End-to-end insurance service processing agent system based on AI voice

Abstract

The invention discloses an end-to-end insurance service processing agent system based on AI voice, and relates to an end-to-end insurance service processing method based on AI voice and an agent system using the method. The agent system identifies the intention of the insurance user, performs multiple rounds of dialogue memory management, and performs context analysis and completion on the current intention; matching the complemented user intention with a configurable business process engine; calling a corresponding external system service to complete business operation, and generating a text response; the text response is synthesized into the voice with the preset timbre through a voice copying model, the voice with the preset timbre is fed back to the user in real time, the real-time voice of the insurance user is recognized, the voice of the voice intelligent agent verbal skill is generated and copied in real time, and an end-to-end real-time collecting and processing closed loop is formed.

Description

AI-powered voice-based end-to-end insurance business processing intelligent agent system Technical Field

[0001] This invention relates to the field of insurance business technology, specifically to an end-to-end intelligent agent system for insurance business processing based on AI voice. Background Technology

[0002] Due to the specialized nature of the insurance industry, insurance companies often experience misunderstandings when serving customers. Although past efforts have focused on improving the generalization of technical terminology and enhancing product system interactivity, the quality of service and the efficiency of resolving customer issues have not been particularly high.

[0003] Existing technologies mainly use a page + technical interface approach to implement specific functions that meet user needs. Many page functions require users to spend a lot of effort to figure out without guidance. At the same time, the page + technical interface interaction method is much inferior to real face-to-face scenario-based collaborative problem-solving in terms of both efficiency and product experience.

[0004] In the traditional model, the primary way users and insurance companies handle business is through a silent process where users navigate to menus on the official application. Insurance company staff cannot intervene in users' actions in real time, and user questions or difficulties are only discovered through various system monitoring metrics. This has led to several long-standing problems: 1. As the company's application becomes more comprehensive, the menu options appear abundant, but in reality, users often only want to solve a specific problem and cannot find the relevant menu; 2. Due to the specialized nature of their business, insurance companies use a lot of technical jargon and terminology, which users may not possess the necessary expertise for, resulting in a lack of empathy and even complaints.

[0005] Based on the above, this invention creates an AI-based method for constructing an end-to-end voice-assisted system that enables panoramic interactive communication between insurance companies and users to meet user business needs. Summary of the Invention

[0006] The purpose of this invention is to provide an AI-based voice-based end-to-end insurance business processing intelligent agent system to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, this invention provides the following technical solution: an AI-based end-to-end insurance business processing intelligent agent system, comprising an AI-based end-to-end insurance business processing method and an AI-based end-to-end insurance business processing intelligent agent system using this method. The method of this invention is implemented through the following sequentially connected and closely coordinated technical modules, comprehensively covering the entire process from interaction to execution: Real-time, high-fidelity voice interaction layer 1. This layer is the direct interface between the intelligent agent system and the user, responsible for bidirectional high-precision information conversion. It integrates a high-accuracy automatic speech recognition engine, converting the user's random, conversational speech stream into text in real time. Its key innovation lies in the integration of a customized voice replication model. This model is trained using exclusive timbre data provided by the insurance company, enabling it to synthesize all text feedback generated by the intelligent agent system (including standard scripts and dynamically generated business content) into a voice with a specific brand identity, natural timbre, and rich emotion, thereby providing a highly humanized, professional, and consistent service experience, breaking through the mechanical feel of traditional TTS technology.

[0008] 2. Accurate and efficient insurance intent understanding core: This core, as the "cognitive brain" of the intelligent agent system, is specifically designed to resolve the contradiction between the diversity of user expression and the rigor of business requirements in insurance scenarios.

[0009] Corpus Engineering Innovation: A dedicated corpus fine-tuning method based on small models was adopted. Specifically, Python and Pandas were used to perform large-scale, automated cleaning and sorting of historical customer service dialogues (recorded in Excel format), building a high-quality corpus of over 20,000 entries. The core innovation lies in reconstructing the corpus using a ShareGPT-style approach, pairing dozens of different user natural language questions expressing the same business intent with a unified, structured standard intent label and slot.

