AI warranty claim coding for OEMs and dealer networks: one model for every dealership

A warranty claim or a repair-order line starts as free text: the customer's concern, the technician's story of complaint, cause and correction. Someone then picks a code. Service advisors, technicians and warranty administrators turn over, each store keeps its own shortcuts, and a code list with hundreds of entries leaves room for honest disagreement.

Claim-validation tools at the dealership help with part of this: they check a claim against the OEM's rules before it's sent. They don't change who decides the code in the first place. The arrangement below puts that decision in the hands of one model trained on the OEM's own coding.

The training data is the OEM's own history: claims, repair-order lines or tech-line cases its warranty analysts or field engineers have already coded. That history is the most consistent version of the coding standard that exists, because the people who wrote the standard applied it.

StayCharted learns from it, checks the examples for contradictions first (the same text under two codes, codes with too few examples to learn), and tests the model on cases it never saw. The report shows how often it's right for each code and which codes it confuses, before any dealer sees an answer.

The model is published once, in the OEM's workspace, and every channel asks that one published version:

A dealership sees the answer, its confidence, and nothing else. Its access code reaches that one model and can ask it questions. It can't see the training data, other models, other dealerships' answers, or the review queue. Dealers don't need a StayCharted account, and the OEM can stop a code at any time.

For the dealership, the change is small: send the text it already writes and use the code it gets back.

One coding standard, applied the same way everywhere. Every store uses the same published model and coding standard. Different descriptions can still produce different answers; the report and review queue help your specialists find and correct those differences.

Every unsure answer in one place. Answers below the confidence cutoff go to the OEM's Review Queue, marked with the channel and the dealership's code name, so a reviewer sees "Mobile app (code: Dealer 1042 service)" next to the case. The specialists who own the standard decide those cases, wherever in the network they came from.

Corrections that improve the model for everyone. Each decision in the Review Queue becomes a training example. When the OEM publishes the next version, every dealership's next answer comes from it, with nothing to roll out store by store.

No gap while a new version trains. The published version keeps answering every dealership while the next one trains, and switches only when the OEM publishes the new one after reading its report.

Start with one of these, for one brand or region, where a code is chosen from free text, the volume is high, and a wrong code costs money or hides a trend:

• Warranty claim coding: the technician's complaint, cause and correction coded to the failure or labor code before the claim is submitted, with the unsure ones checked by the OEM.
• Labor operation code standardization: customer concern and repair-order lines coded to a standard service category or operation, so service data is comparable store to store.
• Technical assistance (tech line) cases: a dealer's request routed to the right engineering group on arrival.
• Field quality reports: dealer reports filed under the OEM's component codes, so quality engineers count from complete, consistent data.
• Parts and damage photos: with picture models, a technician's photo of a returned part or of shipping damage classified the same way.

A dealer group with many rooftops has the same problem on a smaller scale, and the same answer works. The group's fixed-operations team trains one model on the lines its best stores already coded, gives each rooftop an access code or connects the group's own reporting to the API, and works the unsure lines in one Review Queue. The group's service reporting then compares stores on the same codes. Where an OEM's codes apply, the group keeps them; for its own internal categories, it trains on its own.

We tested it on the closest public data: consumer complaints NHTSA received about vehicles. We trained on complaints from earlier years, then asked the models to file 5,000 complaints NHTSA received from January 2025 on, from the driver's own words alone, and compared each answer with where NHTSA filed it.

At the subsystem level, with 70 categories, the StayCharted Dedicated AI Model was sure enough of 71% of complaints to file them on its own, and was right on 94% of those. The other 29% would go to a person. This benchmark did not test repair orders or warranty claims. Your codes and descriptions may be harder or easier; the report on your own data is what decides. Read the benchmark →

Each dealership's access code has its own limits, set by the OEM:

• How many phones can use it (10 by default).
• How many answers a day across those phones (500 by default).
• When it expires (90 days by default, any number of days, or never).
• Photos allowed or not.
• Pause, turn back on, or revoke for good, and remove a single phone.

Every answer counts against the OEM's workspace, so the OEM sees and pays for the network's use in one place. Systems that call the API use the workspace's API keys, which the OEM creates and can revoke.

Access codes require Essentials or Business and the workspace Owner’s permission. Usage draws from workspace allowances. For a large dealer network, contact us to agree capacity and limits before rollout.

• Your own database schema. The OEM's training data, answers and Review Queue are kept in a database schema that holds only its organization's data.
• Encrypted at rest and in transit, hosted on AWS in the United States.
• Used only for your models. The data trains the OEM's models and nobody else's.
• Nothing is collected from dealer systems. StayCharted only answers what it's sent, and never connects to a dealer management system.
• Your own cloud, if you prefer. On the Business plan, models can be trained in your own Amazon Bedrock or Microsoft Foundry account.

Security and your data →

Training in your cloud does not bypass StayCharted: the training data still passes through StayCharted to prepare and run the workflow. See how cloud training works.

1. Week 1: export a few thousand claims or lines your specialists coded in the last two years, with the text and the code. Remove customer names, addresses, phone numbers and VINs.
2. Week 1: train, read the report per code, and pick the confidence cutoff where the answers above it are right often enough to go straight through.
3. Weeks 2 to 3: give access codes to ten dealerships, or connect one dealer-facing system to the API. Your specialists work the Review Queue a few times a week.
4. Week 4: publish the version trained on their corrections, and compare its report with the first.

At the end you'll know, for your own codes, what share of cases can be coded at the dealership with no OEM review, and how often those codes are right.

Start free → · Talk to us about a dealer network pilot →