System and method for improving business through multiple agents

Abstract

The invention provides a system and method for improving business through multiple agents, and relates to the technical field of data analysis, and the system comprises a data layer and a multi-modal data layer for realizing data collection; data processing: preprocessing the data to form a customer portrait; constructing an ontology knowledge base; the vector database vectorizes customer portraits, insurance product information and communication content data; the personalized communication content generated by the application layer is stored, and the personalized communication content is pushed to a contact channel preferred by a customer through an API interface; the AI core layer is used for receiving requests transmitted by different application programs, users or systems; finely adjusting the model through the small model; various atomic energies are provided; and the application layer is used for constructing an intelligent software entity capable of perceiving an environment, making a decision and taking actions by applying a sequence, a dependency relationship and an execution logic of an arrangement definition task. Through multi-agent cooperation, automatic configuration and optimal scheduling of resources are realized, and personalized product and service recommendation is dynamically generated.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to a system and method for improving business through multiple agents. Background Technology

[0002] With the rapid advancement of digital transformation across industries, insurance customer management systems are being upgraded towards personalized and intelligent solutions to unlock the business value potential of existing customers. Traditional IT technologies rely on rule-based recommendations based on customer tags to perform customer data analysis (including policies, interactions, and behaviors), which cannot achieve personalized product and service recommendations or improve policy upgrade conversion rates.

[0003] Moreover, insurance customers seek personalization and emotional resonance, requiring exclusive and warm communication content. Traditional IT technology cannot generate more personalized emotional content; it tends to produce template-based and rigid content output.

[0004] Currently, in response to the dynamic and potential protection needs of customers at different stages of their lives, including personalized service recommendations, we leverage the powerful content creation capabilities of generative AI and multi-agent technology to generate exclusive and warm communication content for customers at key after-sales touchpoints such as birthdays and policy anniversaries. This upgrades standardized customer care into memorable and personalized interactions, thereby strengthening customers' perception of brand value and enhancing their sense of participation and loyalty.

[0005] The existing content generation solutions have the following technical defects: (1) The generation rules are fixed and single: the existing system adopts a static rule engine, which cannot adapt to the diverse business scenario requirements and lacks flexibility and scalability; (2) There is a lack of dynamic response mechanism: the system cannot perceive changes in user behavior, preference shifts and emotional state fluctuations in real time, resulting in the content generation being out of touch with the actual needs of users; (3) The personalized recommendation accuracy is insufficient: traditional collaborative filtering and content matching algorithms have data sparsity and cold start problems, which cannot achieve truly personalized and accurate recommendations; (4) The user operation threshold is high: the existing agent system interface is complex and the configuration is cumbersome, requiring professional technical background to use effectively, which limits the access and application of ordinary users.

[0006] Therefore, how to provide an adaptive content generation system based on multi-agent collaboration that can learn autonomously, adjust dynamically, and make intelligent decisions to dynamically generate personalized product and service recommendations has become an urgent problem to be solved. Summary of the Invention

[0007] To address the aforementioned issues, this invention provides a system and method for improving business through multi-agent collaboration. It aims to achieve automated resource allocation and optimized scheduling through a distributed collaborative mechanism among multiple agents; dynamically generate personalized product and service recommendations based on real-time user profile updates and behavioral analysis; significantly improve system operating efficiency, management intelligence, and user experience, fundamentally overcoming the limitations of existing technical solutions. While ensuring system reliability and cost-effectiveness, it provides an innovative solution for the content generation field.

[0008] To achieve the above objectives, the present invention provides a system for improving business through multiple agents, comprising: a data layer, an AI core layer, and an application layer. The data layer includes modules such as a multimodal data lake, data governance, an ontology knowledge base, and a vector database. The AI ​​core layer includes modules such as a large model service gateway, small models, and atomic capabilities. The application layer includes modules such as application orchestration, an agent framework, and an intelligent computing platform.

