AI's messiest test: insurance claims buried in faxes and 8GB PDFs

The core argument. Denys Linkov, who leads the AI program at Wisedocs, makes a familiar point with an unusually concrete backdrop: AI in insurance is not a model problem. Claims arrive as paper, faxes, and enormous PDFs, often before any data pipeline touches them. He cites a fax market above 7 billion dollars and EHR adoption that only reached 80-90% during COVID. The consequence is that claims processing still depends on manual extraction and human review.

The mess. A single workers' compensation claim can include intake forms from Texas, New York, and California that ask for different fields, doctor notes that contradict each other, missing X-ray reports, and duplicate records that run as high as 50% in some jurisdictions. Linkov describes multi-gigabyte PDFs, one 8GB file in production, and adjusters juggling 200 claims. Each handoff adds delay; the best claims close in weeks, the worst drag for years. That is the loss-ratio problem.

The benchmark. Wisedocs built MLCR with Artificial Analysis to test models on medical claims. Initial accuracy was around 20%; months later frontier models reached about 75% on harder reasoning tiers. The benchmark also measures conciseness and completeness separately, which matters because Linkov reports GPT models score well on accuracy but omit information, while Anthropic models are more complete with slightly lower accuracy. The lesson for teams: do not evaluate on one metric, and do not trust a vendor's single number.

The workflow. Linkov's proposed architecture is layered: OCR and preprocessing, classifiers and extraction models, summarization, human verification, then cross-claim analysis. Human-in-the-loop review is non-negotiable, both for regulatory comfort and because bad upstream data invalidates downstream insights. The more interesting claims are portfolio-level: provider networks, co-occurring diagnoses, treatment cost variation, and year-over-year price increases. None of that is possible until one claim is cleanly extracted.

The trade-offs. The vendor position is visible. Wisedocs sells exactly this pipeline, and the benchmark is self-reported. Linkov is honest about limitations: no industry-wide proof of reduced claim time, and case studies instead of hard numbers. The counterintuitive signal is that plaintiff lawyers adopted AI faster than insurers because their incentive is speed and settlement size. That asymmetry, combined with fraud and duplicate submissions, means slow digital transformation makes insurers more vulnerable, not just less efficient.