Multi-agent-based claim settlement decision method, device, equipment and medium

Abstract

This invention relates to the field of artificial intelligence technology and can be applied to business system platforms such as fintech. It discloses a multi-agent-based claims decision-making method, comprising: acquiring multimodal data and preprocessing the multimodal data to generate structured data; inputting the structured data into a verification agent for verification; when the structured data passes verification, transmitting the structured data to an inference agent; the inference agent performing causal inference and risk identification based on loss assessment standards, risk scoring strategies, and structured data, outputting the compensation amount and risk indicators; inputting the compensation amount and risk indicators into the verification agent for review and quantification, outputting a quantified risk score, and generating a claims decision report based on the quantified risk score. This invention utilizes multiple agents for collaborative processing, generating a claims report after verification, inference, review, and quantification, thereby improving claims efficiency and accuracy, and enhancing system adaptability and security.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a claims decision-making method, apparatus, device, and storage medium based on multi-agent systems. Background Technology

[0002] In the claims processing sector, the initial application of artificial intelligence technology has not yet resolved core technical pain points. Existing solutions have significant shortcomings in processing efficiency, scenario adaptability, risk response, and system security, severely impacting business quality improvement. At the data level, this area involves multiple stakeholders and diverse data types. The modular design of existing systems creates barriers to data interaction, forming "data silos." The conversion between unstructured and structured data relies on manual operation, which is not only inefficient but also prone to introducing human error. Furthermore, there is a lack of effective integration mechanisms between real-time dynamic data generated by emerging technologies and traditional systems, highlighting the data gap problem. At the process level, most systems rely on predefined rule engines, exhibiting a rigid characteristic and lacking business flexibility and adaptability. Faced with complex cases with blurred boundaries, the system cannot make autonomous reasoning decisions; when encountering new business models or special scenarios, manual intervention is necessary. It is difficult to dynamically adjust and self-evolve based on historical cases, real-time feedback, and changes in the external environment, resulting in a large number of complex cases still requiring manual review, hindering automation and intelligent upgrades. Regarding adaptation to new scenarios, changes in business forms brought about by technological innovation make traditional core logic difficult to cope with, and there is a lack of targeted assessment models and decision-making logic for new risk scenarios. Meanwhile, critical data is difficult to share effectively due to factors such as privacy protection and commercial interests, resulting in a lack of sufficient data support for core aspects such as business pricing and risk assessment, thus hindering the precise development of business. At the system collaboration level, there is a prominent contradiction between security and efficiency in multi-agent architectures. Network attacks may tamper with the interaction data between agents, undermining system stability; however, an excessive pursuit of data synchronization will waste network bandwidth and computing resources. Current technologies have not yet formed an effective collaboration mechanism that balances security protection and resource optimization, further affecting the stability and efficiency of business operations.

[0003] At the data level, fintech involves multiple stakeholders, including banks, insurance companies, and third-party institutions, encompassing diverse data types such as transaction records, user profiles, and risk data. Existing modular systems result in significant data interaction barriers, highlighting the "data silo" phenomenon. The conversion between unstructured and structured data relies on manual operation, reducing processing efficiency and increasing operational risks. Furthermore, the lack of effective integration mechanisms between real-time dynamic data generated by emerging technologies and traditional systems exacerbates the data gap, making business decision-making more difficult. At the process level, most platforms rely on predefined rule engines, exhibiting significant rigidity and lacking business flexibility and adaptability. When facing innovative financial scenarios, boundary-setting business cases, or new risk models, they cannot make autonomous reasoning decisions, relying excessively on manual intervention and hindering dynamic adjustment and self-evolution, thus restricting the scaling and intelligent advancement of business. At the system collaboration level, the contradiction between security and efficiency is prominent in multi-agent architectures. Network attacks may tamper with interactive data, threatening financial business security, while continuous data synchronization excessively consumes network bandwidth and computing resources. Existing technologies have not yet formed a collaborative mechanism that balances security protection and resource optimization, affecting the stability and efficiency of fintech services.

[0004] In the data aspect of auto insurance claims, multiple stakeholders are involved, including car owners, repair shops, and insurance companies. This involves diverse data such as policy information, accident footage, and damage assessment records. Existing modular systems suffer from poor data exchange; unstructured data must be manually converted into structured data before it can be integrated into subsequent processes, reducing efficiency and introducing human error. Furthermore, there is a gap between the real-time driving and sensor data generated by intelligent connected vehicles and traditional claims systems, lacking an effective integration mechanism. At the process level, the system relies on a predefined rule engine, exhibiting rigidity. It cannot make flexible decisions in boundary cases where compensation is optional, and it requires manual intervention in new scenarios such as intelligent driving accidents. It struggles to dynamically adjust based on historical cases and real-time feedback, resulting in 70% of complex cases still requiring manual review, hindering automation upgrades. In terms of adapting to new scenarios and system collaboration, traditional claims logic struggles to address the challenges of liability determination arising from "human-machine co-driving," lacking targeted risk assessment and liability determination models. Furthermore, the difficulty in sharing automaker data due to privacy and commercial interests hinders pricing and actuarial support. In multi-agent collaboration, cyberattacks threaten data security, and continuous communication consumes significant resources. Existing technologies cannot balance security and efficiency, further impacting the high-quality development of claims services. Summary of the Invention

[0005] The main objective of this invention is to provide a multi-agent-based claims decision-making method, apparatus, device, and storage medium, aiming to solve the problem of low collaboration efficiency among multiple agents in the prior art when processing claims cases, which affects the efficiency and accuracy of claims processing.

[0006] To achieve the above objectives, the present invention provides a multi-agent-based claims decision-making method, comprising: Acquire multimodal data and preprocess the multimodal data to generate structured data; The structured data is input into the verification agent for verification. When the structured data passes the verification, the structured data is transmitted to the reasoning agent. The reasoning agent performs causal reasoning and risk identification based on loss assessment standards, risk scoring strategies, and structured data, and outputs the compensation amount and risk indicators. The compensation amount and risk indicators are input into the verification agent for review and quantification, a quantitative risk score is output, and a claims decision report is generated based on the quantitative risk score.

[0007] Furthermore, to achieve the above objectives, the present invention provides a multi-agent-based claims decision-making device, comprising: The data processing module is used to acquire multimodal data and preprocess the multimodal data to generate structured data; The data verification module is used to input the structured data into the verification agent for verification. When the structured data passes the verification, the structured data is transmitted to the reasoning agent. The claims reasoning module is used by the reasoning agent to perform causal reasoning and risk identification based on loss assessment standards, risk scoring strategies and structured data, and output the compensation amount and risk indicators. The claims decision module is used to input the compensation amount and risk indicators into the verification agent for review and quantification, output a quantitative risk score, and generate a claims decision report based on the quantitative risk score.

