Most corporate AI projects never make it out of the lab. Studies show 60%-70% of proof-of-concept (PoC) projects are abandoned before reaching actual services, according to Kim Hee-jun, chief technology officer of Kakao Pay Insurance. Speaking with ChosunBiz at the company's Pangyo office, Kim said the problem is that companies run projects with the goal of adopting AI, rather than solving specific pain points with it.
Kakao Pay Insurance's overseas medical expense AI claims review service is built on that principle. The company used AWS's Kiro AI development tool to cut claims review time from 40 minutes to more than two hours per case down to 2-3 minutes. "What customers want most is speedy insurance payouts," Kim said. "To achieve that, we had to solve the issue of the long time and resources required for claims review."
Why overseas claims are slow
Domestic medical expense claims in South Korea follow standardized formats for receipts and itemized statements, and diagnoses are organized by codes. Overseas documents differ in everything from institution names and medical certificate terminology to the types of prescribed drugs. Most are also written in foreign languages. Reviewers had to verify one by one whether a hospital actually exists and whether prescribed drugs were legitimately prescribed.
Kakao Pay Insurance sought to improve this work through AI. A claims expert still makes the final decision, but the company focused on drastically reducing the effort required to reach that decision.
Building for adoption, not demonstration
Kim said many AI PoC projects fail because they are approached as technology tasks detached from real work. "From the perspective of front-line departments, they think, 'Here we go again,' rather than believing it will help their work," he said.
In this project, the team promised it would be "a weapon to reduce your overtime and solve real pain points," and the on-site response was, "Then can we also give lots of input?" The most important part was not "let's adopt AI," but an approach of "let's use technology to solve what is hardest for us right now."
Kakao Pay Insurance runs PoC projects with three principles: clarify the problem to be solved, specify the project's constraints, and use currently available technologies. This approach matters for AI for Insurance teams, where the gap between pilot and production is notoriously wide.
Accuracy and the role of human review
Kim addressed concerns about AI hallucinations in financial services. "AI hallucinations mostly pop up during reasoning, but financial firms' AI uses retrieval-augmented generation (RAG), which derives results based on facts rather than reasoning," he said. "Because it keeps substantiating the grounds for its answers, the accuracy is quite high. Of course there are false positives, but people verify and reference them, so it is different from leaving everything to AI."
The company chose AWS Kiro as its primary AI development environment. Kakao Pay Insurance uses Kiro not only for developers' coding, but also for planners' and front-line departments' work, including eliciting and analyzing requirements.
Expansion plans and other AI services
Kakao Pay Insurance is considering applying for the financial regulatory sandbox program within the year to bring the overseas claims service to customers. Because the system was developed in-house, it can be extended to other claims types with similar pain points.
The company has already deployed other AI services. Its in-house document recognition model identifies customer-submitted documents and extracts key data. It has been applied to mobile phone insurance claims review, with one payout made about 18 seconds after filing. Flight delay compensation for overseas travel insurance has seen a payout in one second.
Kim said it is still early to view AI from an expense-risk perspective. "The right expression is not 'AI replaced people and reduced costs,' but 'thanks to AI, people are doing more valuable work,'" he said. The broader shift toward AI Agents & Automation in claims processing is about changing the customer experience, not just cutting headcount.
Why this matters for insurance professionals
Kim's central argument is that model choice matters less than problem selection. "Big Tech AI firms are rushing to claim 'our model is the most powerful,' but using that model does not make our service the strongest," he said. "It's not the model you used, but where you apply AI to draw out customer value."
For insurance teams, the practical takeaway is to start with the claims process that hurts most - whether that's overseas reviews, document processing, or payout speed - and build AI around that specific friction point. Kim also stressed the importance of trustworthy AI in financial services: RAG to reduce hallucinations, guardrails for stability, and caching to balance expense and response speed. "We live in an era when today's most powerful model becomes outdated in a few months," he said. "Corporations should focus on their own data and domain-specialized technology."
Your membership also unlocks:
