Data Extraction from PDFs for Insurance Carriers: A Step-by-Step Guide

PDF data extraction for insurance carriers cuts claims intake from 4 hours to under 30 minutes per document batch. It is the automated process of reading, parsing, and structuring data from incoming documents so carriers can process policies, claims, and endorsements without re-keying a single field. See our automation guides library for related workflows.

What this workflow looks like before automation

Today, most carriers have someone (an adjuster, an underwriting analyst, or a data entry clerk) opening PDFs and typing what they read into Guidewire, Duck Creek, Majesco, or PolicyCenter. The steps look like this:

1. Receive and open the PDF. An analyst pulls an ACORD form, loss run, or medical bill from email or a shared drive. (5-10 minutes per document)
2. Read and locate the relevant fields. The analyst scans the document to identify what matters: policy number, claimant details, coverage limits, effective dates, loss description. Commercial lines submissions can have 30-50 fields. (10-20 minutes per document)
3. Key data into the core system. Each field gets typed into Guidewire, Duck Creek, or the carrier's policy system by hand. Transposition errors are common at volume: a wrong digit in a policy number, a misspelled claimant name. (20-40 minutes per document)
4. Validate the entry. A second reviewer compares the system record against the source PDF to catch errors before the record moves forward. (5-15 minutes per document)

Total time per document: 40-85 minutes. For a mid-size carrier processing 200 submissions per week, that is 130-280 staff hours per week in data entry before any actual underwriting or claims work begins.

What the automated version looks like

The automated pipeline uses Azure AI Document Intelligence for extraction and Power Automate for orchestration. Here is how a document moves through the system:

1. Intake. Azure AI Document Intelligence ingests the PDF directly from your email inbox, SharePoint folder, or claims portal upload. No manual download required.
2. Document classification. The model identifies the document type: ACORD form, loss run, medical bill, endorsement request. Each type routes to its own extraction model.
3. Field extraction. Structured fields are pulled automatically: policy number, effective dates, coverage amounts, claimant name, date of loss, loss description. Pretrained models cover standard ACORD forms. Custom models trained on your historical documents handle proprietary or specialty-lines formats.
4. System mapping via Power Automate. Extracted fields map to your Guidewire, Duck Creek, or Majesco record schema. Power Automate writes the data to your core system via API connector.
5. HITL checkpoint: low-confidence fields. Any field with a confidence score below your set threshold (typically 85-90%) is flagged and routed to a human reviewer in a structured approval queue. The workflow pauses until a reviewer confirms or corrects that field. Nothing auto-writes to your system of record without meeting the confidence requirement.
6. HITL checkpoint: anomalous formats. Documents with poor scan quality, unfamiliar layouts, or heavy handwriting route to a specialist queue rather than through the automated extraction path. A human handles them directly.
7. Write and log. Validated records write to your core system. An audit log records what the AI extracted, what was human-reviewed, and what was auto-approved, supporting GLBA compliance requirements.

The human stays in the loop at the two points that matter most: uncertain AI output and genuinely difficult documents. Everything else processes without waiting for a person.

What insurance carriers typically save

The savings estimate for this automation is a 70-90% reduction in data entry time. For an insurance carrier, that translates to concrete numbers.

Take a carrier processing 200 PDF submissions per week at 45 minutes per document on average. That is 150 staff hours per week. At a fully loaded cost of $35 per hour for data entry staff, that is $5,250 per week, or roughly $273,000 per year in direct labor.

Cutting entry time by 70-90% turns that 150-hour weekly burden into 15-45 hours. The remaining time goes to the HITL review queue: the documents where analyst judgment is needed, which is where that time should go anyway.

Accuracy improves as well. Manual keying typically produces a 1-4% error rate on high-volume data entry. Errors in claims data (a wrong date of loss, a transposed policy number) can trigger coverage disputes, delayed payments, or state DOI complaints. Automated extraction with confidence scoring brings field-level accuracy above 97% for structured documents, with the HITL checkpoint catching the rest.