[0008] Furthermore, to achieve the above objectives, the present invention also provides a computer device, the computer device including a memory, a processor, and a multi-agent-based claims decision-making program stored in the memory and executable on the processor, wherein when the multi-agent-based claims decision-making program is executed by the processor, it implements the steps of the multi-agent-based claims decision-making method as described above.

[0009] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a multi-agent-based claims decision-making program, wherein the multi-agent-based claims decision-making program, when executed by a processor, implements the steps of the multi-agent-based claims decision-making method described above.

[0010] Beneficial Effects: This invention relates to the field of artificial intelligence technology and can be applied to business system platforms such as fintech. It discloses a multi-agent-based claims decision-making method, comprising: acquiring multimodal data and preprocessing the multimodal data to generate structured data; inputting the structured data into a verification agent for verification; when the structured data passes verification, transmitting the structured data to an inference agent; the inference agent performing causal inference and risk identification based on loss assessment standards, risk scoring strategies, and structured data, outputting the compensation amount and risk indicators; inputting the compensation amount and risk indicators into the verification agent for review and quantification, outputting a quantified risk score, and generating a claims decision report based on the quantified risk score. This invention utilizes multiple agents for collaborative processing, generating a claims report after verification, inference, review, and quantification, thereby improving claims efficiency and accuracy, and enhancing system adaptability and security. Attached Figure Description

[0011] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a schematic diagram of an application environment for a multi-agent-based claims decision-making method according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating an embodiment of the multi-agent-based claims decision-making method of the present invention; Figure 3 This is a flowchart illustrating another embodiment of the multi-agent claims decision-making method of the present invention; Figure 4 This is a schematic diagram of the functional modules of a preferred embodiment of the multi-agent claims decision-making device of the present invention; Figure 5 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention; Figure 6 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0012] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0013] The multi-agent-based claims decision-making method provided in this invention can be applied to, for example... Figure 1In this application environment, the user terminal communicates with the server via a network. The server can obtain multimodal data from the user terminal, preprocess the multimodal data to generate structured data, input the structured data into a verification agent for verification, and when the structured data passes verification, it is transmitted to an inference agent. The inference agent performs causal inference and risk identification based on the loss assessment standards, risk scoring strategies, and structured data, and outputs the compensation amount and risk indicators. The compensation amount and risk indicators are input into the verification agent for review and quantification, outputting a quantified risk score, and generating a claims decision report based on the quantified risk score. This invention uses multiple agents for collaborative processing, generating a claims report after verification, inference, review, and quantification, improving claims efficiency and accuracy, and enhancing system adaptability and security. The user terminal can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will be described in detail below through specific embodiments.

[0014] Please see Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of the multi-agent-based claims decision-making method provided by the present invention. It should be noted that although the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0015] like Figure 2 As shown, the multi-agent-based claims decision-making method proposed in this invention includes the following steps: S100: Acquire multimodal data and preprocess the multimodal data to generate structured data; S200. The structured data is input into the verification agent for verification. When the structured data passes the verification, the structured data is transmitted to the reasoning agent. S300: The reasoning agent performs causal reasoning and risk identification based on loss assessment standards, risk scoring strategies, and structured data, and outputs the compensation amount and risk indicators. S400. Input the compensation amount and risk indicators into the verification agent for review and quantification, output a quantitative risk score, and generate a claims decision report based on the quantitative risk score.

[0016] In this embodiment, as Figure 3As shown, the system first acquires multimodal data, including various types such as ID cards uploaded by car owners, accident photos, damage assessment reports, and insurance policies. Then, it uses Optical Character Recognition (OCR, a technology that converts text in images into editable text) and multimodal large-scale models (such as qwen-vl-max, an AI model with powerful image understanding and text extraction capabilities) to recognize and analyze this data. The system presets exclusive prompt word templates for different material types to guide the model in accurately extracting key fields such as license plate numbers, repair items, and amounts. Simultaneously, it performs data cleaning and standardization operations, ultimately converting unstructured image and text information into standardized JSON (JavaScript Object Notation, a lightweight data exchange format easily parsed and generated by machines) data, providing high-quality data input for subsequent intelligent agent processing.

[0017] The preprocessed structured data is input into the Checker Agent (acting as a quality control "safety valve") for verification. The Checker Agent is equipped with both hard constraints (such as rule-based verification) and soft constraints (such as model-based judgment). It verifies the integrity of the structured data fields and compares key information such as names, amounts, and timestamps (accident dates) across documents, including ID cards, insurance policies, and claims settlement reports. Only when the structured data meets the integrity and consistency requirements and passes verification is it transmitted to the Inference Agent (responsible for core claims logic judgments, essentially the "brain" of claims decisions).

[0018] After receiving structured data, the reasoning agent queries the knowledge base agent (Agentic RAG, the "knowledge heart" of the system, responsible for retrieving relevant knowledge from the specialized knowledge base) to obtain the loss assessment standards, risk scoring strategies, and domain-specific question-and-answer knowledge corresponding to the current case. The knowledge base agent employs a multi-level retrieval strategy: first-level semantic retrieval, second-level business keyword filtering, and third-level verification agent to validate the completeness of the retrieved content, ensuring the accuracy and comprehensiveness of the knowledge provided. Combining the acquired structured data and the retrieved knowledge, the reasoning agent uses a large model (such as qwen-plus) for causal reasoning and risk identification, ultimately outputting the compensation amount and risk indicators.

[0019] The compensation amount and risk indicators are then input into a verification agent for review and quantification. The verification agent first performs a secondary review of the reasonableness of the compensation amount and risk indicators. Simultaneously, the system invokes an operations research-based risk scoring model to dynamically determine the optimal weights for various risk indicators, such as completeness of materials, abnormal amounts, and historical compensation frequency, by solving a constrained optimization problem (e.g., using Python's Pulp library). This process calculates an objective, quantified risk score. Finally, based on the quantified risk score, the case's risk level is determined, and a claims decision report is generated, including a compensation / rejection conclusion, reasoning basis, and risk point analysis, completing the entire process.

[0020] For example, in the auto insurance field, multimodal data preprocessing can quickly integrate data from car owners, repair shops, insurance companies, and other parties. The dual verification mechanism of the verification agent reduces information errors and lowers the risk of fraud. The reasoning agent, combined with a knowledge base, achieves accurate risk identification, and quantified risk scoring makes claims decisions more scientific. For instance, in the face of autonomous driving accidents, by accessing anonymized data provided by automakers, the system can collaboratively assign responsibilities to drivers, automakers, algorithm providers, and other parties, significantly improving claims efficiency and accuracy. This reduces claims review time from several days to hours or even minutes, achieving automated processing for over 80% of cases.

