The 80% of Insurance Documents Nobody Is Reading — And What Upstage Is Doing About It

Most insurance companies have only ever been able to extract structured data from about 10 to 20 percent of the documents they handle. The rest — the messy tables, the multi-column forms, the handwritten submissions, the merged-cell invoices — either get processed manually by third-party vendors or sit untouched in a cabinet somewhere.

That gap is the business Upstage was built to close.

Kasey Roh leads Upstage's US operations. She came to the role through Tesla, Meta, angel investing, and venture capital — a path that gave her an unusually clear-eyed view of which AI companies are building something real and which ones are wrapping an existing model and calling it proprietary. In Episode 152 of InsurTechTalk, she walked me through why traditional OCR fails, why throwing documents into frontier LLMs is not a scalable solution, and what it actually means to build an AI-native document intelligence platform from the ground up.

About Kasey Roh

Kasey Roh heads US operations at Upstage, an enterprise AI company founded in Korea that has been serving Fortune 500 companies across Asia and the US for five years. Before joining Upstage, she worked in corporate finance at Tesla during the Model 3 hypergrowth period, moved to capital strategy at Meta, built an angel investment syndicate within the Facebook alumni community, and joined Prime Mates — described as the YC of Korea — to lead their US fund. Upstage has raised approximately $150 million to date, is backed by SoftBank Ventures, and is growing at triple-digit rates year over year, with IPO conversations underway.

What Upstage Does

Upstage builds enterprise AI infrastructure for document-heavy industries, with insurance as a primary focus. The company serves nearly seven of the top ten insurance carriers in Korea and is expanding rapidly in the US market.

The Two Core Models

OCR model: A next-generation optical character recognition engine enhanced with machine learning for layout recognition and context understanding — reading multi-column documents, tables with merged rows and cells, charts, captions, and forms in the way a human would, not just left to right
Solar LLM: A large language model built from scratch — not fine-tuned from an open-source base — specifically optimized for document understanding, designed to be faster, cheaper, and more accurate than frontier models for document processing tasks

The core product is the marriage of these two: OCR handles extraction with precision, Solar LLM adds business logic, classification, summarization, and structured output. The result is intelligent document processing that does not hallucinate on extraction because it is not guessing — it is reading.

Why Traditional OCR Fails — and Why LLMs Alone Are Not the Answer

This was the sharpest technical insight in the conversation, and it has direct implications for any insurance company currently evaluating AI document solutions.

The OCR Problem

Traditional OCR reads documents left to right without understanding context, layout, or structure
Insurance documents are rarely simple: court forms with checkboxes, SOVs with multiple rows, supplemental applications with no standardized format, claims invoices with merged columns and nested tables
Traditional OCR handles roughly 10 to 20 percent of the document types an insurance company actually deals with
The remaining 80 to 90 percent is either processed manually by third-party vendors or simply never touched

The Frontier LLM Problem

Throwing documents as images into GPT-4, Gemini, or similar frontier models works — for occasional use
At enterprise scale — 50,000 submissions per month, hundreds of thousands of pages per day — the cost of frontier model inference is comparable to or worse than outsourcing to manual processors
Frontier models hallucinate on document extraction, and one hallucination in five attempts is enough to permanently destroy an underwriter's trust in the system
There is no fix available when the model you are using is not yours — you cannot fine-tune what you do not own

The Upstage Approach

Downsize the LLM specifically for document tasks — do not try to solve the entire AI universe, just solve document reading
Use OCR for precise extraction, LLM for context and business logic
Keep the model proprietary so that when errors occur, they can be diagnosed, addressed, and used to improve the core algorithm
Deploy in private cloud environments where data never leaves the enterprise — critical for insurance, where submissions and claims contain trade-sensitive and personally identifiable information

The 80% Opportunity: What Unlocking Untouched Documents Actually Means

Kasey shared a case study from one of the largest carriers in Asia that illustrates what happens when an insurance company finally gets to that 80 percent.

What One Carrier Found in Ten Years of Untouched Documents

The carrier had ten years of claims documents — receipts, prescriptions, attachments — sitting digitized but unreadable in their database
Upstage used idle GPU capacity at night to process the entire archive while daytime capacity ran live claims
After digitizing a decade of records, the carrier discovered data patterns at a granularity they had never seen before — including that patients treated at certain hospitals by certain doctors had a statistically measurable higher incidence of specific cancer diagnoses
Using this insight, the carrier launched two new supplemental health plans — one targeted at breast cancer for women, one at colon cancer for men — based on early detection patterns identified from the historical data
The product line had not existed before. It was created entirely from data the carrier already owned but had never been able to read

Hallucination, Trust, and Why Governance Is Not Optional

One of the most practical points in this conversation was about what happens to enterprise AI adoption the first time a system gets it wrong.

Why One Hallucination Breaks Everything

An underwriter who sees a severely hallucinated result — even once, after five perfect runs — stops using the system immediately
The trust relationship between a knowledge worker and an AI tool is asymmetric: it takes many correct outputs to build confidence and one wrong output to destroy it
The reason the trust breaks is not just the error — it is that the user has no one to call. There is no accountability path. The agent made a decision, it was wrong, and there is no clear owner of that failure
This is why AI in insurance cannot be positioned as replacing human judgment. The agent is not a legal entity. It cannot be held responsible for a bad underwriting decision or a wrongly denied claim. The human must remain the decision maker, with AI as the tool that gets them there faster and more accurately

What Good AI Governance Looks Like in Practice

Human touch points built into every mission-critical step of the workflow
Feedback loops that capture errors and use them to improve the model — not just one-shot automation
Audit trails that allow someone to follow the reasoning of the machine and identify where it went wrong
Private deployment so that confidential data never touches a public model or gets used for training without consent

Is Your AI Investment Compounding?

Kasey offered one of the clearest frameworks for enterprise AI procurement I have heard: treat it like a financial investment. If each successive deployment is not getting easier — if the second rollout does not benefit from the first — you are doing something wrong.

How to Evaluate an AI Vendor

Start with the use case: Know which specific business problem you are solving before evaluating any technology
Ask if the investment compounds: Does each deployment build on the last? Does your data foundation get stronger over time?
Assess proprietary depth: Does the vendor own their model and algorithm, or are they wrapping a third-party API? If they do not own the underlying technology, they cannot fix it when it fails
Prioritize data foundation first: Structured, accessible, high-quality data is the prerequisite for every other AI application — predictive analytics, fraud detection, new product development, all of it flows from this

Key Takeaways

Traditional OCR handles only 10 to 20 percent of insurance documents; the remaining 80 to 90 percent is a largely untapped data asset
Frontier LLMs are not scalable for high-volume document processing — cost and hallucination make them unsuitable as the primary extraction layer
The right approach combines precision OCR with a purpose-built, proprietary LLM — cheaper, faster, and more reliable than frontier models for document tasks
One hallucination destroys enterprise trust; governance, feedback loops, and human accountability are not optional features
The biggest opportunity in insurance AI is not automation — it is unlocking decade-old document archives that contain product, risk, and customer insights nobody has ever seen
Evaluate AI vendors by whether they own their technology — if they cannot fix the model when