[0010] Model training strategy: The aforementioned corpus is divided into training and test sets in an 8:2 ratio. A lightweight basic language model is supervised and fine-tuned using the training set, and a pre-trained basic language model is also supervised and fine-tuned using the training set. Knowledge distillation is performed on the collected question-and-answer data, and the resulting corpus is then used for supervised and fine-tuned on a smaller model. This reduces deployment costs and improves fine-tuning efficiency, resulting in the insurance customer intent understanding model. The obtained model has a small parameter size and fast inference speed, but significantly improved accuracy in recognizing insurance intent. This "multiple questions to one intent" training method enables the model to deeply understand the semantic essence, rather than simply matching keywords, thus achieving both high accuracy and high response efficiency.

[0011] 3. Intelligent and robust multi-turn dialogue management layer: This layer is the "contextual memory center" of the intelligent agent system, ensuring that user needs are not lost in complex multi-turn interactions and gracefully handling topic jumps and interruptions.

[0012] The memory model mechanism stores the complete dialogue text sequence in real time and sets up a high-speed cache to optimize the access performance of the most recent rounds of dialogue, balancing data integrity and retrieval speed.

[0013] Dialogue Summary Agent: This is the core of intelligent dialogue management. It proactively analyzes cached historical dialogues, not simply retelling them, but performing business-oriented contextual analysis and summary generation. It can extract key business entities (including policy numbers and customer names), summarize clearly defined user needs, mark pending tasks, and organize them in a structured format. When a dialogue is interrupted, it can maintain independent summary blocks for each active subtopic, ensuring clear context isolation and seamless transitions.

[0014] 4. Flexible and programmable business agent execution layer: This layer is the "decision-making and execution arm" of the agent system, which translates the understood intent into real business operations.

[0015] Configurable Business Process Engine: Employing a modular and configurable design philosophy, each insurance business (including but not limited to underwriting, claims, and policy maintenance) is broken down into a series of standardized business processing units. Each unit is configured with a unique step code and a clearly defined instruction type: ask (query, requiring voice interaction) or act (execution, requiring a call to backend services).

[0016] Precise matching and chain-driven operation: The structured results (intent tags and slots) of the core intent understanding are mapped to the corresponding business processing unit codes, thereby triggering specific business chains. The engine drives the units to execute according to a preset logical sequence, forming an automated workflow.

[0017] Service Integration and Interruption Recovery: For ACT-type units, standardized interfaces such as the Model Context Protocol (MCP) are used to directly call core business systems and claims engines to complete actual business operations. The intelligent agent system has a built-in business data storage and job switching mechanism. When the dialogue management layer detects a topic interruption, the current process state is fully saved, and the intelligent agent system smoothly switches to handle the new topic. Once the temporary topic ends, the original process can be accurately restored to continue execution, ensuring business continuity and robustness in complex interaction scenarios.

[0018] 5. Formation of an End-to-End Closed Loop: The aforementioned technologies are interconnected, forming a highly automated intelligent closed loop: User voice input → Real-time text conversion → Dedicated small model accurately extracts structured intent → Dialogue summarization agent combines history to complete and disambiguate intent → Orchestration engine matches and drives business chain execution → Substantive business processing is completed by calling backend services through standard interfaces → Result text is generated → Customized voice replication model synthesizes branded voice tone → Real-time feedback to the user. The entire process achieves an intelligent business processing experience of "what you say is what you get".

[0019] Compared with existing technologies, the beneficial effects of this invention are: This invention is a general product construction method based on AI. Through ASR and TTS technologies, it forms an end-to-end real-time acquisition and processing closed loop by recognizing the real-time voice of insurance users and generating and replicating the speech of the voice agent.

[0020] During business dialogue processing, a set of tools and methods for generating user dialogue corpora in the insurance industry were developed. Based on the corpus, a proprietary small model for understanding insurance customer intent was derived through SFT, thereby quickly understanding the user's questions. It is connected to a specially built business processing intelligent agent. Through the combination of the intelligent agent and the MCP interface, the user's intent can be quickly identified and various system services can be called for automatic processing. Attached Figure Description

[0021] Figure 1 is a flowchart of an end-to-end insurance business processing method based on AI voice.