[0021] In the fintech sector, this technology can be applied to scenarios such as credit approval, insurance product pricing, and fraud prevention. Taking credit approval as an example, multimodal data encompasses the applicant's identity verification, income verification, bank statements, asset verification, and credit reports. After preprocessing and transforming this data into structured data, a verification agent checks the authenticity and completeness of the data, such as verifying the consistency between income verification and bank statements. The reasoning agent, based on knowledge base content such as credit policies and risk assessment rules, performs causal reasoning and risk identification regarding the applicant's creditworthiness and repayment ability. The verification agent further reviews the compensation amount and risk indicators, quantifies the risk score, and determines the credit approval result, credit limit, and interest rate level based on the score, generating a credit decision report. This process automates and expedits credit approval while enhancing decision security through precise risk quantification, effectively preventing financial risks. In insurance product pricing scenarios, multimodal data can be used to integrate multi-dimensional user information and continuously optimize pricing models using a knowledge flywheel mechanism to adapt to the risk characteristics of different users, achieving personalized pricing.

[0022] In one embodiment, step S100 includes: S101. Obtain multimodal data uploaded by the user, wherein the multimodal data includes text information data and image information data; S102. Transmit the multimodal data to the material collection agent; S103. The material collection agent calls the optical character recognition engine and the multimodal large model, and combines the material analysis rule template to fuse and extract the multimodal data to generate structured data.

[0023] In this embodiment, users upload multimodal data through the system. This data includes various types such as ID cards, accident photos, damage assessment reports, and insurance policies, encompassing both textual information and unstructured image data, comprehensively covering the core materials required for auto insurance claims. Subsequently, this data is transmitted to the Material Collection Agent (a module specifically responsible for data collection and preliminary processing). This agent serves as the front-end entry point for data processing, undertaking the core responsibility of driving the data parsing process.

[0024] When the material collection agent initiates the data processing flow, it invokes Optical Character Recognition (OCR) technology and a multimodal large model (such as qwen-vl-max). To ensure the accuracy of extracted information, the system presets material parsing rule templates (dedicated prompt word templates) for different material types. For example, the template for ID cards guides the model to focus on key fields such as name, ID number, and expiration date, while the template for damage assessment reports specifies the extraction requirements for core information such as repair items, parts prices, and labor costs. With the synergy of these technologies, OCR technology completes the initial recognition of image text, while the multimodal large model performs semantic understanding and key information filtering based on the prompt word templates. Simultaneously, it cleans the recognition results, removing redundant information, correcting recognition errors, and converting unstructured images and scattered text into standardized JSON format data (JavaScript), ultimately generating structured data to provide high-quality data support for subsequent verification and inference processes.

[0025] For example, in the field of auto insurance, multimodal data such as accident photos, policy scans, and damage assessment reports uploaded by car owners can be processed by the material collection agent using OCR technology and multimodal large models. Key information such as license plate number, accident location, and repair amount can be quickly extracted and transformed into standardized structured data. This avoids errors and delays caused by manual data entry, lays the foundation for the consistency verification of the subsequent verification agent and the risk judgment of the reasoning agent, and greatly improves the automation efficiency of the claims process.

[0026] In the healthcare field, this technology can be applied to scenarios such as medical insurance claims and electronic medical record integration. Multimodal data uploaded by patients, including photos of their medical records, examination reports, payment receipts, and medical insurance participation certificates, can be processed using OCR technology and a multimodal large-scale model by a data collection intelligent agent. Combined with pre-set medical material prompt templates, key fields such as patient name, diagnosis, treatment items, cost details, and medical insurance account information can be accurately extracted and transformed into structured data. This not only solves the problems of diverse medical data formats and tedious manual processing, but also provides accurate data support for compliance verification of medical insurance claims and medical expense calculation. Furthermore, it creates conditions for standardized storage and cross-hospital sharing of electronic medical records.

[0027] In the fintech sector, this technology can be widely applied to scenarios such as credit approval, insurance application, and fraud verification. Taking credit approval as an example, multimodal data uploaded by applicants, including ID photos, income statements, bank statement screenshots, and asset certificates, can be processed by an intelligent data collection agent. Guided by customized prompt templates, it can quickly extract and structure key data such as the applicant's identity information, income amount, transaction details, and asset valuation. This structured data can be directly connected to the risk assessment module of the credit approval system, providing accurate evidence for the applicant's credit rating and repayment ability assessment. In insurance application scenarios, materials uploaded by policyholders, such as health certificates and proof of assets, can be quickly verified after structured processing, improving the efficiency of the application process. Simultaneously, it provides data support for risk screening, helping fintech businesses achieve automated, accurate, and efficient data processing.

[0028] In one embodiment, step S200 includes: S201. Input the structured data output by the material collection agent into the verification agent; S202. The structured data is double-verified by the verification agent according to hard constraint rules and soft constraint rules. S203. When the verification result indicates missing information or inconsistent key information, the material collection agent will remind the user to supplement or correct the structured data. S204. When the verification result shows that the fields are complete and the key information is consistent, the structured data is transmitted to the reasoning agent.

[0029] In this embodiment, the implementation process of structured data verification and transfer is a crucial step in ensuring the accuracy of subsequent claims decisions. Its core relies on a dual verification mechanism of the Checker Agent (an agent module that acts as a "safety valve" for system quality control) to achieve closed-loop management. First, the standardized JSON-formatted structured data output by the Material Collection Agent is directly input into the Checker Agent. This data covers key fields extracted from multimodal materials such as ID cards, insurance policies, and accident photos, including vehicle owner's name, license plate number, damage assessment amount, and accident time. The Checker Agent initiates a dual verification process. Hard Constraint Rules refer to explicit business rule verification, such as whether the materials contain mandatory fields like ID card number and policy number, and whether the amount format conforms to regulations. Soft Constraint Rules, on the other hand, are based on model judgment, using algorithms to evaluate the reasonableness of the data, such as the deviation of the accident loss amount from the average level of similar cases, and the matching degree between repair items and the accident location.

[0030] During the verification process, the verification agent performs precise checks through cross-document comparisons. For example, it verifies whether the name on the ID card matches the policyholder's name, whether the amount on the damage assessment report matches the cost details in the repair record, and whether there are logical conflicts between the accident timestamp and the policy validity period. If the verification results show missing information (e.g., no damage assessment report provided, ID card validity period not filled in) or inconsistencies in key information (e.g., misspelled name, contradictory amounts), the verification agent will feed back the anomaly to the context engineering agent. This agent will then drive the process back to the material collection state. The material collection agent will then prompt the user to supplement missing materials or correct errors through a dialogue. The material collection agent's prompt template will embed the current material collection status, giving it "memory" capabilities for efficient multi-round communication. If the verification results show that the structured data fields are complete and the key information is consistent, meeting the requirements for subsequent processing, the verification agent will transmit the structured data to the inference agent, providing reliable data support for risk reasoning and decision-making.

