The four-hour claim review now takes six minutes.
CNC Catastrophe & National Claims, a US claims services firm, reviews every adjuster-submitted loss report before it reaches the carrier. Damco Solutions built an AI validation platform for that review: the platform ingests a claim file that can run to 500 pages, extracts the claim data, runs 193 validation checks across 76 categories, and returns a structured QA report. A review that took a validator more than four hours now takes the platform about six minutes.
- less validation time per claim: more than four hours of manual review down to about six minutes
- 97.4%less validation time per claim: more than four hours of manual review down to about six minutes
- validation checks in the live rule set, across 76 categories of claim data
- 193validation checks in the live rule set, across 76 categories of claim data
- validation accuracy today, raised from an 87.14% baseline before the extraction pipeline was re-architected
- ~99%validation accuracy today, raised from an 87.14% baseline before the extraction pipeline was re-architected
Processing time is measured on live claims. The 87.14% baseline was measured on the 20-file test set used for the Phase 1 accuracy review. The engagement began in November 2024 and is delivered fixed-price, in phases.
How it works
A claim file becomes a QA report
A flood claim file is an assembly of 250 to 500 pages: estimate XML, preliminary and final reports, valuation documents, photo sheets. Adjusters upload the same files they always have. The platform reads every page before a validator reads any.
The validator reviews findings, not files. Every flag names its page and its rule, and the verdict stays with CNC.
The estimate XML routinely lands about two hours after the first upload, so the pipeline holds the claim open and validates when the document set is complete, rather than failing the claim for a file that has not arrived.
How we did it
Adjuster expertise, encoded
Validating a loss report means holding 76 categories of claim data to rules that experienced adjusters carry in their heads. Four pieces of engineering carried that expertise into the platform.
Extraction is routed by data type.
The first build leaned on a single general model for every field and scored 87.14% on the 20-file test set. The re-architecture routes each data type to the reader suited to it — document AI for key-value fields, a detector that locates and crops tables for a stronger model to read, pattern matching for structured text — and keeps the general model as the fallback for context-heavy prompts. Errors fell from 108 to 25, accuracy rose to 97.02%, and it has climbed toward 99% since.
The rules encode adjuster expertise.
Insurance data defeats naive matching: three carrier names can be one carrier, flood zones arrive in several formats, and depreciation alone has seven sub-categories. Each of the 193 checks was encoded from CNC's own validators.
Photographs are read as evidence.
Much of a loss report is photographs: interior damage, foundation type, flood height. Vision models judge what the photographs show against what the estimate claims, and a photo-quality rule scores every image on five parameters — blurriness, noise, brightness, contrast, occlusion — and names the exact page where quality fails.
A disputed flag becomes a rule improvement.
Adjuster narratives differ file to file, so the same rule can read two similar claims differently. Every false positive is tracked to a root cause, the domain mappings grow to cover it, and the rule set is tuned weekly. Accuracy improves release by release, and CNC's reviewers see it improve.
What it changed
Review capacity stopped setting the ceiling
CNC runs the platform inside its daily claims operation, and review capacity no longer sets the ceiling on claim volume. Throughput is being scaled further as parallel processing lands. The engagement has grown through four phases since November 2024: property claims joined flood, the photo-quality classifier shipped to production, and an analytics dashboard now tracks every validation category.
- of validator time returned on every claim the platform validates
- 3.9 hrsof validator time returned on every claim the platform validates
- claims a month the platform is equipped to process today, about 74 a day
- 1,600claims a month the platform is equipped to process today, about 74 a day
- claims a month once the current scale-up completes, about 222 a day
- 4,900claims a month once the current scale-up completes, about 222 a day
“What would take us four, five, six hours to review a file now is literally taken by Validate almost instant… So we're getting a file back in no more than 20 to 25 minutes, even if there's delays at other systems.”
If a review queue is holding your operation to the pace of manual reading, the first useful conversation is about which checks a machine should run and which judgments must stay with your reviewers.
We will tell you where automation should stop.