人壽醫療理賠已成為保險公司營運與客戶體驗的關鍵接觸點，但現場審查高度倚賴個人經驗，面對大量非結構化醫療文件（病歷、手術紀錄、檢驗報告等）仍呈現「數位化但不智慧化」，易導致同案不同判、延遲給付與對外爭議風險；同時在案件量成長與專業人力稀缺下，本報告進一步探討可落地的人機協作應用方案。
本報告個案分析以 T 保險公司為場域，結合 2022–2024 產業公開統計，呈現其延遲給付問題相較同業更為突出。產業整體延遲率約 2%，T 公司約 7.44%，且延遲案件占比顯著高於其案件量占比；同時問卷回收樣本涵蓋全體線上審查人員約 38%，可支持高頻痛點與耗時結構之描述性分析。
在此基礎上，本報告探討建構「代理式AI醫療理賠審查」的可行性。以智慧理賠輔助系統為核心，採合規優先、輔助而非取代、模組化可分階段導入、過程可視且結果可追溯等原則，並以三層設計串接審查循環：第一層將條款與審查原則（含評議／訴訟見解）規則化，建立公司一致尺度；第二層以多模組嵌入現行流程，整合跨系統資料形成「案件資料包」以降低等資料、找資料與溝通往返成本；第三層以短期／長期／持久記憶沉澱結案經驗，形成可持續優化的規則與治理機制，讓審查經驗轉為組織資產，兼顧效率、品質與風險可控。

Life insurance medical claims have become a key contact point for insurance companies and customer experience. However, the review process still depends heavily on personal experience. When dealing with a large number of unstructured medical documents (such as medical records, surgery reports, and test results), the system is “digital but not intelligent.” This may lead to inconsistent decisions for similar cases, delayed payments, and higher dispute risks. At the same time, with increasing case volume and a shortage of professional staff, this report explores practical human-AI collaboration solutions.
This report uses T Insurance Company as a case study. It combines industry data from 2022 to 2024 and shows that its delayed payment problem is more serious than the industry average. The overall industry delay rate is about 2%, while T Company is about 7.44%. The proportion of delayed cases is also much higher than its share of total cases. In addition, the questionnaire covers about 38% of all online reviewers, which supports the analysis of common problems and time-consuming tasks.
Based on this, the report studies the feasibility of building an “agentic AI medical claims review” system. The core idea is a smart claims support system, with principles such as compliance first, assisting rather than replacing humans, modular design for step-by-step implementation, and clear and traceable processes. The system is designed in three layers.
The first layer turns policy terms and review principles (including dispute and court decisions) into rules to create a consistent standard. The second layer uses multiple modules in the current workflow and integrates data from different systems to build a “case data package,” which reduces the time spent on collecting and checking information. The third layer uses short-term, long-term, and persistent memory to store case experience. This helps improve rules and governance over time, turning review experience into organizational knowledge, while balancing efficiency, quality, and risk control.

中文文獻
（一）政策與實務文件
中華民國人壽保險商業同業公會（2025）。保險業運用人工智慧系統自律規範。
保險業公開資訊觀測站（2026）。人力資源概況（表05020511）與保險申請評議統計（表07090901、表07070710）。
金融監督管理委員會（2024）。金融業運用人工智慧（AI）指引。

英文文獻
（一）國際監理與政策文件
Financial Stability Board. (2025). Monitoring adoption of artificial intelligence and related vulnerabilities in the financial sector. https://www.fsb.org/2025/10/monitoring-adoption-of-artificial-intelligence-and-related-vulnerabilities-in-the-financial-sector/
（二）學術文獻（Journal / ArXiv / Report）
Abuadbba, A., Sultan, N., Nepal, S., & Jha, S. (2026). Human Society-Inspired Approaches to Agentic AI Security: The 4C Framework. arXiv preprint arXiv:2602.01942.
Bandara, E., Hewa, T., Gore, R., Shetty, S., Mukkamala, R., Foytik, P., ... & Loganathan, N. (2025). Towards Responsible and Explainable AI Agents with Consensus-Driven Reasoning. arXiv preprint arXiv:2512.21699.
Dang, Y., Qian, C., Luo, X., Fan, J., Xie, Z., Shi, R., ... & Sun, M. (2025). Multi-agent collaboration via evolving orchestration. arXiv preprint arXiv:2505.19591.
Hua, Q., Ye, L., Fu, D., Xiao, Y., Cai, X., Wu, Y., ... & Liu, P. (2025). Context engineering 2.0: The context of context engineering. arXiv preprint arXiv:2510.26493.
Kahneman, D., Sibony, O., & Sunstein, C. R. (2021). Noise: A flaw in human judgment. Little, Brown Spark.
Kim, Y., Gu, K., Park, C., Park, C., Schmidgall, S., Heydari, A. A., ... & Liu, X. (2025). Towards a science of scaling agent systems. arXiv preprint arXiv:2512.08296.
Raza, S., Sapkota, R., Karkee, M., & Emmanouilidis, C. (2025). TRiSM for Agentic AI: A Review of Trust, Risk, and Security Management in LLM-based Agentic Multi-Agent Systems. ArXiv, abs/2506.04133.
Sajid, M. (2025). Multi-agentic automation for evaluating property claims in underwriting. Open Journal of Applied Sciences, 15, 819–833. https://doi.org/10.4236/ojapps.2025.154055
（三）網路資源（Blog / 技術文章）
Breunig, D. (2025, June 22). How long contexts fail. dbreunig.com. https://www.dbreunig.com/2025/06/22/how-contexts-fail-and-how-to-fix-them.html
Chase, H. (2025, June 23). The rise of “context engineering”. LangChain Blog. https://blog.langchain.com/the-rise-of-context-engineering/
Çelik, T., & Markewich, L. (2025, July 3). Context engineering - What it is, and techniques to consider. LlamaIndex. https://www.llamaindex.ai/blog/context-engineering-what-it-is-and-techniques-to-consider
Gulli, A., Nigam, L., Wiesinger, J., Vuskovic, V., Sigler, I., Nardini, I., Stroppa, N., Kartakis, S., Saribekyan, N., Nawalgaria, A., & Bount, A. (2025). Agents companion. Kaggle. https://www.kaggle.com/whitepaper-agent-companion
Martin, L. (2025, June 23). Context engineering for agents. Lance’s Blog. https://rlancemartin.github.io/2025/06/23/context_engineering/
Schmid, P. (2025, June 30). The new skill in AI is not prompting, it’s context engineering. Philschmid. https://www.philschmid.de/context-engineering
Yan, W. (2025, June 12). Don’t build multi-agents. Cognition. https://cognition.ai/blog/dont-build-multi-agents