[0009] The data layer is used for:

[0010] Data collection is achieved by inputting historical customer data into the multimodal data lake, capturing publicly available customer information, and converting customer service phone conversations into text to obtain the communication content.

[0011] The data governance process preprocesses the collected data to create customer profiles.

[0012] The insurance product ontology, customer and demand ontology, business process ontology, and risk and compliance ontology are constructed in the ontology knowledge base.

[0013] The customer profile, product information in the insurance product entity, and communication content data are vectorized using a vector database.

[0014] The personalized communication content generated by the application layer is stored, and the personalized communication content is pushed to the customer's preferred reach channels through the API interface;

[0015] The AI ​​core layer is used for:

[0016] A unified access service is provided through a large model service gateway, which receives requests from different applications, users or systems and distributes the requests to various model instances through load balancing.

[0017] By fine-tuning pre-trained large or general models using small models, optimized models for specific industries or vertical fields can be constructed.

[0018] IDP provides intelligent document analysis, knowledge graph, speech-to-text, intent recognition, document vectorization, and knowledge recommendation capabilities through atomic capabilities.

[0019] The application layer is used for:

[0020] By applying orchestration to define the sequence, dependencies, and execution logic of tasks, we can develop and build intelligent software entities that can perceive the environment, make decisions, and take actions.

[0021] The intelligent agent framework provides the intelligent software entities with the ability to acquire information, data, and signals from the environment; the ability to identify and invoke external tools, APIs, or services; the ability to arrange the order and dependencies of tasks performed by intelligent software entities; the ability to communicate and collaborate among multiple intelligent software entities; the ability for intelligent software entities to store, retrieve, and manage past experiences, knowledge, and information; and the ability to perform logical reasoning, analysis, and judgment based on acquired information, data, signals, memorized knowledge, and set goals.

[0022] The intelligent computing workbench provides a virtual environment for development, experimentation, and task execution, and offers the ability to store, version control, distribute, and deploy images of computing environments, software configurations, and dependencies. It also provides tools and interface design for building and orchestrating computing or data processing workflows.

[0023] Execute the workflow to generate personalized communication content and send it to customers.

[0024] As a further improvement to the present invention, an infrastructure layer is also included for providing basic cloud computing capabilities.

[0025] As a further improvement of the present invention, the multimodal data lake inputs structured and unstructured data such as customer insurance records, historical interaction data, policy information, past product purchases, service preferences, and communication records through documents and APIs. The policy information includes policy anniversary and birthday. Information of customers on public social media is crawled through web pages, and customer service telephone communication voice is obtained through audio and converted into text to obtain the communication content.

[0026] As a further improvement of the present invention, the insurance product body stores detailed information, coverage, terms, rates, and applicable population of all insurance products;

[0027] The customer and demand ontology stores profile information of customers, including their age, occupation, family structure, income, risk preference, health status, and life stage, and establishes a mapping relationship with potential protection needs.

[0028] The business process entity stores the standard processes, key touchpoints, and communication specifications for insurance marketing and after-sales service.

[0029] The risk and compliance ontology ensures the compliance of all marketing content.

[0030] As a further improvement of the present invention, the data governance preprocesses the massive amount of collected data to remove noise and unify the format. The preprocessing includes data cleaning, date format standardization, and sensitive data desensitization.

[0031] As a further improvement to this invention, a pre-trained large or general model is fine-tuned using a small model to construct an optimized model for a specific industry or vertical field; including:

[0032] Based on the customer life stage identification model, the system automatically identifies the customer's life stage according to data on age, marital status, and children.

[0033] Based on the potential demand forecasting model, combined with the customer's life stage, historical consumption behavior, and risk preference, we predict the customer's potential insurance protection needs.

[0034] Based on the content preference analysis model, we analyze customers' past feedback on different types of content to identify their preferred communication styles and information presentation formats.