[0031] For example, in the field of auto insurance, the structured data uploaded by car owners can be double-verified by a verification intelligent agent, which can quickly identify missing or contradictory information in materials such as ID cards, insurance policies, and damage assessment reports. For instance, it can promptly detect discrepancies between the license plate number and the information registered on the insurance policy, or between the repair items and the damage parts shown in the accident photos. The material collection intelligent agent can then remind users to make corrections, avoiding delays in the claims process due to information issues. This lays the foundation for the subsequent reasoning intelligent agent to accurately determine accident liability and assess claims risks, significantly improving the efficiency and accuracy of claims review.

[0032] In the fintech sector, this technology is widely applicable to scenarios such as credit approval, insurance underwriting, and fraud verification. Taking credit approval as an example, the structured data uploaded by the applicant (such as ID information, income statements, bank statements, and asset certificates) is verified by an intelligent verification system. Hard constraint rules check for missing key supporting documents and completeness of information, while soft constraint rules determine whether the income amount matches the bank statement details, whether the asset valuation is reasonable compared to the applied credit limit, and whether the credit record is consistent with the description of repayment ability. If information is missing or contradictory, the material collection intelligent system reminds the applicant to supplement and correct it, ensuring the authenticity and reliability of the approval data and providing an accurate basis for subsequent risk assessment. In insurance underwriting scenarios, the structured data uploaded by the policyholder, such as health declarations and proof of assets, is verified to identify false or missing information, helping insurance companies accurately assess underwriting risks and ensuring the compliance and security of insurance business.

[0033] In one embodiment, step S300 includes: S3011. When the reasoning agent receives structured data, it initiates a query request to the knowledge base agent. S3012. The knowledge base agent uses a multi-level retrieval strategy to obtain the damage assessment standards and risk scoring strategies for the current case in the knowledge base according to the query request, and transmits the damage assessment standards and risk scoring strategies to the reasoning agent. S3013. The reasoning agent performs causal reasoning and risk identification based on the loss assessment standards, risk scoring strategies and structured data, and outputs the compensation amount and risk indicators.

[0034] In this embodiment, the Inference Agent acts as the "brain" of claims decision-making, and its core workflow revolves around accurately acquiring professional knowledge and deeply integrating data and rules. When the Inference Agent receives structured data verified by the Checker Agent, it immediately initiates a query request to the Agentic RAG (the "knowledge heart" of the system), specifying its requirement to obtain loss assessment standards, risk scoring strategies, and relevant domain Q&A knowledge matching the current case. The Agentic RAG responds to the query using a three-tiered, multi-level retrieval strategy: Level 1 semantic retrieval, based on natural language understanding technology, quickly matches knowledge entries in the expert knowledge base that are semantically related to the case; Level 2 business keyword filtering further focuses on core elements, such as case type, accident scenario, and relevant insurance types, filtering out more targeted content; Level 3, the Checker Agent verifies the completeness of the recalled knowledge, ensuring that the loss assessment standards transmitted to the Inference Agent are complete and the risk scoring strategies are unambiguous.

[0035] After receiving the damage assessment standards and risk scoring strategies, the reasoning agent combines structured data (such as vehicle owner information, accident details, damage assessment amount, repair items, etc.) and invokes a large model (such as qwen-plus) to initiate the causal reasoning and risk identification process. The causal reasoning stage focuses on the "matching logic between facts and rules," for example, judging the reasonableness of repairs based on the collision parts and repair items extracted from accident photos, and determining whether the repairs are within the coverage scope based on the accident time and policy validity period. Risk identification revolves around potential risk points, such as identifying high-frequency claim risks through historical claim frequency rules, determining whether the damage assessment amount exceeds a reasonable range through abnormal amount rules, and investigating fraud suspicions through material consistency association rules. Finally, the reasoning agent outputs the compensation amount and risk indicators, including "case compliance conclusion, risk point details, and rule matching basis," providing core evidence for subsequent review and quantitative scoring.

[0036] For example, in the field of auto insurance, the reasoning agent obtains accurate damage assessment standards (such as the prices of parts for different car models and repair labor costs) and risk scoring strategies (such as single-vehicle accident risk coefficients and cross-regional claim risk thresholds) through a knowledge base agent. Combined with structured accident information and damage assessment data, it quickly completes core judgments such as "whether the accident meets the compensation conditions, whether the repair items are necessary, and whether there is a risk of fraud." For instance, in the case of autonomous driving accidents, the reasoning agent can invoke dedicated damage assessment standards and liability allocation rules, combined with anonymized data provided by the automaker (such as the activation status of the autonomous driving system and sensor data), to infer the liability ratios of the driver, the automaker, and the algorithm provider, outputting clear risk identification results. This provides support for subsequent quantitative scoring and claims decisions, significantly improving the efficiency and accuracy of handling complex cases.

[0037] In the fintech sector, this technology can be widely applied to scenarios such as credit approval risk assessment, insurance underwriting, and anti-fraud verification. Taking credit approval as an example, the reasoning agent receives the applicant's structured data (such as income statements, bank statements, credit records, and asset information) and queries the knowledge base agent for credit policy standards (such as credit limit caps for different income levels and rules for handling credit delinquencies) and risk scoring strategies (such as debt-to-income ratio thresholds and risk level classifications based on the number of credit delinquencies). Through causal reasoning, it determines the match between the applicant's repayment ability and credit limit, and the compliance of their credit status with credit eligibility. It also identifies and eliminates risks such as false income statements and hidden liabilities, outputting the compensation amount and risk indicators, including "credit compliance conclusion, preliminary risk level assessment, and rule matching details." In insurance underwriting scenarios, the reasoning agent can combine the insured's structured health data with underwriting rules to infer underwriting risks, providing a scientific basis for underwriting decisions and helping fintech businesses achieve automated and precise risk assessment.

[0038] In one embodiment, step S400 includes: S401. Input the compensation amount and risk indicators into the verification agent, and review the compensation amount and risk indicators in combination with the loss assessment standard and risk scoring strategy; S402. Input the approved compensation amount and risk indicators into the operations research risk scoring model, and quantify the compensation amount and risk indicators by solving a constrained nonlinear integer programming problem, and output a quantitative risk score. S403. Based on the expert knowledge base, risk levels are divided according to the quantitative risk score, and claims decision and reasoning reports are generated based on the risk levels.