[0022] Figure 2 is a system block diagram of an end-to-end insurance business processing intelligent agent system based on AI voice. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Please refer to Figures 1 and 2. In this embodiment of the invention, the AI ​​voice-based end-to-end insurance business processing intelligent agent system includes an AI voice-based end-to-end insurance business processing method and an AI voice-based end-to-end insurance business processing intelligent agent system using this method. This embodiment focuses on the overall operation of the intelligent agent system, illustrating how each technical module works collaboratively to form a complete, coherent, and intelligent business processing closed loop.

[0025] 1. Intelligent Agent System Initialization and Voice Interaction Startup: After the intelligent agent system starts up, each module completes its initialization. The voice interface module loads the automatic speech recognition engine and the speech synthesis engine. The core of the speech synthesis engine is a speech replication model, which has been pre-trained using target timbre audio data provided by the insurance company through a neural network. This model can synthesize human-like speech with a specific timbre that conforms to insurance service specifications. When a user initiates a voice call or presses the voice button through a terminal device (including but not limited to a mobile app or smartphone), the intelligent agent system enters real-time voice interaction mode.

[0026] 2. First processing stage: speech-to-text and preliminary intent understanding. The user begins to speak, and their speech stream is captured in real time.

[0027] Speech Recognition: The automatic speech recognition unit in the voice interface module works immediately, converting continuous speech signals into corresponding text content. This process is real-time, ensuring a low-latency interaction.

[0028] Text Input and Initial Intent Extraction: The converted dialogue text is fed into the intent understanding module. The core of this module is the insurance customer intent understanding model.

[0029] The model was built following a corpus fine-tuning method for small models. Before deployment of the agent system, the following steps were completed for training: 1. Corpus engineering: Using Python and Pandas scripts, a large number of Excel-formatted dialogue records were exported from the historical customer service system and cleaned to remove irrelevant information and noise.

[0030] 2. Data Reconstruction: Adopting the ShareGPT style, hundreds or thousands of different user questions expressing the same business intent (including but not limited to "policy inquiry") are paired with standardized tags of the intent to form a standardized corpus pair of "multiple questions to one intent", and divided into training set and test set in an 8:2 ratio.

[0031] 3. Model fine-tuning: Using the training set, a lightweight base language model is fine-tuned under supervision to obtain a specialized model with a small number of parameters, fast inference speed, and high accuracy in intent recognition for the insurance domain.

[0032] At this moment, the model receives the current user's dialogue text, and based on the knowledge it has trained, it quickly analyzes and outputs a structured intent recognition result, which includes standardized intent labels and related semantic slot information.

[0033] 3. Second processing stage: Dialogue context management and intent completion. The dialogue management module is activated at the same time as or shortly after the intent recognition model is working.

[0034] Historical storage: The memory model of this module begins to work. It associates and stores the original text input by the user in the current round with the subsequent system responses in real time, on a round-by-round basis. Simultaneously, to ensure efficient retrieval of historical information in multi-round dialogues, the system establishes a high-speed cache for the current session, prioritizing the storage of the most recent N rounds of dialogue content.

[0035] Context summarization generation: To understand the position of the current utterance within the overall dialogue, the dialogue summarization agent is initiated. It reads historical dialogues from the cache (which may be empty if it's the first round), not simply listing them, but performing contextual discrimination.

[0036] Its specific tasks are: to identify and extract key insurance business entities appearing in the dialogue (including but not limited to customer names, ID numbers, and mentioned policy numbers), to summarize clearly defined user requests, and to mark the pending tasks proposed by the system (including but not limited to "requires supplementary materials" and "requires confirmation of certain materials"). Then, it organizes the above information into a structured, condensed dialogue summary.

[0037] Final determination of intent: The initial intent recognition result, along with the generated structured dialogue summary, is fed into a decision point (including but not limited to secondary analysis of the intent model itself or a simple logical layer). The dialogue summary provides key background information, thereby completing, disambiguating, or confirming the current intent, forming the final user intent that takes into account the context.