[0035] As a further improvement to this invention, atomic capabilities are provided to enable IDP intelligent document analysis, knowledge graph, speech-to-text conversion, intent recognition, document vectorization, and knowledge recommendation, for the following purposes:

[0036] Based on IDP's intelligent classification and analysis capabilities, unstructured documents provided by customers are analyzed, and customer intent is understood based on intent recognition capabilities.

[0037] Leveraging knowledge graph capabilities for complex relationship reasoning, and utilizing knowledge recommendation capabilities to recommend products to customers based on the reasoning results;

[0038] Multimedia data is processed based on speech-to-text capabilities, and vector representations are generated for text and images based on document vectorization capabilities.

[0039] As a further improvement of the present invention, the intelligent framework continuously senses changes in customer status through a perception system, updates customer memory through a memory management system, decomposes and makes decisions on tasks based on preset business rules and the outputs of various models in the AI ​​core layer, and coordinates multiple intelligent agents to jointly complete complex tasks through multi-intelligent collaboration.

[0040] As a further improvement of the present invention, the application orchestration connects the AI ​​capabilities and business processes of the AI ​​core layer and the application layer to form an automated and intelligent marketing workflow;

[0041] The intelligent computing workbench provides visualization tools, which business personnel and engineers use for model training, marketing workflow development, and data source management.

[0042] During the execution of the marketing campaign, generative AI is used to create content based on customer profiles, potential needs, communication preferences, and triggering scenarios, generating exclusive and personalized communication content, including insurance advice.

[0043] As a further improvement of the present invention, the multimodal data lake stores the generated personalized communication content and synchronizes it to different outreach channels;

[0044] Personalized communication content can be pushed to customer-preferred channels via API.

[0045] The present invention also provides a method for improving business through multiple agents, comprising the steps of:

[0046] The large model service gateway receives requests from business systems and distributes the data to different large models for processing.

[0047] Based on the requested business content, the large model retrieves the business process from the ontology knowledge base and calls the application orchestration to connect the business process with AI capabilities to form a marketing workflow.

[0048] The marketing workflow is executed according to the preset date, and the small model is called to update the latest customer profile and predict the customer's potential protection needs.

[0049] Use generative AI to generate marketing messages and product recommendations;

[0050] Choose distribution channels based on customer preferences and send marketing messages and product recommendations to customers.

[0051] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0052] This invention leverages the powerful capabilities of an AI factory to upgrade traditional standardized marketing processes into a highly personalized end-to-end experience. From in-depth customer data insights and accurate prediction of potential needs to intelligent generation of exclusive content and precise multi-channel outreach, it ultimately enhances customer loyalty and drives business growth. This personalized marketing approach not only improves customer satisfaction but also significantly increases the efficiency and conversion rate of insurance marketing.

[0053] This invention's data layer supports unified processing of multiple data sources such as documents, web pages, APIs, and audio, achieving multimodal data fusion. By constructing an ontology knowledge base specifically for the insurance industry, it enables accurate business understanding. An AI core layer is set up, atomically decomposing AI capabilities and arranging and combining them to form a complete business process. Multiple professional intelligent agents work collaboratively to provide comprehensive services. Based on customer profiles, it provides personalized insurance content generation services, achieving a personalized customer experience. Attached Figure Description

[0054] Figure 1 This is a schematic diagram of a system for improving business through multiple agents, as disclosed in one embodiment of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0056] The present invention will now be described in further detail with reference to the accompanying drawings:

[0057] like Figure 1 As shown, the present invention provides a system for improving business through multiple agents, comprising: a data layer, an AI core layer, an application layer, and an infrastructure layer. The data layer includes modules such as a multimodal data lake, data governance, an ontology knowledge base, and a vector database. The AI ​​core layer includes modules such as a large model service gateway, small models, and atomic capabilities. The application layer includes modules such as application orchestration, an agent framework, and an intelligent computing platform.