[0039] In this embodiment, the review and quantification of compensation amounts and risk indicators are crucial closed-loop links to ensure the accuracy and objectivity of claims decisions. This process relies entirely on the dual functions of the Checker Agent and the quantification capabilities of the operations research model. First, the compensation amount and risk indicators output by the Inference Agent, which include "case compliance conclusions, risk point details, and rule matching basis," are input back into the Checker Agent to initiate a second review process. At this point, the Checker Agent, in conjunction with previously invoked loss assessment standards and risk scoring strategies, comprehensively checks the logical consistency and rule matching accuracy of the compensation amount and risk indicators. For example, it verifies whether risk point identification is omitted, whether the matching between the compliance conclusion and loss assessment standards is rigorous, and whether rule citations are accurate. This ensures that the compensation amount and risk indicators have no logical loopholes or rule misinterpretations. Only compensation amounts and risk indicators that pass the review will proceed to the quantification stage.

[0040] Next, the approved compensation amount and risk indicators are input into the Operations Research Risk Scoring Model (ORM), a risk assessment tool based on mathematical optimization theory. The core of this model is solving a constrained nonlinear integer programming problem (an optimization problem where the objective function or constraints contain nonlinear functions and the decision variables are only integers). Using tools such as Python's Pulp library, the optimal weights for various risk indicators, such as completeness of materials, degree of anomalous amount, historical compensation frequency, and severity of risk points, are dynamically determined. This means that, based on the specific circumstances of the case, key risk indicators are given higher weights, while secondary indicators are given lower weights, thereby achieving accurate quantification of compensation amounts and risk indicators. Ultimately, an objective and comparable quantitative risk score (such as a 0-100 point system) is output.

[0041] Finally, based on preset rules in the Specialized Knowledge Base (a database storing professional knowledge such as loss assessment standards, risk level classification rules, and claims policies), the system classifies risk levels according to quantitative risk scores, typically into three levels: high, medium, and low. For example, a score above 80 is considered a high-risk case, 40-80 a medium-risk case, and below 40 a low-risk case. Then, a corresponding claims decision report is generated based on the risk level: low-risk cases directly generate an "Agree to Pay" decision and a detailed report; medium-risk cases generate a "Conditional Pay" or "Further Investigation" decision; and high-risk cases generate a "Deferred Pay" or "Reject Pay" decision. The report clearly states the quantitative risk score, risk level, risk point details, rule matching basis, decision conclusion, and core reasons, ensuring the decision-making process is traceable and explainable.

[0042] For example, in the field of auto insurance, this effectively solves the problems of vague risk assessment and strong subjectivity in traditional claims decision-making. The verification agent combines auto insurance damage assessment standards (such as the compensation ratio for different accident types and the upper limit of parts prices) and risk scoring strategies (such as the risk coefficient of single-vehicle accidents and the risk threshold for claims in different locations) to review the compensation amount and risk indicators, ensuring that the identification of risk points (such as abnormal damage assessment amount and mismatch between repair items and accidents) and compliance judgments are unbiased. The operations research model, through quantitative processing, transforms qualitative and quantitative information such as "accident severity, material completeness, and historical claims records" into specific scores, and then classifies risk levels based on a specialized knowledge base. High-risk cases (such as scores above 85 points, large abnormal amounts, and historical fraud records) trigger manual review and intervention, while medium- and low-risk cases (such as scores below 60 points, complete materials, and minor risk points) directly generate compensation decision reports. This ensures that the risk of high-risk cases is controllable and improves the processing efficiency of medium- and low-risk cases, shifting claims decision-making from "experience-based judgment" to "data-driven".

[0043] In the fintech business, it can be adapted to core scenarios such as credit approval risk assessment, insurance underwriting, and anti-fraud verification. Taking credit approval as an example, the verification agent combines credit policy standards (such as debt-to-income ratio thresholds and credit delinquency handling rules) and credit risk scoring strategies (such as implicit debt identification rules and characteristics of false income certificates) to review the compensation amount and risk indicators (such as "the applicant's repayment ability meets the standards, and there is a risk of credit delinquency"). The operations research model quantifies indicators such as "income stability, debt level, credit status, and number of risk points" into risk scores, and classifies them into high, medium, and low risk levels based on a specialized knowledge base. High-risk levels (such as a score of 80 or above, excessive debt, and...) Multiple overdue records generate a "reject credit" decision; medium-risk levels (e.g., a score of 60-80 with minor credit flaws) generate a "reduce credit limit and increase interest rate" decision; low-risk levels (e.g., a score below 40 with excellent qualifications and no risk points) generate a "agree to credit" decision and a detailed report. In insurance underwriting scenarios, quantitative scoring can be used to classify the insured's health risk and property risk levels, generating decisions such as underwriting, rejecting, or underwriting with increased premiums. This helps fintech businesses achieve efficient decision-making under the premise of controllable risk, balancing business development and risk prevention.

[0044] In one embodiment, step S403 further includes: S4031. Pre-construct a wrongful conviction understanding agent; S4032. Obtain historical claims data and input the historical claims data into the Wrongful Claims Understanding Agent for information extraction to obtain historical wrongful claims; S4033. Analyze the historical wrong claims, generate error types, generate modification plans based on the error types, and store the error types and corresponding modification plans in the expert database; S4034. Obtain the preset insurance terms and deconstruct the insurance terms to extract the core claims data; S4035. Generate a compensation strategy based on the core compensation data, store the core compensation data and the corresponding compensation strategy in the knowledge base, and update the knowledge base agent based on the core compensation data and the corresponding compensation strategy.

[0045] In this embodiment, the continuous evolution of the expert knowledge base and the dynamic updating of the knowledge base agent are achieved through two core paths: "error-driven optimization" and "term-driven iteration." The entire process relies on the synergy of the Badcase Understanding Agent (an agent module specifically designed to analyze claims errors and generate optimized solutions) and AI technology to ensure the accuracy and timeliness of the system's knowledge structure. First, the Badcase Understanding Agent needs to be pre-built. This agent possesses the capabilities of historical data mining, error identification, and solution generation, serving as the core carrier for bottom-up knowledge updates. Subsequently, the system acquires massive amounts of historical claims data, including structured information on past cases, payout amounts and risk indicators, claims decision reports, and subsequent verification and confirmation of erroneous claims records. After inputting this data into the Error Understanding Agent, the agent uses information extraction algorithms to filter out historical claims errors. These errors are then analyzed in depth and categorized into error types, such as misjudgments due to omissions in material verification, deviations in the application of risk scoring strategies, misinterpretations of damage assessment standards, and improper division of responsibility in autonomous driving cases. For different error types, the Error Understanding Agent combines business logic and expert experience to generate targeted modification plans, such as optimizing the field verification rules of the verification agent, adjusting the indicator weights of the operations research risk scoring model, and supplementing damage assessment standards for autonomous driving scenarios. Finally, after the error types and corresponding modification plans are confirmed by insurance experts, they are automatically or semi-automatically stored in the Specialized Knowledge Base (a core database storing professional knowledge such as damage assessment standards, risk rules, and optimization plans).

