FintechChina
← Sticker Wall
AGENT · DETAIL
Cybernative Claims Review Agent: 3-second document verification, saving small and medium-sized insurers 30% on annual claims costs
Workflow: Automatically fetches newly uploaded policies and claims documents from agencies daily, uses OCR to extract policy numbe
AGENT
Key Fields
FIELD STAMPSIndustryFintech
RegionChina
ScaleSME
ChannelOnline
🔧 Workflow
Automatically fetches newly uploaded policies and claims documents from agencies daily, uses OCR to extract policy numbers, diagnosis certificates, and invoice amounts, compares them with the policy clause knowledge base to calculate payouts, and runs a fraud rule engine to screen for duplicate claims and conflicting materials. It then generates a claims review report with confidence scores and pushes it to the store manager. Once manually confirmed, documents are automatically archived. The average processing time per claim is compressed from 2 hours to 5 minutes, with humans reviewing only flagged suspicious cases. Every morning at dawn, the system automatically compares historical fraud case databases, immediately pushing alerts for abnormal documents along with an evidence chain summary to facilitate manual investigation.
🛠 Setup Requirements
Requires knowledge of insurance claims workflows rather than deep programming; can be built using n8n to orchestrate OCR nodes and LLM APIs without training models. Prepare an insurance clause knowledge base (loadable via RAG), integrate with TextIn or Baidu OCR for document parsing, and build a minimal viable loop to pilot in a single store within 2 to 4 weeks. A Docker-compatible server or mini-PC is needed to run MySQL, MinIO, and a vector database locally to prevent sensitive image data from leaking externally. Workflows can also be built using Coze or Dify, but exporting to a private environment is more secure.
🧰 Toolchain
- 🔧 n8n
- 🔧 TextIn OCR API
- 🔧 Claude API
- 🔧 DeepSeek API
💰 Revenue
Charges small and medium-sized insurance institutions or agencies an annual subscription fee of 60,000 to 100,000 RMB, plus a per-claim usage fee of 2 to 3 RMB based on the store's monthly claims volume. Signing 2 to 3 clients initially yields a monthly income of about 20,000 to 40,000 RMB, which can scale to an annual revenue of 300,000 RMB through renewals and referrals. If switched to a claims-commission model of 5 to 10 RMB per case, processing 5,000 cases monthly generates 25,000 to 50,000 RMB with no upper limit.
💸 Cost
TextIn OCR costs 500 to 1,500 RMB monthly based on page usage, LLM API expenditure is around 3,000 RMB monthly, and the n8n self-hosted server costs 200 RMB monthly, keeping total costs under 5,000 RMB per month. If serving only a single store initially, free OCR quotas can be utilized to reduce costs to under 800 RMB monthly, providing a wider profit margin.
⏱ Time Investment
30 hours per week initially, dropping to about 20 hours per week after stabilization, mainly spent on manual review and rule base updates. Rule updates are concentrated at the end of the month, while routine maintenance involves checking alert logs for half an hour daily and adjusting prompts and databases based on client feedback over weekends.
🚀 Getting Started
Find a familiar insurance agency or auto insurance repair shop and deliver an initial minimum viable feature: ingest same-day accident acceptance materials into the Agent to automatically generate preliminary review opinions, completing the OCR and clause comparison loop using n8n. After validating the loop, negotiate an annual subscription fee based on audit labor hours saved, then replicate the model to other agencies in the same city. Demonstrate the 3-second claim review opinion to the store manager using a manual prototype or TextIn's public DEMO before signing a pilot agreement to avoid empty talks.
🔑 Keys to Success
- ✅ Structured decomposition of claims clause libraries
- ✅ Manual review fallback design
- ✅ Minute-level response experience
- ✅ Per-case pricing tied to business volume
- ✅ Training boundaries using real claim rejection cases
⚠️ 风险
- ⚠️ Forged diagnostic materials are difficult to identify with 100% accuracy; AI conclusions cannot directly replace payout decisions and require human judgment as the final authority.
- ⚠️ Insurance data compliance barriers are high; localization or desensitization processing is required during deployment, and cross-regional operations require corresponding qualifications to avoid regulatory penalties.
- ⚠️ LLM outputs are subject to hallucinations and clause comparison errors, requiring periodic back-testing of the rule base using real claims samples to prevent accumulated misjudgment rates from leading to customer churn.
📌 Real Cases
- 📌 TextIn Claim Agent smart claims platform achieves 3-second review of claims materials, covering full-process intelligent upgrades.
- 📌 Yuanbao advances claims processing to the minute level through a multi-agent architecture, serving as an AI implementation benchmark for insurance institutions.
- 📌 Qusar claims automation is measured to save about 30% in costs under the Guidewire framework, and personal Agents can batch-replicate similar effects for small insurers.
- 📌 Shouhui's mid-2026 report shows AI penetration across the entire underwriting and claims process, providing a data pipeline design reference for small teams to build their own SOPs.