[0058] Data layer:

[0059] (1) A multimodal data lake includes documents, pages, APIs, and audio. Documents: refer to various text files, such as reports, contracts, emails, PDFs, and other structured or unstructured text information; Pages: refer to data obtained from the Internet, including page content, HTML structure, link information, etc.; APIs: refer to data obtained through application programming interfaces (APIs), which are usually structured data used for interaction and data transmission between systems; Audio: refer to various sound data, such as speech, music, ambient sound, etc., which can be raw audio files or processed audio streams.

[0060] Multimodal data lakes are used for: inputting historical customer data, capturing publicly available customer information, and converting customer service phone conversations into text to obtain communication content, thereby achieving data collection;

[0061] The personalized communication content generated at the application layer is stored, synchronized to different outreach channels, and pushed to the customer's preferred outreach channels via API interface;

[0062] Specifically, the multimodal data lake inputs structured and unstructured data such as customer insurance records, historical interaction data, policy information, past product purchases, service preferences, and communication records through documents and APIs. Policy information includes policy anniversary dates and birthdays. Information about customers on public social media is crawled from web pages, and audio of customer service phone conversations is obtained and transcribed into text to obtain the communication content.

[0063] (2) Data governance includes data collection, data cleaning, data standardization, and data anonymization. Data collection: refers to the process of collecting raw data from various sources, and is the starting step of data governance; data cleaning: aims to identify, correct, or delete inaccurate, incomplete, or irrelevant data to improve data quality; data standardization: converts data into a uniform format and structure to ensure data consistency and compatibility; data anonymization: uses technical means to modify or obscure sensitive information to protect privacy and comply with data regulations.

[0064] Data governance is used to: preprocess collected data to create customer profiles;

[0065] Specifically, the massive amounts of collected data are preprocessed to remove noise and standardize the format. Preprocessing includes data cleaning, date format standardization, and sensitive data anonymization.

[0066] (3) The ontology knowledge base includes insurance product ontology, customer and demand ontology, business process ontology, and risk and compliance ontology.

[0067] Insurance product ontology: refers to the formal and structured description and definition of insurance product-related concepts, attributes, and relationships, so that computers can understand and process insurance product information;

[0068] Customer and demand ontology: refers to a systematic description of customer information, customer needs, and the relationship between customers and products / services, in order to better understand and meet customer needs;

[0069] Business process ontology: refers to the formal modeling of business processes, activities, participants, inputs and outputs within an enterprise or organization to facilitate the analysis, optimization and automation of business processes;

[0070] Risk and compliance ontology: refers to a structured description of the types of risks, risk events, compliance requirements, and regulatory provisions involved in the business, used for risk management and compliance review.

[0071] In addition, the ontology knowledge base also includes ontology definition, visualization, version management, and manual correction.

[0072] Ontology definition: refers to the process of creating, writing and standardizing an ontology, including determining the scope, concepts, attributes and relationships of the ontology, and expressing them using a specific language (such as OWL);

[0073] Visualization: refers to presenting the complex structures and relationships in the ontology knowledge base in a graphical and intuitive way to help users better understand and browse ontology content;

[0074] Version management refers to tracking, controlling, and managing different versions of an ontology knowledge base to ensure that the ontology maintains consistency and traceability during its evolution.

[0075] Manual correction refers to the manual review and correction of errors or inaccuracies that may arise from automated methods during the ontology construction or application process, in order to improve the accuracy and quality of the ontology.

[0076] The ontology knowledge base is used to: construct insurance product ontology, customer and demand ontology, business process ontology, and risk and compliance ontology within the ontology knowledge base;

[0077] Specifically,

[0078] The insurance product entity stores detailed information on all insurance products, including coverage, terms, rates, and applicable populations; the customer and demand entity stores profile information on customers' age, occupation, family structure, income, risk preference, health status, and life stage, and establishes a mapping relationship with potential protection needs; the business process entity stores standard processes, key touchpoints, and communication norms for insurance marketing and after-sales service; and the risk and compliance entity ensures the compliance of all marketing content.

