How to Clean Insurance Claims Data Before Adding AI

Insurance claims data is often described as unstructured, and the machine-learning fixes people reach for are the wrong place to start. The harder truth from Wisedocs AI lead Denys Linkov is that a claim is large, inconsistent and sometimes shared with the wrong patient, so document AI only works after extraction and deduplication are auditable.

Why insurance claims data is the real blocker for AI

Insurance claims data is the hardest input in most enterprise AI programs because a claim arrives as a stack of inconsistent documents before it becomes a record. Denys Linkov, who leads the AI program at Wisedocs, a medical claim processing company, placed the problem upstream of model choice in an AI Engineer talk published on October 18, 2026.

The claim runs through a document pipeline that rarely ends in one clean file. A hospital formats its own bill, a doctor writes a progress note in a personal style, and the intake forms differ by state. Linkov showed three workers compensation intake forms from Texas, New York and California. Each asked for different fields.

That inconsistency survives into the systems built to fix it. According to the Centers for Disease Control and Prevention, only about 9.4 percent of non-federal acute care hospitals reported at least a basic electronic health record system in 2008. Adoption reached the high 80s and low 90s only around the COVID-19 pandemic, so paper workflows still sit upstream of many claims.

The business stakes are visible in one ratio. Linkov described a loss ratio near 50 percent, the share of collected premiums paid out on claims, as a typical reference point for the sector. Improving that number is what motivates the data work.

What the Wisedocs MLCR benchmark actually measures

The Wisedocs MLCR benchmark measures whether a language model can answer medical claim questions correctly, completely and concisely. Wisedocs built it with Artificial Analysis, an independent model evaluation service, and released six tiers of questions that rise from basic facts to reasoning over a whole claim.

The benchmark released a public dataset covering the first three, easier tiers, while the harder tiers stayed private and hosted with Artificial Analysis. Linkov gave a practical reason: a private set limits overfitting to a known test. The transcript does not state a release date beyond the October 2026 talk.

Scores are graded on three axes against a human-verified answer, with a calibrated set of judges and a triple-judge verification step. The Wisedocs blog post and a follow-up post describe the calibration method. Linkov warned that an uncalibrated judge quietly propagates upstream errors into downstream analysis.

One result shows how fast the target moves. Linkov said the highest accuracy Wisedocs saw on its harder reasoning tiers at release was about 20 percent, and a few months later reached roughly 75 percent at very high reasoning levels. Those figures compare different model configurations at different dates, so they describe benchmark movement rather than a fixed capability of any single model.

Cleaning insurance claims data starts with duplicates and co-mingled records

Cleaning insurance claims data begins with duplicate detection and patient matching, because both problems distort every analysis that follows. Linkov said duplicate records can reach 50 percent or more of a body of claims in some jurisdictions, and some duplicates are deliberate because claimants re-send files to chase a slow response.

Co-mingled records carry a higher risk. A parent and child with similar names can share a file, and decisions get made with another patient's history inside the record. Linkov tied that to HIPAA exposure, the US health privacy rule, and to claims that collapse once a wrongly attached document is discovered.

The rest of the cleanup list is familiar to anyone who has processed documents: poor scans, handwriting, copy-paste between visits and patients, and electronic medical record systems that every hospital configures differently. Linkov described hospitals in Toronto, Canada, running fresh implementations each time an EMR is installed.

Here is the staged structure Linkov described for handling the input, from raw file to verified summary.

How a claim document pipeline should be layered

The document pipeline Linkov described layers cheaper extraction first and reserves AI judgment for the top, where the remaining questions are genuinely hard. Each layer is checked before it feeds the next, and a human verifies the summary before any downstream analysis runs.

Steps in the pipeline as described:

Cross-claim analysis is where the portfolio value sits

Cross-claim analysis turns cleaned single-claim data into portfolio decisions about cost, network and treatment patterns. Linkov described a mid-size insurer collecting about 1 billion dollars in premiums across several lines of business and paying out roughly 20,000 claims a year. At that volume, per-claim accuracy alone does not explain where money leaks.

The questions that become answerable are concrete. Are patients traveling outside their local network for treatment, and is the treating provider the right one? Which lawyers, doctors, chiropractors and physiotherapists work together repeatedly? Which diagnoses co-occur with a given injury, and how do providers document them differently?

Cost analysis adds a second layer. Linkov gave an example of a 10 percent year-over-year increase for a physiotherapy treatment at a given doctor and asked whether that pattern holds across the jurisdiction. That comparison is only possible when the same fields are extracted consistently from every claim.

He also pointed to reporting on roughly 1 billion dollars of additional billings as an example of what weak processing allows. The claim is his framing in the talk, not a figure the benchmark measured.

What insurance claims data work means for regulated deployment

Regulated deployment of AI in claims rests on human review and auditability, and neither one is the same as compliance certification. Linkov said Wisedocs always requires a human in the loop to make the decision, and it surfaces conflicts as pointers, for example a notice that page two and page 20 show different diagnoses from two doctors.

He estimated that a human in the loop removes about 70 percent of the regulatory concerns customers raise. That is his working estimate from customer conversations, not a measured compliance result. An application still needs its own audit trail, access controls and jurisdiction-specific review.

Linkov named decision timelines as a real constraint. In Texas, he said a claim can require an answer within hours, while other states set 7 or 21 days. A 60-year-old worker with 30 years of medical history may need records reconstructed across name changes and moves, which makes complete retrieval matter as much as fast inference.

Limits of what one benchmark proves

One benchmark proves what its own configuration measured, and the MLCR results do not establish that a model handles every real claim. Linkov said the benchmark uses clean data to ask its questions, while production claims arrive dirty. Everything in the earlier cleaning stages exists to close that gap.

Model comparisons need the same restraint. Linkov observed during the document analysis runs ChatGPT models, OpenAI's model family, were strong on accuracy but left out key pieces of information, while Anthropic models were more complete with somewhat lower accuracy. Those are impressions from one benchmark setup, not a general ranking.

Adoption claims also stay narrow. Linkov said plaintiff-side firms have taken to AI faster than insurers because their compensation rises with speed and settlement size. He did not offer adoption data, and the talk does not support a claim about the wider industry.

FAQ

• What is the MLCR benchmark? The MLCR benchmark is a medical claim question set that Wisedocs built with Artificial Analysis to score model answers on accuracy, completeness and conciseness against human-verified answers. Six tiers exist, and the first three were released publicly while the harder tiers stayed private.

• Why is an 8 GB PDF a problem for AI? A single claim file can include images, X-ray reports and scans, so it exceeds what any context window holds at once. Linkov described seeing a PDF of that size in the Wisedocs database, which forces a page-level extraction pipeline before a model can answer questions about the claim.

• Can duplicate claim records be fixed with software alone? Detection software can flag repeats and matching errors, but Linkov noted the behavior has incentives, since claimants re-send files to speed a slow process. The fix combines detection with process changes that remove the reason to resend.

• Does a human in the loop make an AI claims system compliant? No. Human review and an auditable pipeline address many practical concerns, but compliance also depends on the organization's controls, jurisdiction and record-handling rules.

• Does the Wisedocs pipeline use large language models for everything? No. Linkov said the pipeline uses OCR, entity extraction models and classifiers, with vision-language models and summarization reserved for later stages, because cost, accuracy and hallucination risk differ by task.

Source video