[0046] Simultaneously, the system acquires pre-set insurance terms (including newly released and updated auto insurance terms), and uses AI technology to deconstruct the insurance terms, removing redundant expressions and accurately extracting core claims data, such as coverage scope, compensation ratio, exclusions, maximum coverage amount, and compensation requirements for special scenarios (such as intelligent driving and cross-regional claims). Based on this core claims data, it further generates implementable claims strategies, clarifies the claims process, judgment standards, and implementation details for different case types, and then stores the core claims data and corresponding claims strategies in a specialized knowledge base after expert review and confirmation. The knowledge base agent (Agentic RAG) is also updated synchronously to ensure that the knowledge base agent can obtain the latest knowledge corresponding to the terms in real time and accurately recall relevant rules and strategies in subsequent case processing.

[0047] For example, in the auto insurance sector, it effectively addresses the pain points of traditional claims systems, such as lagging knowledge updates and difficulty in adapting to new scenarios. The error-understanding agent analyzes historical errors to optimize damage assessment standards and risk scoring strategies. For instance, it corrects errors in damage assessment caused by misidentification of accident photos, improves the rules for assigning responsibility between automakers and drivers in cases involving autonomous driving, and reduces the recurrence of similar errors. Furthermore, the framework deconstruction of insurance terms and the generation of compensation strategies enable rapid response to term updates. For example, after adding a new clause for compensation for damage to new energy vehicle batteries, the system can quickly extract core data, formulate compensation strategies, and update the knowledge base agent, ensuring that claims decisions always align with the latest policy requirements and enhancing the system's adaptability to complex scenarios and policy changes.

[0048] In the fintech business, it can be widely adapted to scenarios such as optimizing credit approval rules, updating insurance underwriting standards, and iterating anti-fraud strategies. Taking credit business as an example, the error understanding agent analyzes historical credit approval errors, can identify error types such as "inaccurate verification of income certificates" and "misinterpretation of credit records," generate optimization solutions (such as strengthening the cross-verification rules of bank statements and income certificates), and incorporate them into the specialized knowledge base; when credit policies, regulatory requirements, and other "insurance clause-type" documents are updated, AI decomposes and extracts core data (such as adjusting the debt-to-income ratio threshold and expanding the scope of inclusive finance credit support), generates corresponding credit approval strategies, updates the knowledge base intelligence agent, and ensures that credit decisions are in line with the latest policies and risk control requirements; in the insurance underwriting scenario, it can optimize underwriting rules and improve underwriting efficiency and accuracy by quickly responding to changes such as updates to health insurance clauses and adjustments to the definition of critical illness, helping fintech businesses achieve dynamic evolution of the knowledge system and continuous strengthening of risk control.

[0049] In one embodiment, step S300 further includes: S3021. Obtain de-identified information through a responsible intelligent agent; S3022. Determine a preliminary responsibility allocation strategy based on the de-identified information; S3023. Real-time detection of the local state of the intelligent agent, and comparison of the local state with a dynamic threshold; S3024. When the local state is greater than the dynamic threshold, the preliminary responsibility division strategy is encrypted. S3025. The responsibility assignment agent transmits the encrypted preliminary responsibility assignment strategy to the reasoning agent through an adaptive event-triggered control protocol. S3026. The reasoning agent performs causal reasoning and risk identification based on the loss assessment standards, risk scoring strategies, structured data, and encrypted preliminary liability division strategies, and outputs the compensation amount and risk indicators.

[0050] In this embodiment, when the Inference Agent performs causal reasoning and risk identification, for complex scenarios such as intelligent driving, it needs to incorporate the liability allocation information provided by the Liability Determination Agent (an intelligent agent module specifically adapted for the liability division of intelligent driving accidents), and ensure the security and effectiveness of data transmission throughout the process. First, the Liability Determination Agent obtains anonymized information provided by the automaker through a secure interface (referring to data that has undergone anonymization processing, removing privacy and commercially sensitive content while retaining core valid information, such as the activation status of the autonomous driving system, sensor operating data, and decision logs), avoiding privacy and commercial interest risks caused by the leakage of raw data. Subsequently, the Liability Determination Agent, in conjunction with industry standards such as the "Technical Specification for Determining Insurance Compensation for Traffic Accidents of Intelligent Connected Vehicles," interacts with the human-machine co-driving liability determination model, comprehensively analyzes the causes of the accident based on the anonymized information, determines the preliminary liability allocation strategy for relevant parties such as the driver, automaker, and algorithm provider, and clarifies the liability ratio or attribution of each party.

[0051] Simultaneously, the system monitors the local state of each agent (including the responsible agent and the reasoning agent) in real time. Here, "local state" refers to key indicators such as the agent's current data synchronization level and operational stability. The dynamic threshold is a critical value that is adjusted in real time based on factors such as the network environment and data importance, used to determine whether data interaction is necessary. When the error in the agent's local state exceeds the dynamic threshold, it indicates insufficient data synchronization or a security risk, requiring encryption of the initial responsibility allocation strategy (using encryption technology to protect the data and prevent tampering or theft during transmission). If the local state error does not exceed the dynamic threshold, transmission can proceed according to the normal process, but encryption is the default security mechanism to ensure the security of core data.

[0052] The liability determination agent transmits the encrypted preliminary liability allocation strategy to the inference agent via an Adaptive Event-Triggered Control Protocol (an ad hoc communication protocol that exchanges data only when preset conditions are met, rather than real-time continuous transmission). This protocol avoids the waste of network bandwidth and computing resources caused by continuous communication and uses encryption technology to resist network threats such as denial-of-service attacks and false information injection attacks, ensuring the security of the collaborative process. Finally, the inference agent integrates the damage assessment standards, risk scoring strategies, structured data (such as accident details and damage assessment amounts), and the encrypted preliminary liability allocation strategy, and calls a large model (such as qwen-plus) to perform deep causal reasoning. For example, it combines the liability allocation results to determine the responsible party for compensation, calculates the compensation amount based on the damage assessment standards and liability ratios, and identifies potential claims risks caused by ambiguous liability allocation and disputes involving multiple parties. Finally, it outputs the compensation amount and risk indicators, including the basis for liability determination, compensation logic, and details of risk points.