[0079] (4) A vector database is a database specifically designed for storing, managing, and querying vector data. In the field of AI, especially machine learning and deep learning, various types of data (such as text, images, audio, and video) often need to be converted into high-dimensional numerical vectors through embedding techniques. These vectors can capture the semantic information and features of the data, and in multidimensional space, similar data have closer vector distances. Vector databases are an important infrastructure for data management in the AI ​​era. They organize and retrieve data in a unique way, greatly enhancing the ability of AI models to process unstructured data and perform intelligent applications.

[0080] Vector databases are used to vectorize customer profiles, product information in insurance product entities, and communication content data.

[0081] AI Core Layer:

[0082] (1) Large model service gateway, including unified access, load balancing, authentication, monitoring and billing, model adapter and intelligent routing.

[0083] Unified access: refers to providing a unified entry point for different applications, users or systems to access large model services. This simplifies client integration and ensures that all requests are processed through a controlled channel.

[0084] Load balancing: responsible for distributing incoming requests to multiple large model instances in the backend to optimize resource utilization, maximize throughput, reduce response time, and prevent a single model from being overloaded;

[0085] Authentication, monitoring, and billing: Authentication verifies the identity and permissions of the requester, ensuring that only authorized users or systems can access the model service; monitoring tracks the performance, availability, and health status of the model service in real time and collects key metrics; billing records and statistically analyzes the usage of the model service, providing data support for subsequent cost calculations.

[0086] Model adapter: Responsible for converting requests from different sources or formats into a unified format that large models can understand and process, and vice versa. It solves compatibility issues between different models or interfaces;

[0087] Intelligent routing: Based on certain strategies (such as request type, user permissions, model load, cost-effectiveness, etc.), requests are intelligently routed to the most suitable large model instance or service to optimize performance, cost, or user experience.

[0088] Large model service gateway, used for:

[0089] A unified access service is provided through a large model service gateway, which receives requests from different applications, users or systems and distributes the requests to various model instances through load balancing.

[0090] During the execution of the marketing workflow by the large model, generative AI is used to create content based on customer profiles, potential needs, communication preferences and triggering scenarios, generating exclusive and warm personalized communication content, including insurance advice.

[0091] (2) Small models, including fine-tuned small models and industry vertical models.

[0092] Fine-tuning small models refers to small models that are further trained (fine-tuned) using task- or domain-specific datasets, building upon pre-trained large or general models. Through fine-tuning, these small models can be better adapted to the needs of specific scenarios while maintaining a small size and efficient inference speed.

[0093] Industry-specific vertical models: These are models specifically built and optimized for a particular industry or vertical field (such as finance, healthcare, law, education, etc.). These models are typically trained using the industry's expertise, terminology, and data, thereby exhibiting higher accuracy and professionalism within that field.

[0094] Small models are used for:

[0095] By fine-tuning pre-trained large or general models using small models, optimized models for specific industries or vertical fields can be constructed.

[0096] Specifically, small-scale models are used to construct customer life stage identification models, potential demand prediction models, and content preference analysis models for:

[0097] Based on the customer life stage identification model, the system automatically identifies the customer's life stage according to data on age, marital status, and children.

[0098] Based on the potential demand forecasting model, combined with the customer's life stage, historical consumption behavior, and risk preference, we predict the customer's potential insurance protection needs.

[0099] Based on the content preference analysis model, we analyze customers' past feedback on different types of content to identify their preferred communication styles and information presentation formats.

[0100] (3) Atomic capabilities, including IDP intelligent document analysis, knowledge graph, speech-to-text, intent recognition, document vectorization and knowledge recommendation capabilities.

[0101] IDP Intelligent Document Analysis refers to intelligent document processing capabilities. It combines artificial intelligence technologies (such as OCR, NLP, machine learning, etc.) to automatically identify, extract, classify, and process information in unstructured or semi-structured documents, thereby improving the efficiency and accuracy of document processing.