Faster intake means faster claims decisions. Shortening FNOL (First Notice of Loss) intake from 2 days to under 4 hours affects customer retention and compliance timelines in states that require formal acknowledgment within specific windows after first notice.

The tools we use to build this

We build PDF extraction pipelines for insurance carriers on three Microsoft services. Here is what each does and why it fits your compliance requirements:

Azure AI Document Intelligence

Microsoft's OCR and form recognition service. Pretrained models cover ACORD forms, invoices, and standard financial documents. For custom layouts (specialty lines submissions, proprietary claim forms, loss runs from prior carriers) we train custom models on your historical documents. Because it runs in Azure, your documents stay within your cloud tenant. That matters for GLBA compliance: data does not transit through a third-party SaaS platform you do not control. See the Azure AI Document Intelligence documentation for the full list of supported document types.

Power Automate

Handles workflow orchestration: routing documents from intake to Document Intelligence, mapping extracted fields to your core system, managing the HITL approval queue, and writing audit logs. Power Automate connects natively to Microsoft 365 and SharePoint, with API connectors for Guidewire, Duck Creek, and other major insurance platforms. For health lines carriers subject to HIPAA, Power Automate data handling is covered under Microsoft's Business Associate Agreement.

Azure AI Foundry

For carriers that need a reasoning step on top of field extraction (categorizing a free-text loss description, summarizing a lengthy medical record for a claims adjuster) we add a language model step using Azure AI Foundry. The entire pipeline stays inside the Microsoft Azure boundary, simplifying state DOI and HIPAA compliance reviews. Learn more about our AI agent services for insurance carriers.

Where this breaks down

Here is where automated PDF extraction has real limits. Buyers who have been oversold on AI tend to find out the hard way, so we are direct about it.

Handwritten documents

Document Intelligence handles printed and typed text accurately. Handwritten FNOL forms, field adjuster notes, or handwritten endorsements reduce extraction accuracy significantly. If more than 20% of your incoming documents are handwritten, expect a larger HITL queue and a weaker ROI case. Fully handwritten documents should route directly to human review.

Complex multi-page loss runs

A loss run from a large commercial account can be 40 pages with inconsistent formatting across multiple prior carriers. First-pass accuracy on less structured layouts is lower and custom model training time increases. Budget extra time for UAT on complex document types before go-live.

Poor scan quality

Faxed documents, double-sided copies on a slow feeder, or documents with heavy notations reduce OCR quality. We recommend a scan quality threshold: documents below a minimum resolution or with excessive image noise route to human review automatically rather than through the extraction pipeline.

State filing and regulatory requirements

Some state DOI requirements mandate human review before a claim is formally acknowledged or a policy is issued. Automation does not remove those requirements. Your compliance team needs to identify where human sign-off is a regulatory requirement, not just an internal quality gate. See the NAIC guidance on data and technology in insurance for relevant compliance context.

How long to build and what it costs

A standard implementation (Document Intelligence connected to Power Automate, writing to one core system, with a HITL approval queue) takes 6-10 weeks to build and test. This covers document model training on your specific forms, system integration, and UAT with your claims or underwriting team.

A more complex build with multiple document types, multiple downstream systems, and a Copilot Studio interface for the HITL review queue runs 14-20 weeks.

Typical project cost: $40,000-$250,000 depending on scope, number of document types, and integration complexity. See our full PDF data extraction cost guide for a detailed breakdown by project type.

Related work we have done

We have built document extraction and workflow automation for carriers and adjacent regulated industries. Our work in this area covers commercial lines submission intake, FNOL processing, and medical bill parsing for health lines carriers. While we do not publish all client details, we are happy to walk through comparable builds on a call. For a broader view of document automation across industries, see our automation guides library or contact us directly to discuss your document types and volume.

Does PDF extraction automation require replacing your existing policy system?

No. The automation layer sits in front of your existing Guidewire, Duck Creek, or Majesco system, not inside it. Azure AI Document Intelligence reads the incoming PDFs, Power Automate maps the extracted fields, and the data writes to your core system via API connector. Your existing system stays in place. No migration required.