< + > Healthcare’s Paper-to-EDI Bridge Was Built on OCR, It’s Time to Replace It With Vision-Language Models

The following is a guest article by Shaurya Sahay, Co-Founder of Hewto.ai, an AI company building AI products for the US healthcare industry

Healthcare standardized on EDI for claim submission decades ago. The X12N 837 is HIPAA-mandated, and the 2025 CAQH Index estimates that nearly 98% of medical claims are submitted electronically — among the highest adoption rates of any HIPAA administrative standard. On paper, the paper-claim problem appears solved.

It isn’t.

At the scale U.S. healthcare operates — billions of claims annually — even the remaining fraction translates into millions of paper claims every year. These claims are concentrated among smaller providers, workers’ compensation and auto medical claims, secondary submissions, and workflows involving attachments.

The fragility of the system became evident when Change Healthcare went offline in February 2024, forcing providers and payers to revert almost overnight to paper, fax, and manual workarounds. The incident exposed how fragile the bridge from paper to EDI really is.

For the last two decades, that bridge has largely been built on OCR: mailroom intake, scanning, OCR extraction, manual review and correction, and finally generation of the 837. The process works — until the document is skewed, photocopied into a low-contrast black-and-white scan, or handwritten, which describes much of what actually arrives in the mailroom.

“0” gets interpreted as “O,” “i” as “l,” and a line gets captured as “1”. Every failed extraction creates a manual touchpoint, and those touchpoints are where the operational cost still resides.

Most health plans rely on outsourcing partners to manage paper claims. These partners optimize extraction for cost efficiency, often with no accuracy guarantees on non-critical fields and limited guarantees even on line-item data. As a result, human review remains deeply embedded in the workflow.

Vision-language models fundamentally change this equation.

Imagine a product powered by a custom-trained vision model that reads paper claims the way an experienced claims processor would — handling skewed scans, poor image quality, low contrast, and mixed handwriting and print, while generating a clean, structured 837 output. The workflow shifts from labor-intensive correction of OCR output to exception-only review for the small percentage of claims that genuinely require human intervention.

A modern paper-to-EDI system typically follows a five-step extraction process:

1. Ingest claim files from the mailroom or scanning systems
2. Detect the claim type — HCFA, UB, or Dental
3. Identify alignment issues and correct skew using feature detection
4. Extract each field using multiple models
5. Flag fields for manual review if validation rules fail or extracted values do not match across models

The reason custom-trained computer vision models outperform off-the-shelf OCR systems is simple: they are trained specifically for healthcare document edge cases.

For example:

• The model understands that it should capture only the handwritten or printed claim data, not the red instructional labels on the form
• It can infer low-ink or faded characters the way a human operator would
• In overlapping-text scenarios, it can determine whether the foreground or background text is the relevant value

In practice, computer vision models are especially effective for claims that are:

• Skewed or poorly aligned
• Low-contrast or low-ink scans
• Handwritten
• Photocopied multiple times
• Mixed-format documents containing both print and handwriting

The operational impact is significant. Custom-trained vision models can reduce paper claim processing times from minutes to under ten seconds per claim, while maintaining high accuracy even on poor-quality and handwritten documents.

However, the quality of the model depends heavily on the quality of the training data. Vision model training requires large, clean datasets consisting of field-level image snippets paired with accurately captured values. The more representative the training set — especially across edge cases — the better the model performs in production.

Healthcare spent the last two decades digitizing claims submission through EDI. The next phase is modernizing the paper-to-EDI bridge itself.

OCR built the first generation of that bridge. Vision-language models will build the next one.

Modern platforms such as Hewto.ai are using custom-trained vision-language models to automate paper-to-EDI conversion even on poor-quality handwritten scans.