[0102] Knowledge graphs are a method of representing knowledge using a graph structure, where nodes represent entities (such as people, places, and events) and edges represent relationships between entities. They can store and organize large amounts of complex knowledge in a structured way, facilitating computer understanding and reasoning.

[0103] Speech-to-text (STO) refers to the technology of converting human speech signals into editable and searchable text content. This typically involves acoustic and language models and is widely used in scenarios such as voice assistants, meeting recording, and voice input.

[0104] Intent recognition refers to the technology of identifying the true purpose or intent implied in user input (usually natural language text or speech). For example, in a chatbot, it determines whether the user wants to "check the weather" or "order food".

[0105] Document vectorization refers to the process of converting text documents into numerical vectors (usually high-dimensional vectors). These vectors can capture the semantic information of the documents, allowing the similarity between documents to be calculated using vector distance, which is very useful in fields such as text search, clustering, and recommendation.

[0106] Knowledge recommendation refers to the ability to recommend relevant knowledge, information, or documents from a knowledge base based on a user's interests, behavior, or current context. It utilizes knowledge graphs, user profiles, and recommendation algorithms to provide personalized knowledge services.

[0107] Atomic capabilities are used for:

[0108] Based on IDP's intelligent classification and analysis capabilities, unstructured documents provided by customers are analyzed, and customer intent is understood based on intent recognition capabilities.

[0109] Leveraging knowledge graph capabilities for complex relationship reasoning, and utilizing knowledge recommendation capabilities to recommend products to customers based on the reasoning results;

[0110] Multimedia data is processed based on speech-to-text capabilities, and vector representations are generated for text and images based on document vectorization capabilities.

[0111] Application layer:

[0112] (1) Application orchestration, including workflow orchestration, agent development, Tools / MCP, and application release monitoring.

[0113] Application orchestration refers to the process of designing, executing, and managing a series of tasks or activities (i.e., workflows). It defines the sequence, dependencies, and execution logic of tasks to automate and coordinate complex business processes or system operations.

[0114] Intelligent agent development refers to the ability to develop and build intelligent software entities or programs (i.e., intelligent agents) that can perceive their environment, make decisions, and take actions. These intelligent agents can be autonomous, responsive, and capable of performing specific tasks.

[0115] Tools / MCP refers to the various auxiliary tools needed for developing and managing applications.

[0116] Application release monitoring refers to the ability to continuously track and monitor an application's performance, health status, user experience, and the release process itself after its release. This helps to identify and resolve issues promptly, ensuring the application runs stably.

[0117] Application orchestration, used for:

[0118] By applying orchestration to define the sequence, dependencies, and execution logic of tasks, we can develop and build intelligent software entities that can perceive the environment, make decisions, and take actions.

[0119] Specifically, application orchestration connects the AI ​​capabilities and business processes of the AI ​​core layer and application layer to form an automated and intelligent marketing workflow;

[0120] (2) Intelligent agent framework, including perception system, tool calling framework, task orchestration engine, multi-agent coordination, memory management system, task decomposition engine, reasoning decision engine and reflection and self-optimization.

[0121] Perception system: refers to the ability of an intelligent agent to acquire information, data, and signals from its environment. This includes understanding and recognizing the state and changes of the external world through sensors, API interfaces, data inputs, etc.

[0122] Tool invocation framework: This refers to the ability of an intelligent agent to recognize and invoke external tools, APIs, or services to complete specific tasks. This allows the intelligent agent to extend its functionality, interact with external systems, and perform more complex tasks;

[0123] Task orchestration engine: Responsible for coordinating and arranging the order and dependencies of a series of tasks executed by an agent based on preset logic, rules, or the current context. It ensures that tasks are executed according to the correct process and conditions;

[0124] Multi-agent collaboration refers to the ability of multiple agents to communicate, cooperate, and coordinate with each other to jointly achieve a complex goal. This typically involves task allocation, information sharing, and conflict resolution mechanisms.