[0038] 4. Third processing stage: Intelligent business process matching and execution. The structured user intent, after being completed and confirmed, is sent to the business intelligence module.

[0039] Process Matching: The core of this module is a configurable business process engine. During system deployment, this engine, through a process configuration management interface, decomposes and configures various insurance businesses (including but not limited to insurance application, claims, changes, and inquiries) into a series of business processing units in a low-code or graphical manner. Each unit has a unique step code and a defined instruction type (`ask` or `act`). The engine matches the received user intent (including tags and slots) with predefined rules, mapping it to one or more specific step codes, thereby activating the corresponding business processing chain.

[0040] Chain-driven: If the instruction type of the currently active unit is `ask`, the engine generates the query script text according to the unit's configuration. This text is directly passed to the speech interface module, ready for speech synthesis.

[0041] If the instruction type is `act`, the engine, through the service integration module, calls the corresponding external business system services (including but not limited to core business systems, claims calculation engines, and databases) according to predefined model context protocols and other interface specifications. After the call is completed, the engine receives the business result data returned by the service.

[0042] State Management and Interruption Handling: The engine continuously manages the state and data of the process while executing the business chain. The intelligent agent system has a built-in mechanism for temporary transfer and recovery of business workflows. When the dialogue management module detects that a new intent input by the user does not match the main goal of the currently executing process (for example, a user suddenly inquires about a new product during a claims process), it sends an interruption signal to the business engine. The engine then pauses the current chain, temporarily stores the complete execution state (including but not limited to the current step and collected data), and then matches and switches to a new process chain to handle the temporary topic based on the new intent. Once the temporary topic process is completed, the intelligent agent system can trigger a recovery operation through user guidance or automatic logic, allowing the engine to accurately continue executing the original business process from the temporary point.

[0043] 5. Fourth processing stage: Result generation and humanized voice feedback. When the business process engine reaches the node where feedback to the user is required, it will generate the final text response.

[0044] Speech Synthesis: This text response is sent to the speech synthesis unit of the speech interface module. This unit invokes the loaded speech replication model.

[0045] Voice replication: Based on the insurance company's proprietary "customer service voice," this model synthesizes the input text into natural, fluent, and expressive speech. This process is not a mechanical reading of text, but rather imbues the voice with emotional color and professional friendliness, achieving the interactive effect of a voice-based digital human.

[0046] Real-time feedback: The synthesized speech stream is played back to the user in real time through an audio output device.

[0047] 6. Closed-Loop Cycle and Continuous Optimization: After hearing the response from the intelligent agent system, the user may initiate another round of dialogue. The intelligent agent system then returns to step 2, starting a new cycle of "speech recognition - intent understanding (combined with historical summaries) - business execution - voice feedback." The dialogue data, intent recognition results, and business execution logs generated throughout this process can all serve as feedback data, used for periodic iterative fine-tuning and optimization of the insurance customer intent understanding model and the dialogue summary intelligent agent, making the intelligent agent system increasingly intelligent with use.

[0048] The working principle of this invention is as follows: 1. Construction of insurance user intent recognition. Insurance users' communication styles are diverse, and given insurance regulations, insurance companies need to maintain a serious approach when generating communication styles. In this process, a corpus fine-tuning method based on a small model is developed, balancing the accuracy and efficiency of intent recognition. The entire intent model construction steps are as follows: First, Python and Pandas are used to organize the customer service corpus collected from historical Excel files to improve corpus quality. A script is used to segment the corpus, dividing the overall corpus into an 80% training set and a 20% test set.

[0049] The overall corpus style adopted the ShareGPT method, with a total of more than 20,000 corpora. The training and fine-tuning this time adopted the method of "multiple similar questions corresponding to a unified question", which is different from the previous training and fine-tuning methods, thus making the model's intent recognition more accurate.