[0053] For example, in the auto insurance field, after obtaining anonymized data from automakers, the liability-determining intelligent agent accurately classifies the responsibilities of each party in "human-machine co-driving" scenarios. Through adaptive event-triggered control protocols and encryption processing, the liability allocation strategy is securely and efficiently transmitted to the reasoning intelligent agent. The reasoning intelligent agent combines this strategy with damage assessment standards and structured data to quickly complete claims logic reasoning. For example, it clarifies the proportion of compensation that automakers should bear in accidents caused by autonomous driving system malfunctions, identifies the risk of claims delays that may result from liability disputes, significantly improves the efficiency and accuracy of claims for intelligent driving accidents, and simultaneously protects automakers' data privacy and the security of the claims process.

[0054] In the fintech sector, this approach can be widely applied to risk assessment and liability allocation in multi-party collaborative projects, such as joint loan approval and cross-border financial transaction dispute resolution. Taking joint loans as an example, the liability-determining agent can obtain anonymized business data (such as credit limits and risk control records) from each cooperating financial institution. Combining this data with cooperation agreements and industry regulatory rules, it initially determines the liability ratio of each party in risk events such as loan delinquency and bad debts. After state detection and encryption, the liability allocation strategy is transmitted to the inference agent via an adaptive event-triggered control protocol. The inference agent integrates credit policy standards, risk scoring strategies, borrower structured data (such as credit records and proof of repayment ability), and liability allocation strategies to infer loan approval results and risk-sharing schemes. This identifies bad debt risks caused by unclear responsibilities among multiple parties, ensuring the compliance and security of joint loan operations. In cross-border financial transactions, by assigning responsibilities among parties in exchange rate fluctuations and compliance reviews, it improves the efficiency of transaction dispute resolution while protecting cross-border transaction data security and privacy.

[0055] In one embodiment, a multi-agent-based claims decision-making device is provided, which corresponds one-to-one with the multi-agent-based claims decision-making method described in the above embodiments. (Refer to...) Figure 4 , Figure 4 This is a schematic diagram of the functional modules of a preferred embodiment of the multi-agent claims decision-making device of the present invention. The modules include a data processing module 10, a data verification module 20, a claims reasoning module 30, and a claims decision-making module 40. Detailed descriptions of each functional module are as follows: The data processing module 10 is used to acquire multimodal data and preprocess the multimodal data to generate structured data; The data verification module 20 is used to input the structured data into the verification agent for verification. When the structured data passes the verification, the structured data is transmitted to the reasoning agent. The claims reasoning module 30 is used by the reasoning agent to perform causal reasoning and risk identification based on the loss assessment standards, risk scoring strategies and structured data, and output the compensation amount and risk indicators. The claims decision module 40 is used to input the compensation amount and risk indicators into the verification agent for review and quantification, output a quantitative risk score, and generate a claims decision report based on the quantitative risk score.

[0056] In one embodiment, the data processing module 10 includes: The data acquisition unit is used to acquire multimodal data uploaded by the user, including text information data and image information data; Data transmission unit, used to transmit the multimodal data to the material collection agent; The data processing unit is used to call the optical character recognition engine and the multimodal large model through the material collection agent, and combine the material analysis rule template to fuse and extract the multimodal data to generate structured data.

[0057] In one embodiment, the data verification module 20 includes: The data transmission unit is used to input the structured data output by the material collection agent into the verification agent; A dual verification unit is used to perform dual verification on the structured data by the verification agent based on hard constraint rules and soft constraint rules; The inconsistency verification unit is used to remind the user to supplement or correct the structured data when the verification result indicates missing information or inconsistency of key information; The consistency verification unit is used to transmit the structured data to the reasoning agent when the verification result shows that the fields are complete and the key information is consistent.

[0058] In one embodiment, the claims reasoning module 30 includes: The query request unit is used to initiate a query request to the knowledge base agent when the reasoning agent receives structured data; The rule invocation unit is used by the knowledge base agent to obtain the damage assessment standards and risk scoring strategies of the current case in the knowledge base according to the query request using a multi-level retrieval strategy, and to transmit the damage assessment standards and risk scoring strategies to the reasoning agent; The compensation amount and risk indicator unit is used by the reasoning agent to perform causal reasoning and risk identification based on the loss assessment standard, risk scoring strategy and structured data, and output the compensation amount and risk indicator.

[0059] In one embodiment, the claims decision module 40 includes: The secondary review unit is used to input the compensation amount and risk indicators into the verification agent, and review the compensation amount and risk indicators in combination with the loss assessment standard and risk scoring strategy; The risk assessment unit is used to input the approved compensation amount and risk indicators into the operations research risk scoring model. It quantifies the compensation amount and risk indicators by solving a constrained nonlinear integer programming problem and outputs a quantitative risk score. The claims decision-making unit is used to classify risk levels based on the quantitative risk score using a knowledge base, and to generate claims decision and reasoning reports based on the risk levels.

[0060] In one embodiment, the claims decision-making unit further includes: Pre-build a wrong case understanding agent; Obtain historical claims data and input the historical claims data into the Wrongful Claims Understanding Agent for information extraction to obtain historical wrongful claims; The historical wrong claims are analyzed to generate error types, and modification plans are generated based on the error types. The error types and corresponding modification plans are then stored in the expert database. Obtain the preset insurance terms and break them down into their framework to extract core claims data; A compensation strategy is generated based on the core compensation data, the core compensation data and the corresponding compensation strategy are stored in the knowledge base, and the knowledge base agent is updated based on the core compensation data and the corresponding compensation strategy.

[0061] In one embodiment, the claims reasoning module 30 further includes: The de-identified information unit is used to obtain de-identified information through a responsible intelligent agent; The responsibility allocation unit is used to determine a preliminary responsibility allocation strategy based on the de-identified information. A state comparison unit is used to detect the local state of the agent in real time and compare the local state with a dynamic threshold. A data encryption unit is used to encrypt the preliminary responsibility allocation strategy when the local state is greater than the dynamic threshold. An encrypted transmission unit is used to transmit the encrypted preliminary responsibility allocation strategy to the reasoning agent via an adaptive event-triggered control protocol. The encrypted reasoning unit is used by the reasoning agent to perform causal reasoning and risk identification based on loss assessment standards, risk scoring strategies, structured data, and encrypted preliminary liability allocation strategies, and outputs the compensation amount and risk indicators.

[0062] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used for communication with external user terminals via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a multi-agent-based claims decision-making method on the server side.

[0063] In one embodiment, a computer device is provided, which may be a user terminal, and its internal structure diagram may be as follows: Figure 6 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a multi-agent-based claims decision-making method on the user side. In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Acquire multimodal data and preprocess the multimodal data to generate structured data; The structured data is input into the verification agent for verification. When the structured data passes the verification, the structured data is transmitted to the reasoning agent. The reasoning agent performs causal reasoning and risk identification based on loss assessment standards, risk scoring strategies, and structured data, and outputs the compensation amount and risk indicators. The compensation amount and risk indicators are input into the verification agent for review and quantification, a quantitative risk score is output, and a claims decision report is generated based on the quantitative risk score.

