How I redesigned a dealership warranty-claims workflow at Meridian Auto Group using Claude and Tekion's DMS API — turning a 2.4-hour paperwork slog into a 14-minute conversation the service advisor actually enjoys.
Every claim was a paper chase: pull the repair order from the DMS, type 12 fields into a warranty portal, copy/paste part numbers, look up the contract's covered components, calculate the customer's deductible, submit, wait. Two-and-a-half hours of clicking and cross-referencing.
Then the adjuster might come back and say "that's not covered." The advisor re-opened the contract PDF, re-read the §5.2 language, and resubmitted. Mrs. Chen's coffee was cold.
Multiply this by 80 advisors × 6 claims a day × 40 dealerships. The math hurt.
12+ fields re-typed from the repair order into the warranty portal — by hand.
Claims sat in the adjuster's queue while the customer's car sat on the lift.
30% of claims came back denied for §-language the advisor didn't catch.
Denied claims meant another hour of resubmissions per advisor per week.
Labor hours and part markups drifted outside the negotiated matrix — nobody noticed.
Mrs. Chen left a 2-star review. Her next car wasn't a Subaru, and it wasn't ours.
3 dealership GMs, 2 claims execs, the Tekion integration lead. 90 minutes. Recorded.
30 min each. Claude transcribed, tagged, and clustered themes into a ranked friction map.
Observed actual claim adjudication. Found the "where is §5.2 again?" moment in 11 of 12.
Fed Claude 7 contract templates (800 pages). Asked: "What patterns predict denial?"
Most denials weren't judgment calls. They were the adjuster recognizing that "labor hours for line item X exceed the rate matrix for plan Y." A language task. A pattern-matching task. The exact thing modern LLMs are good at.
The remaining 30% — edge cases, ambiguous fault, customer disputes — would stay human. But the 70%? AI could handle it in seconds, leaving the adjuster to spend their judgment where judgment actually mattered.
The repair order arrives as a PDF (or a Tekion DMS hook fires). AI extracts the claim and pre-fills 12+ form fields.
{vin, jobs[], parts[], labor_hrs}AI compares the submitted claim against the contract's covered components and the negotiated pricing matrix — per job — and recommends approve, adjust, or deny.
{verdict, $rec, reason}AI flags overcharges, drifted labor rates, and uncovered exclusions before the claim reaches the adjudicator — so denials stop being a surprise.
{flags[], adjustments[]}| Customer | VIN | Plan | Submitted | Status |
|---|---|---|---|---|
| Chen, P. | JF2GTACC1MH… | Premium Care Plus | $1,847 | AI Analyzing |
| Hayes, T. | 5YJ3E1EA4LF… | Comprehensive Care | $612 | Approved |
| Velasquez, J. | 1FTFW1ET5DK… | Tire & Wheel | $390 | Adjudicator |
The headline. Service advisors got 2 hours of their day back. Customers got their cars back the same morning.
Same advisor, same day, 12× the output. We didn't add headcount — we removed friction.
Coverage analysis caught the gaps up-front. Advisors stopped resubmitting claims they should never have submitted.
AI caught drifted labor hours and over-matrix parts pricing every shift. Multiplied by 14,000 claims, the math compounded fast.
NPS jumped 38 points in the first quarter. "I was out in 20 minutes" became the most-quoted Google review phrase.
Adjudicators agreed with the AI recommendation 94% of the time. The 6% they overrode? Those were the judgment calls — exactly where we want humans deciding.
42 components, all token-driven. KPI cards, claim table rows, AI hero card, file-upload zone,
per-job breakdown rows, stat badges, action bars. Each ships with a .theme-dark variant for after-hours triage.
WCAG 2.2 AA across light and dark themes. AI-generated content always carries a
aria-live="polite"
announcement when it appears, so screen-reader users hear the reasoning at the same
moment sighted users see it.
I build AI-augmented UX for teams who measure success in minutes saved, not slides shown. If you have a process that's eating your team alive, let's talk.