[0050] 2. Memory Model. During user collaboration with the insurance company, many unexpected situations may arise, such as "temporary interruptions" or "users changing the subject." A multi-turn dialogue memory model is constructed for analyzing multi-turn conversations between users and the insurance company, ensuring that user requests are not overlooked and that user actions are not mishandled. In the multi-turn dialogue process, user history is stored and retrieved in real time. To improve access speed, a certain range of historical dialogues is cached. The retrieved historical dialogues are analyzed using a dialogue summary agent to determine context and identify new intentions. 3. Business Workflow. Based on interactive business processing, a business workflow is constructed, building an automated business workflow chain through business configuration. Simultaneously, interactive business processing may encounter interruptions and other situations; temporary topic shifts are handled through business data storage and workflow switching. A business process engine, designed with independent business units, is implemented during user dialogue.

[0051] By configuring each business processing step in a modular fashion, a complete configuration engine is formed, with each business processing unit configured in the manner described above. When the user's intent is recognized to a specific business unit, the AI ​​model and the engine will match them to the corresponding code and find the process in the specific process engine. stepCode represents the recognized step, and instruct is divided into forms such as ask and act, where act is used for the execution of specific MCP services.

[0052] The content from ask will be processed into audio reproduction, generating speech terms and pronunciations, which will then be pushed to the front end.

[0053] 4. Voice Replication. During the conversation, in addition to accurately recognizing the user's voice, the insurance company's responses must also be of a high standard in terms of real-time performance, accuracy, and handling attitude. The entire product and methodology, based on a voice replication model and supplemented by its own voice timbre, forms a digital voice avatar specifically responsible for text-to-speech replication.

[0054] This invention applies to the health insurance industry to achieve user business needs through panoramic interactive end-to-end AI voice assistance. In this new model, through panoramic interactive end-to-end real-time voice dialogue, insurance companies can quickly identify user needs and provide services. The backend translates the identified customer needs into actual code execution and simultaneously uses voice replication technology to push specified scripts to the user, achieving a "face-to-face" business processing. This model offers the following advantages: 1. Reduced user service window: Previously, users needed to spend time finding service windows; now, a single dialogue window can resolve all business needs. 2. Improved efficiency: Reduced manual input by users; accelerated customer service processing efficiency through big data and voice technology; and atomic processing capabilities for business processes centered on the MCP (Multi-Channel Processing).

[0055] The technology of this invention integrates the corpus fine-tuning of AI large models, speech recognition and speech replication, and multiple agent technologies to form a complete set of general construction methods.

[0056] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An end-to-end insurance business processing method based on AI voice, characterized in that, Includes the following steps: S1. Real-time voice interaction: Converts real-time voice streams input by users into conversational text through automatic speech recognition technology; S2. Insurance User Intent Recognition: The dialogue text is input into the insurance customer intent understanding model, which is a small-parameter model obtained through supervised fine-tuning based on insurance domain corpus, used to quickly and accurately identify the user's insurance business intent; S3. Multi-turn dialogue memory management: The user's historical dialogue information is stored and managed through a memory model; when the intent understanding model identifies the current intent, the dialogue summarization agent performs contextual analysis and completion of the current intent based on the historical dialogue information in the memory model; S4. Intelligent execution of business workflow: Match the completed user intent with a configurable business process engine; the business process engine is composed of multiple independently configurable business processing units linked together. After a successful match, the business process engine drives the execution of the corresponding business processing chain. S5. Service Invocation and Voice Response: When executing the business processing chain, the corresponding external system service is invoked to complete the business operation and generate a text response; through the voice replication model, the text response is synthesized into a voice with a preset timbre and fed back to the user in real time.

2. The end-to-end insurance business processing method based on AI voice according to claim 1, characterized in that, The method for constructing the insurance customer intent understanding model, namely the corpus fine-tuning method based on the small model, includes the following steps: S2.1, Corpus collection and preprocessing: Using Python and Pandas tools, the Excel-formatted corpus collected from historical customer service dialogues is sorted and cleaned to improve the corpus quality; S2.2, Corpus segmentation and dataset construction: The overall corpus is divided into a training set and a test set using a script, with the training set accounting for 80% and the test set accounting for 20%, forming a dataset containing more than 20,000 high-quality corpus entries; S2.3, Corpus style reconstruction: The corpus is reconstructed using the ShareGPT style, reconstructing multiple similar user questions targeting the same standard intent into normalized corpus pairs mapped to that single standard intent; S2.4, Supervised fine-tuning training: The pre-trained basic language model is supervised fine-tuned using the training set, knowledge distillation is performed on the sorted question-and-answer data, and the obtained corpus is then supervised fine-tuned on the small model to obtain the insurance customer intent understanding model.