[0064] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Acquire multimodal data and preprocess the multimodal data to generate structured data; The structured data is input into the verification agent for verification. When the structured data passes the verification, the structured data is transmitted to the reasoning agent. The reasoning agent performs causal reasoning and risk identification based on loss assessment standards, risk scoring strategies, and structured data, and outputs the compensation amount and risk indicators. The compensation amount and risk indicators are input into the verification agent for review and quantification, a quantitative risk score is output, and a claims decision report is generated based on the quantitative risk score.

[0065] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and user side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0066] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0067] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0068] It should be noted that if any software tools or components not belonging to this company appear in the embodiments of this application, they are merely illustrative examples and do not represent actual use. The embodiments described above are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A claims decision-making method based on multi-agent systems, characterized in that, Includes the following steps: Acquire multimodal data and preprocess the multimodal data to generate structured data; The structured data is input into the verification agent for verification. When the structured data passes the verification, the structured data is transmitted to the reasoning agent. The reasoning agent performs causal reasoning and risk identification based on loss assessment standards, risk scoring strategies, and structured data, and outputs the compensation amount and risk indicators. The compensation amount and risk indicators are input into the verification agent for review and quantification, a quantitative risk score is output, and a claims decision report is generated based on the quantitative risk score.

2. The multi-agent-based claims decision-making method as described in claim 1, characterized in that, The process of acquiring multimodal data and preprocessing the multimodal data to generate structured data includes: Acquire multimodal data uploaded by users, including text information data and image information data; The multimodal data is transmitted to the material collection agent; The material collection agent invokes an optical character recognition engine and a multimodal large model, and combines the material analysis rule template to fuse and extract the multimodal data, generating structured data.

3. The multi-agent-based claims decision-making method as described in claim 1, characterized in that, The step of inputting the structured data into the verification agent for verification, and transmitting the structured data to the inference agent when the structured data passes the verification, includes: The structured data output by the material collection agent is input into the verification agent; The verification agent performs dual verification on the structured data based on both hard and soft constraint rules. When the verification result indicates missing information or inconsistencies in key information, the material collection agent will remind the user to supplement or correct the structured data. When the verification result shows that the fields are complete and the key information is consistent, the structured data is transmitted to the reasoning agent.

4. The multi-agent-based claims decision-making method as described in claim 1, characterized in that, The reasoning agent performs causal reasoning and risk identification based on loss assessment standards, risk scoring strategies, and structured data, and outputs the compensation amount and risk indicators, including: When the reasoning agent receives structured data, it initiates a query request to the knowledge base agent; The knowledge base agent uses a multi-level retrieval strategy to obtain the damage assessment standards and risk scoring strategies for the current case in the knowledge base according to the query request, and then transmits the damage assessment standards and risk scoring strategies to the reasoning agent. The reasoning agent performs causal reasoning and risk identification based on loss assessment standards, risk scoring strategies, and structured data, and outputs the compensation amount and risk indicators.

5. The multi-agent-based claims decision-making method as described in claim 1, characterized in that, The process involves inputting the compensation amount and risk indicators into the verification agent for review and quantification, outputting a quantitative risk score, and generating a claims decision report based on the quantitative risk score, including: The compensation amount and risk indicators are input into the verification agent, and the compensation amount and risk indicators are reviewed in conjunction with the loss assessment standards and risk scoring strategies. The approved compensation amount and risk indicators are input into the operations research risk scoring model. The compensation amount and risk indicators are quantified by solving a constrained nonlinear integer programming problem, and a quantitative risk score is output. Based on the expert knowledge base, risk levels are divided according to the quantitative risk score, and claims decision and reasoning reports are generated based on the risk levels.

6. The multi-agent-based claims decision-making method as described in claim 1, characterized in that, The method of classifying risk levels based on the expert knowledge base and the quantitative risk score, and generating claims decision and reasoning reports based on the risk levels, also includes: Pre-build a wrong case understanding agent; Obtain historical claims data and input the historical claims data into the Wrongful Claims Understanding Agent for information extraction to obtain historical wrongful claims; The historical wrong claims are analyzed to generate error types, and modification plans are generated based on the error types. The error types and corresponding modification plans are then stored in the expert database. Obtain the preset insurance terms and break them down into their framework to extract core claims data; A compensation strategy is generated based on the core compensation data, the core compensation data and the corresponding compensation strategy are stored in the knowledge base, and the knowledge base agent is updated based on the core compensation data and the corresponding compensation strategy.

7. The multi-agent-based claims decision-making method as described in claim 1, characterized in that, The process of the reasoning agent performing causal reasoning and risk identification based on loss assessment standards, risk scoring strategies, and structured data, and outputting compensation amounts and risk indicators, also includes: Obtain de-identified information through a responsible intelligent agent; A preliminary responsibility allocation strategy is determined based on the anonymized information; The local state of the agent is detected in real time, and the local state is compared with a dynamic threshold. When the local state is greater than the dynamic threshold, the preliminary responsibility allocation strategy is encrypted. The assigning agent transmits the encrypted preliminary responsibility allocation strategy to the reasoning agent via an adaptive event-triggered control protocol. The reasoning agent performs causal reasoning and risk identification based on loss assessment standards, risk scoring strategies, structured data, and encrypted preliminary liability allocation strategies, and outputs the compensation amount and risk indicators.

8. A claims decision-making device based on multi-agent systems, characterized in that, The multi-agent-based claims decision-making device includes: The data processing module is used to acquire multimodal data and preprocess the multimodal data to generate structured data; The data verification module is used to input the structured data into the verification agent for verification. When the structured data passes the verification, the structured data is transmitted to the reasoning agent. The claims reasoning module is used by the reasoning agent to perform causal reasoning and risk identification based on loss assessment standards, risk scoring strategies and structured data, and output the compensation amount and risk indicators. The claims decision module is used to input the compensation amount and risk indicators into the verification agent for review and quantification, output a quantitative risk score, and generate a claims decision report based on the quantitative risk score.

9. A computer device, characterized in that, The computer device includes a memory, a processor, and a multi-agent-based claims decision-making program stored in the memory and executable on the processor, wherein the multi-agent-based claims decision-making program, when executed by the processor, implements the steps of the multi-agent-based claims decision-making method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a multi-agent-based claims decision-making program, which, when executed by a processor, implements the steps of the multi-agent-based claims decision-making method as described in any one of claims 1-7.

Images