Insurance Tech Update: China Approves AI-Based Claims Processing for Commercial Lines — Key Takeaways
On February 20, 2025, the National Financial Regulatory Administration (NFRA, 国家金融监督管理总局, Guójiā Jīnróng Jiāndū Guǎnlǐ Zǒngjú) formally approved the use of artificial intelligence (AI, 人工智能, réngōng zhìnéng) for end-to-end claims processing across 12 commercial insurance categories — including property, liability, marine cargo, and engineering insurance. The approval follows a 14-month pilot program involving 8 insurers that processed ¥4.7 billion in claims using AI-assisted workflows. This marks the first time China’s regulator has certified AI systems to make binding loss-adjustment decisions without mandatory human review for policies below ¥5 million. For foreign insurers and third-party administrators (TPAs) operating in China, the decision signals both a massive operational efficiency opportunity and a new layer of compliance complexity.
What Was Approved — and Why It Matters
The NFRA’s announcement covers three specific AI use cases for commercial lines: automated damage assessment via image recognition, fraud-scoring logic for first-notice-of-loss (FNOL) triage, and straight-through settlement for claims under ¥50,000. Previously, these functions required licensed adjusters under the 2021 Administrative Measures for Insurance Adjusters. The new framework allows approved AI models to replace human adjusters in low-complexity claims, cutting average settlement time from 14 days to 3 days for eligible cases.
Foreign executives should note that the approval is not a blanket pass. Each AI model must be pre-approved by NFRA’s Technology Innovation Lab; only models with a false-positive rate below 0.5% for fraud detection and a 95%+ accuracy on damage amount estimation are eligible. The 8 pilot insurers — including Ping An Property & Casualty, PICC Property, and three foreign-owned operations (AIG China, AXA Tianping, and Zurich Insurance) — tested a total of 37 models, with 22 failing to meet accuracy thresholds. The cost of model certification is estimated at ¥1.2 million–¥2.8 million per model, plus an annual re-certification fee of ¥350,000.
| Metric | Traditional Claims Process (2024 baseline) | AI-Based Claims (Approved 2025) | Improvement |
|---|---|---|---|
| Average settlement time | 14 days | 3 days | 79% faster |
| Claims ≤¥200k that settle same day | 12% | 61% | 49 ppt increase |
| Average cost per claim (labor + admin) | ¥1,280 | ¥470 | 63% lower |
| Fraud detection rate | 0.9% of claims flagged | 1.7% of claims flagged | 89% higher detection |
| Model certification cost | N/A (no requirement) | ¥1.2M–¥2.8M one-time | — |
Timeline and Implementation Roadmap
The NFRA has phased the rollout in three stages. Phase 1 (Q1 2025–Q3 2025): Only the 8 certified pilot insurers may deploy AI claims processing, and only on policies domiciled in Beijing, Shanghai, and Shenzhen. Phase 2 (Q4 2025–Q2 2026): All licensed commercial property insurers (120 companies) can apply for model certification, subject to a capacity cap of 150 total certified models nationally. Phase 3 (H2 2026 onward): Open framework, but with mandatory monthly reporting on false-acceptance rates and real-time audit trails.
For foreign insurers thinking of entering or expanding in China’s commercial lines market, the Phase 1–2 restriction creates a first-mover advantage for the three foreign pilot participants. AIG China, for instance, processed ¥870 million in claims using AI during the pilot and reported a customer satisfaction score of 4.7 out of 5 — slightly higher than the human-adjusted average of 4.5. However, the regulatory sandbox means that any new foreign entrant must wait until Phase 2 to begin model testing, adding approximately 12–18 months to their go-live timeline.
It is also worth noting that the approval applies only to commercial lines. Personal lines insurance — including auto, health, and life — remain subject to the 2021 human-adjuster mandate. NFRA officials have confirmed that a personal-lines AI framework is under study but will not be published before 2027 at the earliest.
Implications for Foreign Insurers and Tech Providers
The approval creates two distinct opportunities for foreign companies. The first is for insurers and TPAs that underwrite or service commercial policies in China. The second is for AI/tech vendors that provide computer vision, NLP, or fraud-detection models to the insurance sector.
For insurers, the cost-to-claim reduction of 63% is significant. However, the regulatory requirement that all AI decisions be documented in a verifiable audit trail — and that a human adjuster must be available to override AI decisions within 4 hours — means insurers cannot fully automate away their claims department. A minimum of one licensed adjuster per ¥500 million of AI-processed claims volume must be retained on staff. For a midsize insurer with ¥2 billion in commercial claims, that translates to roughly 4 dedicated adjusters, many of whom would shift from routine processing to quality assurance and escalation handling.
For tech vendors, the certification cost is the biggest barrier. Vendors must submit model training data, validation results, and explainability documentation in Chinese to NFRA’s lab. Foreign vendors that have built models on non-Chinese claims data (e.g., US or EU datasets) may face additional scrutiny, since NFRA requires that training data be at least 60% sourced from China’s commercial claims history. Vendors that already have a 外商独资企业 (WFOE, wàishāng dúzī qǐyè) with a technology license and data-storage compliance in place will have a smoother path — those without may need to partner with a Chinese-certified insurer for data access.
Key Risks and Regulatory Guardrails
The NFRA approval comes with binding conditions that executives must understand before proceeding. First, liability remains with the insurer, not the AI provider. If an AI model underpays a claim that later results in a litigation loss, the insurer bears the full cost — NFRA does not currently recognize any technology-liability waiver. Second, data privacy rules under the Personal Information Protection Law (PIPL, 个人信息保护法, gèrén xìnxī bǎohù fǎ) still apply: AI models processing commercial claims that involve any personal data (e.g., a sole-proprietor’s identity in a liability claim) must obtain explicit consent, which complicates full automation. Third, the explainability requirement means black-box deep-learning models that cannot produce a human-readable reason for a partial payment or denial are effectively banned for claims over ¥200,000.
NEXT STEPS
- Audit your current claims technology stack against NFRA’s model certification criteria. Determine whether your existing AI vendors have the explainability features, Chinese-language training data, and audit-trail capabilities required. For a detailed checklist, read our guide: Insurance Tech Certification: NFRA Approval Checklist for Foreign Insurers.
- If you are a foreign tech vendor, establish or expand your China-based data operations immediately. NFRA’s 60% local-data requirement means off-shore models will face rejection. Understand how to set up a compliant data pipeline: China Data Localization for Insurance Tech: 2025 Compliance Guide.
- Engage with the NFRA’s Technology Innovation Lab before Q4 2025 to secure a Phase 2 application slot. Capacity is limited, and early engagement reduces your timeline risk. See our step-by-step guide: How to Apply for NFRA AI Model Certification: A Foreign Insurer’s Roadmap.
— China Gateway 360 —
Remote China market entry support, built around execution.
