# Insurance Claims Are AI's Messiest Data Problem — Denys Linkov, Wisedocs

## Executive summary

The insurance claims industry faces significant data challenges due to its reliance on unstructured formats (handwriting, faxes, 8GB PDFs) and complex, non-standardized processes. Denys Linkov outlines a three-phase approach to leveraging AI: 1) Cleaning data (addressing duplicates and poor digitization); 2) Analyzing information within a single document; and 3) Generating cross-claim insights across an entire portfolio. Successful AI deployment requires robust digital transformation, including structured data lakes and human-in-the-loop validation, to handle the complexity of medical records and mitigate risks like fraud and regulatory non-compliance.

## Key takeaways

- Data Cleanliness is the Prerequisite for AI: Claims data is plagued by issues like duplicate records (sometimes exceeding 50%), mixed patient records (violating HIPAA), and poorly digitized inputs. Solving data cleanliness requires technical implementations to detect and manage these issues before AI can be effective.
- Three Phases of AI Application in Claims: The process must move from basic data cleaning to within-document analysis, and finally to cross-claim insights. Cross-claim analysis allows insurers to identify trends in loss ratios, geographical patterns, and provider networks across their entire portfolio.
- Benchmarking AI Performance: The MLCR (Medical Claims Record) benchmark was developed to measure AI effectiveness across three axes: conciseness, accuracy, and completeness, using a human-verified answer as the gold standard. This process is crucial for building confidence in automated systems.

## Technical details

- Data Ingestion and Processing Pipeline: The recommended pipeline involves multiple layers: OCR and image processing for underlying records, followed by NLP models and Vision-Language Models (VLMs) for extraction. A summary layer sits on top, which is then human-verified before running advanced analysis.
- Data Structure and Complexity: Claims often contain massive, unstructured files (e.g., 8GB PDFs) containing images, X-rays, and ultrasounds. Extracting context from these requires advanced techniques, as standard context windows are insufficient.
- AI Model Limitations and Safety: AI models, particularly LLMs, are prone to hallucinations. To maintain regulatory compliance and accuracy, the system must incorporate a 'human-in-the-loop' mechanism for all critical decisions.

## Practical implications

- Implementing robust data pipelines can reduce claim processing time and improve patient outcomes by automating the handling of complex, multi-source medical records.
- Cross-claim analysis allows insurance companies to identify systemic cost inflation, potential fraud, and opportunities for underwriting improvements across different lines of business.
- Focusing on digital transformation and data governance (e.g., building a clean data lake) is necessary to move beyond basic AI applications and achieve true enterprise-wide intelligence.

## Topics

AI, Insurance Technology, Data Engineering, Medical Records, Digital Transformation, Data Governance, Wisedocs, MLCR Benchmark

Source: https://www.youtube.com/watch?v=W4pnzOURwHQ