Processing 60+ Insurance Document Variants with Azure AI Document Intelligence

The Challenge

Every carrier and insurer uses a different insurance certificate format. At Corcentric we handled 60+ variants — different field positions, fonts, table structures, and even scanned vs digital PDFs.

Manual entry was costing ~3 minutes per document. At volume, that's thousands of hours annually.

Azure AI Document Intelligence

We used two approaches:

1. Prebuilt Invoice Model (for standard e-invoices)

Azure's prebuilt prebuilt-invoice model covered ~70% of our documents out of the box.

2. Custom Model Training (for the other 30%)

For carrier-specific certificates, we trained custom extraction models:

var client = new DocumentAnalysisClient(
    new Uri(endpoint), 
    new AzureKeyCredential(apiKey)
);
 
var operation = await client.AnalyzeDocumentAsync(
    WaitUntil.Completed,
    modelId: "insurance-cert-v3",
    document: insuranceCertStream
);
 
var result = operation.Value;
 
foreach (var document in result.Documents)
{
    var policyNumber = document.Fields["PolicyNumber"].Value.AsString();
    var coverageAmount = document.Fields["CoverageAmount"].Value.AsDouble();
    var effectiveDate = document.Fields["EffectiveDate"].Value.AsDate();
    
    await _insuranceRepository.UpsertAsync(new InsuranceCertificate
    {
        PolicyNumber = policyNumber,
        CoverageAmount = coverageAmount,
        EffectiveDate = effectiveDate
    });
}

Handling Low-Confidence Extractions

We implemented a confidence-threshold routing system:

• >90% confidence → auto-processed
• 70–90% → flagged for human review in dashboard
• <70% → queued for manual entry with AI-prefilled form

Results

• 85% reduction in manual data entry hours
• 60+ document variants handled by 4 custom models
• Average processing time dropped from 3 min → 8 seconds
• ROI achieved within 3 months of deployment