ACORD Form Extraction

The @openinsure/acord package provides typed data models for ACORD standard forms and an AI extraction pipeline that converts uploaded PDFs and images into structured submission data.

---

Supported Forms

Form	Title	Use
ACORD 125	Commercial Lines Application	GL, Property, Umbrella new business
ACORD 200	Commercial Lines — Workers Compensation	WC new business
ACORD 126	Commercial General Liability Section	Standalone GL supplement
Supplemental	Program-specific supplements	Additional risk questions per program

---

Data Model

Every extracted form produces a typed object:

import type { ACORD125, ACORD200, ExtractionResult } from '@openinsure/acord';

interface ACORD125 {
  // Applicant
  namedInsured: string;
  mailingAddress: Address;
  naicsCode?: string;
  annualRevenue?: number;
  yearsInBusiness?: number;

  // Coverage requested
  requestedLines: ('GL' | 'PROPERTY' | 'UMBRELLA' | 'CRIME' | 'INLAND_MARINE')[];
  occurrenceLimit: number;
  aggregateLimit: number;
  deductible: number;
  effectiveDate: string;

  // Loss history
  lossHistory: LossYear[];

  // Extraction metadata
  extractionConfidence: number; // 0.0 – 1.0
  rawText: string; // Full OCR text for manual review fallback
  sourceFile: string; // R2 key of the uploaded PDF
}

extractionConfidence is an aggregate score (0–1) calculated per field and averaged across the form. A score below 0.75 triggers a review flag in the UW queue — an underwriter manually verifies the highlighted fields before the submission is created.

---

Extraction Pipeline

1. Upload — Producer uploads a PDF via the portal or it arrives via the mailbox ingest cron 2. OCR — The PDF is rendered to images via the Browser binding and passed to Workers AI (Llama 3 vision) for OCR text extraction 3. Field extraction — Claude (claude-haiku-4-5) maps OCR text to the ACORD125 / ACORD200 schema using a structured output prompt 4. Validation — Zod schema validation catches missing required fields and type mismatches 5. Confidence scoring — Per-field confidence is calculated based on regex match strength and Claude’s self-reported uncertainty 6. Result — Structured ExtractionResult stored in PlanetScale; source PDF stored in R2

---

API

Submit a form for extraction

POST /v1/submissions/:id/extract-acord
Authorization: Bearer <token>
Content-Type: multipart/form-data

file=@acord125.pdf
formType=ACORD125

Response:

{
  "extractionId": "ext_01J8...",
  "status": "processing",
  "estimatedSeconds": 15
}

Poll extraction status

GET /v1/submissions/:id/extract-acord/:extractionId

{
  "status": "complete",
  "extractionConfidence": 0.87,
  "reviewRequired": false,
  "data": {
    "namedInsured": "Acme Roofing LLC",
    "naicsCode": "238160",
    "annualRevenue": 2500000,
    ...
  }
}

Auto-populate submission

Once extraction is complete with confidence ≥ 0.75, the submission fields are auto-populated:

POST /v1/submissions/:id/apply-extraction/:extractionId

This merges the extracted data into the draft submission. Low-confidence fields are marked with ⚠ Review in the producer portal form.

---

Review Queue

Extractions with confidence < 0.75 appear in the UW Review Queue under Extraction Review. The UW:

1. Sees the original PDF alongside the extracted fields
2. Corrects any flagged fields inline
3. Clicks Approve — the corrected data is merged and the submission proceeds

---

Manual ACORD Input

If no PDF is uploaded, producers can enter ACORD data directly through the submission form in the producer portal. The form follows the ACORD 125/200 field layout and validates the same schema.

---

Adding a New Form Type

1. Add the form’s TypeScript interface in packages/acord/src/types/{form}.ts 2. Create the extraction prompt in packages/acord/src/prompts/{form}.ts 3. Add the form to the SUPPORTED_FORMS registry in packages/acord/src/index.ts 4. Write extraction tests using fixtures in packages/acord/src/__fixtures__/ (sample PDFs + expected output) 5. Update this page with the new form in the Supported Forms table