By Christopher Frankland | InsurTech360 & BionicAgent.com
When Kasey Roh talks about Upstage, there’s an unmistakable sense of purpose – a belief that enterprise AI should move beyond hype into measurable impact. From its early days in Seoul to its growing presence across the U.S., Upstage has carved out a distinct path: building domain-specific language models and intelligent tools that help organizations transform the way they work with documents, data, and decisions.
Phase One: Precision, Not Promises
Founded in October 2020, Upstage began its journey with a focus on a single, unglamorous but foundational challenge – optical character recognition (OCR). As Roh recalls:
“It took us about a year to launch our first OCR product. Our first customer was one of the largest insurance carriers in Korea – with nearly $90 billion under management. Within six months, we realized our accuracy rates were 20-30 percent higher than any competing vendor.”
For an industry where margins are thin and volume is high, that difference was seismic. Roh points out that a few percentage points in accuracy can translate into millions of dollars saved annually for large carriers processing millions of claims and policy documents.
Word spread quickly. Within a year, Upstage’s document-intelligence solutions were being used by Korea’s top five life and property insurers – a testament to the power of enterprise-grade precision and speed.
Phase Two: The LLM Revolution
Then came the turning point. Late 2022 saw the explosion of generative AI and large language models. For Upstage, this wasn’t a disruption – it was an inflection point.
“With generative AI, we were able to overcome the limitations of traditional OCR. You no longer need rigid templates or rules. As long as you know the data fields you want to extract, the LLM can dynamically interpret and extract them – regardless of layout or format.”
That leap enabled Upstage to create an LLM-based extraction engine that removed one of the biggest barriers in enterprise AI: dependency on manual template creation. Where teams once spent months building schema maps and IT pipelines for each document type, Upstage’s LLM can now adapt in real time – reading, understanding, and structuring data across hundreds of formats.
Roh explains that this system even incorporates a confidence score – automatically flagging uncertain extractions for human review. Those corrections are then fed back into the model through an automated feedback loop, continuously improving accuracy over time.
“We built a feedback loop that automatically updates the underlying schema. The system learns from human validation, making every correction part of its continuous improvement process.”
Phase Three: Building for Scale – and for Humans
Despite Upstage’s rapid evolution toward end-to-end automation, Roh is clear-eyed about the balance between AI and human judgment:
“We can build the perfect infrastructure, but technology will never replace business context. Many errors come not from AI limitations but from unclear business intent. Technology needs to work in partnership with humans, not replace them.”
This philosophy – that AI should augment rather than replace – sits at the core of Upstage’s design ethos. The company is now developing systems where business users can inject their own rules, logic, and guidelines directly into the AI engine.
“We’re making it easy for business users to embed their logic into our platform. That’s how we go beyond extraction – by allowing business and technology to collaborate in real time.”
Phase Four: Redefining the Enterprise Workforce
When asked how this technology is changing teams inside carriers, Roh pauses – not because she needs time to think, but because the answer is transformational.
“The biggest change we’re seeing is within data teams. Traditionally, data scientists spent time building schema and annotation models for each OCR project. Now they’re using our engine to monitor confidence scores and focus on higher-level data analysis and insight generation.”
In other words, Upstage is helping data scientists get back to being data scientists. Rather than manually tagging fields, they’re refining models, analyzing performance, and unlocking insights that drive strategic value – a shift that embodies what many are calling the “bionic workforce.”
Phase Five: Global Expansion and the U.S. Play
Having proven its model in Asia, Upstage set its sights on the U.S. In 2025, the company closed a Series B bridge round that included AWS and AMD Ventures as strategic investors, signaling a major push into North America.
“Our mission is to bring that same level of document intelligence to the U.S. market – especially industries like insurance and financial services that still rely heavily on manual document processing. We’re ready to scale.”
At ITC Vegas, Upstage’s presence was impossible to miss. Beyond its booth buzz and live demos, the company stood as a case study for how AI can move from research to revenue – from technical promise to enterprise production.
The Road Ahead: Toward Agentic AI in Enterprise Workflows
When asked about what’s next, Roh doesn’t default to buzzwords. Instead, she talks about responsibility, transparency, and collaboration.
“Confidence scores will keep rising. Automation will expand. But we can’t forget that technology is only as good as the instructions we give it. Our goal is to create systems where AI and humans co-author decisions – with traceability and trust built in.”
That vision aligns closely with the future many of us in the insurance and enterprise AI space are building toward – one where agents are not replaced but reinforced by intelligent systems.
Conclusion: Intelligence in Action
From OCR to LLM to agentic AI, Upstage’s journey is a blueprint for what enterprise AI can be when precision, purpose, and partnership align. Under Kasey Roh’s leadership, the company is not just digitizing documents – it’s redefining the relationship between humans and machines in the enterprise workflow.
For industries like insurance that still wrestle with manual documents and fragmented data, the message is clear: the future of work is intelligent, collaborative, and already here – and Upstage is helping build it.
Kasey Roh (Head of US):
Kasey Roh is the U.S. CEO of Upstage AI, based in the San Francisco Bay Area. Before joining Upstage, she was an early-stage venture investor focused on AI startups, including Upstage itself.
She previously held strategic roles at Meta and Tesla during their hyper-growth phases, where she managed over $10 billion in capital across Capital Strategy and Corporate Finance. Kasey brings a unique blend of operating and investing experience at the intersection of AI, innovation, and cross-border market expansion.
She is also an active angel investor and co-leads Iterative Ventures, a syndicate she founded with fellow Meta alumni.
Upstage:
Upstage is building the future of enterprise AI through domain-specific language models and intelligent tools designed for high-impact business use. With deep expertise in LLM development, document AI, and agentic workflows, the company enables regulated industries and global enterprises to unlock the full potential of generative AI – securely, efficiently, and at scale.
