AI reads, a person decides: redesigning claims review
카카오 AI 돛 Summit 26 | [카카오페이손해보험] AI가 읽고, 사람이 판단해요 - 보상 심사 재설계 여정
B · Useful, limited Kakao Pay Insurance's claims engineering lead gives a clear first-hand account of an on-premises document pipeline with humans kept in the decision; credible design detail, but no volume, accuracy or cost figures.
Summary based on the full transcript·Original: about 24 minutes to watch
In short
A 24-minute talk at Kakao's AI Summit 26 (published 21 October 2026) by Bae Seong-yul ("Lucas"), who leads the claims server development team at Kakao Pay Insurance. Having already cut customer effort on claims, the team turned to the adjusters who review overseas travel medical claims: phone photos of foreign, multilingual hospital documents. Open-weight models served on-premises now classify, parse and cross-check the documents, and parallel agents produce a review report. Adjusters keep interpretation, exceptions and the final decision.
Key numbers
- 20-40 minadjuster time per overseas medical claim before the redesign
- within minutestime to produce a review report with the new pipeline (no exact figure given)
- hundreds of pageslength of work manuals adjusters searched with Ctrl+F before conversion to text
As reported by the source.
Korean·Video·opens youtube.com
What it reports
- Before: adjusters read, translated and checked each document by hand, searched the web to verify hospitals, matched prescriptions to diagnoses and checked every receipt line. One case took 20, 30 or 40 minutes, and an added document restarted the review.
- Pipeline: intake to a bucket; pre-processing (orientation, quality, PDF splitting); document classification (diagnosis, receipt, prescription) and parsing to JSON; prior claim history; role-split agents running in parallel; a final review report.
- Models are open-weight and served on internal infrastructure, so no external API is called, which the speaker says matters under financial regulation.
- Result: work that took tens of minutes now completes "within minutes, with more accurate output". Screens were not shown because of personal data.
- Principle: human-in-the-loop is a design element. AI reads, finds and transcribes; people interpret context, handle exceptions and carry responsibility because money moves.
- Knowledge step: manuals, review guidelines and policy terms existed as slide decks exported to PDF. The team converted them to text as a single source, then built chunking, Amazon Titan embeddings, hybrid keyword and semantic search, and per-chunk hashes for incremental updates. Graph RAG and ontology are next.
- Testing step: Playwright browser tests generated from natural-language scenarios, so business staff as well as QA write test cases.
- All team members use coding agents; an internal "AI gateway" standard governs cross-team model access.
How they got it
Conference talk with architecture slides and a demo video; no metrics beyond the before-state time per case. The transcript is from Korean auto-generated captions.
Limits
- No figures for claim volume, accuracy, adjuster time saved, error rates or cost.
- "More accurate" is asserted, not measured in the talk.
- Company conference; AWS is tagged in the video description.
Why it matters for companies
The talk is a template for regulated document review: keep models on-premises, structure messy inputs before reasoning over them, and give people the decision rather than a rubber stamp. The knowledge step applies widely. Converting slide-deck manuals into a single text source came before any retrieval system, and it is the step many firms skip. Claims, HR case management and procurement review share the same pattern.
Key passages in translation
Korean original
그러다 보니까 이것들이 20분, 30분, 40분씩 걸려요.
So these end up taking 20, 30, 40 minutes each.
Korean original
금융권 규제 안에서 외부 API 호출 없이 내부 어 환경에서 이런 것들을 서빙하는 것이 굉장히 중요하기 때문입니다.
Because under financial-sector regulation it is very important to serve these internally, without calling external APIs.
Korean original
그 분석과 어 처리 과정의 완성은 결국 사람이 한다라는 측면으로 접근을 했습니다.
We approached it on the basis that the analysis and processing are ultimately completed by a person.
Summary based on the full transcript. Read the original for the full argument and data.Watch the original on youtube.com