Insurance Claims Are AI's Messiest Data Problem — Denys Linkov, Wisedocs
Read the talk
Insurance Claims Are AI’s Messiest Data Problem
Denys Linkov of Wisedocs follows a medical claim from inconsistent records through extraction, evaluation and portfolio analysis—and explains why people still need to verify the result.
From a talk by Denys Linkov
At a glance
Ideas worth remembering
Duplicate, mixed and missing records can change a claim’s outcome; extracting text is only the beginning of cleaning the file.
MLCR evaluates accuracy, completeness and conciseness separately. A correct answer can still omit information a reviewer needs.
Cross-claim analysis can expose treatment, cost and professional relationships, but its usefulness depends on reliable extraction from individual claims.
Wisedocs layers preprocessing, specialized models, human-verified summaries and further analysis; people remain responsible for decisions.
AI arrives before digitization is finished
A request to find a doctor with top-half outcomes for a patient’s ankle injury sounds like a database query. It first requires reliable data about the injury, doctors, outcomes and a recent time period. Denys Linkov, who leads the AI program at Wisedocs, uses that example to show how far an insurer may be from asking a seemingly straightforward question.
The problem starts before AI. Paper records and fax still move between medical offices, legal departments and insurers, while electronic records do not all capture the same information. Linkov shows workers’ compensation intake forms from Texas, New York and California. What changes from one form to another? Each jurisdiction has its own fields and layout; even a predictable form can require different information from another state’s form, and versions can change.
Paper also makes comparison across claims difficult: organizing one stack does not make its contents easy to aggregate with thousands of others. Linkov’s broader point is that cloud migration, data infrastructure, governance and security remain unfinished in many enterprises. A useful AI application needs a defined outcome and the people and systems capable of supporting it.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A claim accumulates handoffs and contradictions
Linkov focuses on complex medical claims in property and casualty insurance: a person might be injured at work, in a car accident or by a defective product. He says adjudication can take weeks in the best case and years in the worst, depending on complexity, regulation and litigation. Treatment may already have been delivered while the claim continues.
A worker reports an injury and files a claim, then may visit an emergency room and other providers. Diagnosis, specialist opinions, coverage checks, utilization review, treatment, billing and payment introduce handoffs and possible branches. Specialists can disagree; procedures and providers may be in or out of network. Linkov’s illustrative claims are synthetic, but he describes them as close to production cases.
Each handoff can add a differently formatted bill or progress note, often delivered as a PDF, email or fax. One assessment might even contain three different results for a patient’s range of motion. Resolving the claim requires deciding what those differences mean, not merely converting the pages to text.
Litigation adds another set of participants and incentives. Linkov says plaintiff-side firms have incentives to move cases quickly and pursue larger settlements, while insurers must also deal with repeated or potentially fraudulent submissions. He returns to the claimant’s experience: the goals include good care, lower costs and a quicker process, even when the administrative system has many competing concerns.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Clean the record before asking it questions
Linkov organizes the work into three levels: clean the data, analyze an individual claim, then look across claims. The first level has several distinct failure modes.
- Duplicates: Repeated follow-ups can send the same material through the system again. Linkov says duplicates can reach 50% or more in some jurisdictions; that is a context-dependent observation, not a rate for all claims.
- Mixed identities: Records for people with similar names, such as a parent and child, can be commingled. A decision made on another person’s medical history can damage the claim and create privacy risk.
- Unreadable or inconsistent source data: Poor scans, handwriting, copy-paste between visits or patients, and differently configured electronic medical record systems make extraction harder. Even electronic data may be sent onward as a fax or PDF when it is difficult to retrieve.
What does duplication look like at the receiving end? Linkov presents illustrative duplicate records to show why a recipient needs detection systems even when the repeated documents originate elsewhere. Detecting repeats can prevent the same material from being processed over and over, including when a claim is resent with different dates.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Find what is missing as well as what conflicts
Size is its own technical problem. Linkov recalls finding an 8 GB PDF in Wisedocs’ database. A claim can contain accident images, X-ray reports, ultrasounds and other records; a file that large can challenge a PDF viewer as well as a model’s context window. It cannot be treated like a ten-page document casually pasted into a chat.
A progress note might say an X-ray was booked, while the X-ray itself is absent from the claimant package. Flagging that gap prompts a follow-up: was the record omitted, did a fax fail, or is there another explanation? The answer can affect assessment of care and its relationship to billing. An adjuster handling as many as 200 claims at once cannot reliably keep every such follow-up in memory.
Conflicting diagnoses call for a different response. An insurer may seek an independent medical evaluation, in which another doctor reviews the file, including when it wants to check the amount to set aside for a claim. That adds expertise, but also another provider and another handoff to manage. A system can highlight the conflict; the medical judgment still needs a reviewer.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Measure accuracy, completeness and conciseness separately
Wisedocs and Artificial Analysis built MLCR to evaluate questions about medical claims, including the kinds of missing records and conflicting diagnoses that complicate review. Its six tiers range from identifying an injury and its date to questions such as how many surgeries occurred or what capacity a patient has. A question can itself assume a fact that the record does not contain, creating an opportunity for hallucination.
The first three, easier tiers have an open-source dataset; a private set is hosted with Artificial Analysis. Responses are assessed for accuracy, completeness and conciseness against human expert-validated answers, with calibrated judges and triple-judge verification. These dimensions matter separately: an answer can be correct in what it says while leaving out information the reviewer needs.
How different are accuracy and completeness in practice? In the benchmark results Linkov discusses, GPT models were strong on accuracy but omitted key information, while Anthropic’s models were more complete with slightly lower accuracy. He says completeness remained difficult across most providers. The comparison makes the choice of metric consequential: selecting only for correct statements can hide missing parts of an answer.
Linkov reports that the highest accuracy on the benchmark’s reasoning levels rose from 20% at release to 75% a few months later with very high reasoning levels. Those are his reported benchmark results, not a general clinical accuracy rate. Cost per suitable answer also matters when choosing a model, and Wisedocs retains human review. Calibrating each step matters because an incorrect extraction can contaminate everything built on it.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Turn individual records into portfolio questions
Once information is extracted from individual claims, it can be combined across a portfolio. Linkov’s illustrative mid-sized insurer collects $1 billion in premiums and may pay 20,000 claims a year or more. It can ask which injuries, geographies, doctors and claim types affect its loss ratio—the relationship between claims paid and premiums collected.
A redacted California subset offers a relationship question: which lawyers, doctors, chiropractors and physiotherapists work together across claims? The connections can help an analyst investigate the story behind a claim. A separate geographic question asks how far patients travel for treatment and whether they receive it locally. These are starting points for investigation, not proof that a particular relationship or journey is improper.
The same extracted data can group co-occurring conditions. For a back injury, several diagnoses may map to ICD codes; comparing their occurrence across doctors’ write-ups can raise questions about why the records differ. This portfolio view depends on the earlier extraction step: errors within one claim can become misleading patterns when aggregated.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Compare treatment costs with outcomes
A comparison of physiotherapy charges asks a more subtle question than which doctor charges least. Linkov suggests that a provider who helps someone recover in five treatments rather than ten could justify a higher price per treatment. In his example, one doctor’s charge for a type of treatment rose 10% year over year. The comparison invites the next question: is that change specific to the provider or occurring across the jurisdiction?
With sufficiently reliable data, an insurer can explore counterfactuals: what might change if it adjusted a provider network, directed patients differently or changed billing? It can also compare trends across lines such as auto injury, workers’ compensation and general liability, and use the information to examine underwriting assumptions. These are potential decisions enabled by analysis; a detected pattern alone does not establish the outcome of a policy change.
The desired result is not just a cleaner data lake. Linkov ties it to shorter claim duration, faster return to work, treatment and patient experience. He reports significant savings in Wisedocs’ work, while later limiting the claim: the company cannot yet say it has reduced average claim time across the industry.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Build the analysis in verified layers
In the Q&A, Linkov explains how Wisedocs structures the records. PDF preprocessing uses OCR and images to extract underlying information. NLP, extraction models, vision-language models and classifiers then help structure it; a conventional entity-extraction model can be suitable for some fields. A summary layer sits above those models and is verified by people before additional analysis.
What reaches the later analysis, and where does verification occur? The layers progressively concentrate information, with models and people checking it at various points. MLCR asks questions using clean data; real claim files need the preceding cleanup and extraction work before that benchmark-like task is possible. Access to portfolio insights also depends on how each customer permits information to be shared across teams and lines of business.
From claim files to reviewed analysis
Read the diagram as text
- Claim records. PDFs containing text and images.
- Preprocessing. OCR and image processing extract underlying information.
- Structuring models. Extraction, NLP, vision-language models and classifiers.
- Summary. Document summaries are verified by people.
- Additional analysis. Higher-level questions use the prepared information.
- Claim records → Preprocessing: process records.
- Preprocessing → Structuring models: text and images.
- Structuring models → Summary: structured information.
- Summary → Additional analysis: reviewed summary.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Align the fields, keep decisions reviewable
Onboarding starts with sample files and a comparison between what the customer extracted and what Wisedocs extracted. Even a date field needs agreement: does the business want the injury date or the date the report was submitted? If standard cases align, the work can move to a broader pilot and integrations with upstream providers and internal systems. Linkov notes that some capabilities are available immediately and others are not.
Wisedocs’ described system requires a human to make the decision. For example, it can point to differing diagnoses on page 2 and page 20 for a reviewer to inspect. Linkov favors an auditable pipeline and explains its data and operation to customers’ governance and regulatory teams. He considers human review helpful for regulatory concerns; that judgment is not a substitute for assessing the rules in a particular jurisdiction.
Model choice is also a cost and accuracy decision. Wisedocs has trained its own models and uses classification and extraction alongside more complex model-based analysis. Linkov’s point is practical: a task does not necessarily need the most expensive reasoning approach merely because it appears in an AI pipeline.
The last complication is time. Linkov describes jurisdiction-dependent response deadlines and the work needed to reconstruct a long medical history: a worker may have changed names, cities or providers over decades. An initial notice is only the beginning of that search. He reports customer-specific improvements in speed and cost, but says Wisedocs cannot yet claim an industry-wide reduction in average claim time. Making the first document easier to process does not make the rest of the history disappear.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
Related talks
- Your Agreements Are a Database You Can't Query — Hiral Shah, Docusign & Sean Sodha, NVIDIA
Examines another document-extraction problem where preserving structure is essential to querying the result.
- Enterprise Agents Have a Structure Problem - Ishita Daga, Tesla
Explores how inconsistent business definitions and missing context limit enterprise data agents.
Full transcript
Loading transcript…