cars.
The briefCredit · unclaimed03
Read every DEG inquiry
An open table of all public DEG inquiries, more than 40,000 by 2026 going by the inquiry numbers, giving each one's estimating system, make and model, operation, kind of error (missing time, wrong time, missing part, unclear note), outcome and days to resolve, with a short report on which errors recur across models and systems.
The briefCredit · unclaimed
Open 01 · the brief
Test set for missed work
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of at least 200 collision estimates, with owners' names, VINs and addresses removed and the shops' and owners' consent, each paired with its final invoice, including supplements, and the carmaker procedure pages for the repair. Two trained estimators mark, each on their own, every operation the first estimate missed. Any estimate-review tool, or a general AI model, can then be scored on what it finds and what it flags wrongly.
Why it is still open
First estimates are often incomplete: Mitchell reported that 42.05% of original estimates written with its software in the first quarter of 2018 were supplemented at least once. Review tools are sold without published accuracy, among them DataTouch's P-Pages AI, which marks not-included operations on an insurer's estimate, and VaraFix's platform. The closest data, National AutoBody Research's BillableGenie, is a paid subscription listing line items insurers paid, not a test set, and a search of Hugging Face found car-damage photo sets but no estimate line items.
Built with what exists
Estimates exported in the CIECA standard format, which the CCC, Mitchell and Audatex systems all read and write; the SCRS Guide to Complete Repair Planning as the labeling checklist; an AI model that answers fixed yes-or-no questions about each operation from the estimate and procedure text (see the Note "A thousand questions, one pass"); a public GitHub repository for data and scoring code.
First step
Ask one shop to export 20 closed repair orders that had supplements, redact them, and draft the labeling sheet from the SCRS guide.
Done looks like
200 estimates from at least three shops and 20 makes, the two estimators' agreement reported as Cohen's kappa, and each tool scored for recall (missed operations found) and false flags by operation type, with the scoring code public.
Open 02 · the brief
Guide times on wrecked cars
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A filmed, timed study of bolt-on replacement jobs (front bumper cover, front fender, front door shell, headlamp, liftgate) done by students or technicians on damaged cars, compared with the times the estimating systems give for the same model, with a DEG inquiry filed wherever a measured time differs by more than 20%. Only outer bolt-on parts, with no welding, airbag parts, high-voltage parts or structural work, and no electric or hybrid cars.
Why it is still open
MOTOR says its times are based on new, undamaged parts "installed on new, undamaged vehicles". Thatcham built its British times from stopwatch studies and now works mostly from teardowns of cars in its own workshop and standard allowances. Corrections reach the databases one model at a time through the DEG, as with inquiry 40734's change from 10.2 to 17.5 hours, and a search found no published set of times measured on damaged cars set against guide times. MOTOR's own shop-floor visit in 2025 was described as development work, with no results published.
Built with what exists
Damaged cars already held by a school collision program or a salvage yard; ordinary hand tools; head-mounted cameras; CVAT, an open-source video annotation tool, to mark the start and end of each step; the DEG's inquiry form.
First step
Ask a community-college collision program to time five bumper-cover replacements on one model this month.
Done looks like
30 operations on at least 10 models, each timed at least three times by different people, published with the videos, a table of measured against guide times, and the DEG inquiry numbers and outcomes.
Open 03 · the brief
Read every DEG inquiry
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open table of all public DEG inquiries, more than 40,000 by 2026 going by the inquiry numbers, giving each one's estimating system, make and model, operation, kind of error (missing time, wrong time, missing part, unclear note), outcome and days to resolve, with a short report on which errors recur across models and systems.
Why it is still open
The DEG shows its inquiries one at a time in a web table that can be filtered by estimating system and make. Its administrator reports totals from time to time: more than 1,000 changes from inquiries in 2024, and 30% of 2025's inquiries leading to some change, 40% to 50% for those about CCC. A search found no public table of the inquiries themselves and no count by kind of error across years.
Built with what exists
A bulk export requested from the DEG's administrators, or its public pages collected within its terms; an AI model to sort each inquiry into error types, checked by hand on a sample (see the Note "A thousand questions, one pass"); Python with pandas; a public GitHub repository.
First step
Write to the DEG's administrator to ask for a bulk export and permission to publish it.
Done looks like
Every public inquiry in the table, a hand check of 300 random rows agreeing with the AI labels at least 95% of the time, and a report of the 20 most common error types by system.