3. The method for constructing an insurance customer intent understanding model according to claim 2, characterized in that, In step S2.3, the "reconstruction of multiple similar user questions into normalized corpus pairs mapped to a single standard intent" is specifically achieved by constructing an "instruction-response" format, where the "instruction" is a variety of user natural language questions, and the "response" is structured text containing standardized intent labels, thereby improving the accuracy of the model's intent recognition.

4. The end-to-end insurance business processing method based on AI voice according to claim 1, characterized in that, The memory model in step S3 works in the following ways: S3.1, Real-time storage of historical dialogues: During the multi-turn dialogue between the user and the system, the complete text content of each round of dialogue is stored in real time; S3.2, High-speed caching mechanism: A high-speed cache area is established for the current dialogue session, and the most recent historical dialogue content within a certain range is stored in the cache to improve access speed. S3.3, Context discrimination and summary generation: The dialogue summary agent performs context discrimination on the historical dialogues obtained from the cache and generates a condensed dialogue summary for identifying new user intentions or completing the current intention. The specific method for the dialogue summary agent to generate the condensed dialogue summary includes: identifying key business entities in the dialogue, extracting the clearly defined user requests, marking to-do items, and organizing them in a structured format to form a context for subsequent intention analysis.

5. The end-to-end insurance business processing method based on AI voice according to claim 1, characterized in that, The construction and execution of the business process engine in step S4 includes: S4.1, Business unit configuration: Decomposing the insurance business into multiple independent and configurable business processing units, each unit is configured with a unique step code and defines an instruction type, which includes "ask" for querying user information and "act" for performing system operations; when the business processing unit with instruction type "ask" is executed, the generated query text will be used as input and transmitted to the voice replication model to generate the corresponding voice script, and pushed to the front end for user interaction; when the business processing unit with instruction type "act" is executed, the corresponding external business system service is called through the predefined model context protocol interface to complete the specific business operation; S4.2, Intent and code matching: Mapping the identified user intent to a specific step code, thereby locating and activating the corresponding business processing unit in the engine; S4.3, Chained execution and state management: Based on the configuration order and logic of the business processing units, an automated work chain is formed and driven to execute, while managing the data and state during the business execution process.

6. The end-to-end insurance business processing method based on AI voice according to claim 5, characterized in that, The method also includes a temporary transfer and recovery mechanism for business workflows: when a user initiates a temporary topic or the conversation is interrupted during the execution of a business process, the execution of the current business processing chain is paused, the current business data and status are temporarily stored, and the process is switched to handle the temporary topic. Once the temporary issues are resolved, the execution of the original business processing chain will be resumed from the temporary state.

7. The end-to-end insurance business processing method based on AI voice according to claim 1, characterized in that, The voice replication model is a voice synthesis model that uses a neural network and is trained using target timbre data provided by the insurance company to generate customized voice digital human responses that meet insurance service specifications.

8. An end-to-end intelligent agent system for insurance business processing based on AI voice, characterized in that, To implement the method of any one of claims 1 to 7, the intelligent agent system comprises: a voice interface module for real-time voice input and output, integrating an automatic speech recognition unit and a speech synthesis unit; an intent understanding module with a built-in insurance customer intent understanding model for processing user input text and recognizing business intent; a dialogue management module integrating a memory model and a dialogue summary intelligent agent for maintaining dialogue context, managing historical information, and completing user intent; a business intelligent agent module including a configurable business process engine for matching intent, executing business chains, and calling external services; and a service integration module for connecting to and calling various external business services.

9. The intelligent agent system according to claim 8, characterized in that, The speech synthesis unit in the speech interface module integrates a speech replication model trained according to the method described in claim 7.

10. The intelligent agent system according to claim 8, characterized in that, The intelligent agent system also includes a process configuration management interface, which is used to define, modify and deploy the business processing units and their logical relationships in the business process engine in a low-code or graphical manner.

Images