OCR + LLM: The Architecture That Actually Scales AI in Insurance | Kasey Roh Upstage AI

Episode 152 of InsurTechTalk Kasey Roh--Head of US, Upstage AI: Building the “Perfect Marriage” Between OCR and LLMs.What happens when AI stops trying to boil the ocean—and instead solves one of the most painful, expensive problems in enterprise workflows?In this episode, I sit down with Kasey Roh, Head of US at Upstage AI, to unpack how a focused, infrastructure-first approach to AI is quietly transforming industries like insurance, healthcare, and manufacturing.We start somewhere unexpected—yoga, discipline, and managing chaos in the age of AI—and quickly move into one of the most important conversations in enterprise tech today: why most AI implementations fail to scale, and what it actually takes to build systems that compound.Upstage isn’t just another “AI wrapper.” They’ve built both their own OCR engine and a proprietary LLM (SOLAR), designed specifically for document understanding—arguably one of the most overlooked bottlenecks in enterprise operations.Key topics we cover:1. Why traditional OCR only solves ~10–20% of enterprise document workflows2. The hidden cost problem of LLMs at scale (tokens, GPUs, inference)3. Why hallucinations kill enterprise trust—and how to design around them4. The “OCR + LLM” architecture as a pragmatic middle ground5. Turning 10 years of unused documents into entirely new insurance products6. Why AI in enterprises is not about automation—but accountability7. The real test for AI adoption: does your investment *compound*?We also go deep on:1. Evaluating AI vendors in a world full of “AI-washing”2. The importance of owning core models vs. relying on wrappers3. Human-in-the-loop systems as a feature—not a limitation4. The future of agentic workflows and governanceThis is a must-watch for operators, investors, and anyone trying to separate signal from noise in the AI hype cycle.Timestamps:00:00 – Yoga, chaos, and entering the AI world04:30 – What Upstage actually does (OCR + proprietary LLM)12:00 – Why enterprise OCR is fundamentally broken20:10 – The cost vs. accuracy tradeoff in LLMs28:45 – Unlocking value from “dead” enterprise data38:20 – AI vs. humans: accountability and trust47:10 – Evaluating AI vendors (what actually matters)55:00 – The compounding framework for AI investmentsAbout Kasey Roh:Formerly at Tesla and Meta, Kasey transitioned from corporate finance and venture investing into operating, now leading Upstage’s US expansion. Upstage has raised ~$150M and is scaling rapidly across global enterprise markets.