Vehicle damage claims AI

A customer takes a photo of their damaged car. The AI reads the make, model, color and damage, and gives a rough repair-cost range. Written rules then recommend the next step, and a reviewer makes the call: approve the estimate, ask the customer for better photos, or send the claim to an adjuster.

• Live prototype: https://vehicle-damage-claims-ai.vercel.app
• Evaluation: the Evaluation page in the app, and eval/README.md

What you asked for

You asked for	Where to find it
Accept a photo by upload or URL	+ Add photos: a photo, a folder per claim, or a link
Make, model and color	Top of the assessment card. If it can't tell, it says so instead of guessing
Damage summary	Assessment card, e.g. "Left rear door dent with scraping"
A rough repair estimate	Assessment card: a likely range, the hours and parts behind it, and possible extras listed separately
Setup, architecture, tools, next steps	Setup, Architecture, Why these tools, Next steps
Evaluation	Evaluation

How it works for a claims team

The tool supports the first decision a reviewer makes on a claim. Getting that right means simple claims move faster, customers are asked for the right photos the first time, and serious claims reach an adjuster earlier.

Route	What happens
Ready to approve	The reviewer, a desk appraiser, approves the repair estimate if it's within their approval limit. Payment and booking a repair follow, and the shop can send a supplement if it finds more damage
Request more evidence	The reviewer texts or emails the customer which photos to retake, with an upload link. The claim waits until they reply
Adjuster / total loss	The claim goes to a field adjuster or the total loss unit, plus the fraud team (SIU) if a photo matches a past claim

• Approving commits the estimate, as a desk appraiser does today, up to an approval limit (a setting standing in for the reviewer's authority limit). Payment follows the carrier's usual process; nothing is paid automatically.
• If the AI call fails, the claim goes to manual triage, which is today's normal process.
• Each claim answers its own question ("Why it's ready to approve", "Why it needs more evidence", "Why it goes to a field adjuster") in one line, such as "Because: car can't be driven (S2) and signs of structural damage (S4)". One click shows each rule in full, what the AI saw, and where each fact came from (the policy, the claim form, the AI, the photo checks), with the policy and photo checks underneath.
• The reviewer can change the amount or the route. Changes go back through the rules, and a route change needs a reason.
• The reviewer can ask a colleague for a second opinion. The claim waits under Waiting until they reply, then comes back to the inbox with the reply in the case thread (the demo plays the reply on request). Handing it off ends it on the reviewer's side.
• Each claim has a case thread: the first review, every action, the reviewer's comments, and questions to an assistant that explains the photos but can't change the route. Whoever picks the claim up next sees the whole history.
• Each decision is logged with the recommended route and the reviewer's route, so you can see how often they agree (more).

Setup

The prototype is live at https://vehicle-damage-claims-ai.vercel.app. Click Load demo queue, or drag in the demo-images/ folder (each subfolder is one claim).

To run it locally (Node 22 or later):

npm install
cp .env.example .env.local     # add ANTHROPIC_API_KEY
npm run dev                    # http://localhost:3000
npm test                       # routing rules and URL safety; no API key needed
npm run eval                   # re-run the labelled set; needs a key and costs money

To deploy your own copy, import the repo into Vercel and add ANTHROPIC_API_KEY. Optional settings are in .env.example.

Architecture and data flow

The prototype today

flowchart LR
  rev(["Reviewer"])

  subgraph BR["Browser"]
    ui["Worklist and assessment<br/>state lives in the tab"]
  end

  subgraph VC["Vercel function · nothing stored"]
    direction LR
    prep["① Prepare photos<br/>rotate, resize"]
    checks["② Photo checks<br/>brightness, sharpness,<br/>reused photo"]
    rules["④ Routing protocol<br/>locked rules,<br/>configurable limits"]
  end

  ai["③ Claude<br/>one call per claim<br/>sees photos only"]

  rev --> ui
  ui -- "photos" --> prep --> checks --> ai --> rules
  checks -. "check results" .-> rules
  ui -. "claim and policy details" .-> rules
  rules == "route, reasons, estimate" ==> ui

  classDef person fill:#e0e7ff,stroke:#4f46e5,color:#1e1b4b,stroke-width:1.5px
  classDef screen fill:#f1f5f9,stroke:#475569,color:#0f172a,stroke-width:1.5px
  classDef code fill:#dcfce7,stroke:#16a34a,color:#052e16,stroke-width:1.5px
  classDef model fill:#ffedd5,stroke:#ea580c,color:#431407,stroke-width:1.5px
  class rev person
  class ui screen
  class prep,checks,rules code
  class ai model
  style BR fill:#f8fafc,stroke:#94a3b8,color:#334155
  style VC fill:#f8fafc,stroke:#94a3b8,color:#334155

Color key: green is plain code, orange is the AI, purple is people.

1. Prepare photos. Fix rotation and resize. Links are downloaded by a safe fetcher first.
2. Photo checks. Plain code checks brightness, sharpness, black and white, and whether the photo matches one from a past claim.
3. AI extraction. One Claude call per claim. It fills in a fixed form: the vehicle, each damaged part with how bad it is and whether it needs repair or replacement, what the photos show, and any risk signs.
4. Routing rules. Written rules pick the route from the AI's answers, the photo checks and the claim details.

Key design choices

This is the beginning of the README. Read the rest at the original. Read the full README ↗