[0125] Memory management system: responsible for the agent's storage, retrieval, and management of past experiences, knowledge, and information. This includes short-term memory (such as the current dialogue context) and long-term memory (such as learned knowledge and experience);

[0126] Task decomposition engine: responsible for breaking down a large, complex task into smaller, more manageable, and executable subtasks, which helps the agent solve complex problems step by step and improve efficiency;

[0127] Reasoning and decision-making engine: It is the brain of the intelligent agent, responsible for logical reasoning, analysis and judgment based on perceived information, knowledge in memory and set goals, so as to make the best action decision;

[0128] Reflection and self-optimization refer to the ability of an intelligent agent to evaluate its own behavior, decisions, and outcomes, learn from them, and adjust its strategies, knowledge, or parameters to continuously improve its performance and efficiency.

[0129] Intelligent agent framework, used for:

[0130] The intelligent agent framework provides intelligent software entities with the ability to acquire information, data, and signals from the environment; the ability to identify and invoke external tools, APIs, or services; the ability to arrange the order and dependencies of tasks performed by intelligent software entities; the ability to communicate and collaborate among multiple intelligent software entities; the ability for intelligent software entities to store, retrieve, and manage past experiences, knowledge, and information; and the ability to perform logical reasoning, analysis, and judgment based on acquired information, data, signals, memorized knowledge, and set goals.

[0131] Specifically,

[0132] The intelligent framework continuously senses changes in customer status through a perception system, updates customer memory through a memory management system, decomposes and makes decisions based on preset business rules and the outputs of various models in the AI ​​core layer, and coordinates multiple intelligent agents to complete complex tasks through multi-intelligent collaboration.

[0133] (3) Intelligent computing workbench, including workspace, operator development SDK, image management, data source management and workflow development.

[0134] Workspace: refers to a self-contained, configurable virtual environment provided to users for development, experimentation, and task execution, where users can manage projects, files, and resources;

[0135] Operator Development SDK: This refers to a software development kit used for developing and integrating "operators" (i.e., encapsulated units for specific computational tasks or algorithms). It provides a set of interfaces, libraries, and tools to facilitate developers in creating reusable and pluggable computational modules;

[0136] Image management refers to the ability to manage the storage, version control, distribution, and deployment of "images" (usually container images, such as Docker images) of computing environments, software configurations, and dependencies. This ensures consistency between development and runtime environments.

[0137] Data source management refers to the unified connection, configuration, authentication, and management of various data sources (such as databases, file storage, API interfaces, etc.). This enables the workbench to easily access and utilize the required data;

[0138] Workflow development refers to providing tools and interfaces that enable users to design, build, and orchestrate complex computational or data processing workflows. This typically involves defining task sequences, data flow, and conditional checks to automate the entire process.

[0139] The intelligent computing workbench is used for:

[0140] The intelligent computing workbench provides a virtual environment for development, experimentation, and task execution, and offers the ability to store, version control, distribute, and deploy images of computing environments, software configurations, and dependencies. It also provides tools and interface design for building and orchestrating computing or data processing workflows.

[0141] Specifically, the intelligent computing workbench provides visualization tools, which business personnel and engineers use for model training, marketing workflow development, and data source management.

[0142] Execute the workflow to generate personalized communication content and send it to customers.

[0143] Infrastructure layer, used for:

[0144] Provides basic cloud computing capabilities.

[0145] The business layer is used for:

[0146] Specific business scenarios for insurance agents.

[0147] Example:

[0148] A system for improving business through multiple agents, applying the present invention, takes customer birthday outreach as an example to complete the entire process of insurance marketing, including:

[0149] Step 1: Data Input and Processing

[0150] Multimodal data lake (data input):

[0151] Documents / API: Input structured and unstructured data such as customer insurance application records, historical interaction data, policy information (policy anniversary, birthday, etc.), past product purchases, service preferences, and communication records.

[0152] Web / Audio: Capture customer information on public social media (interests, recent topics of interest, etc.) and voice-to-text content of customer service phone conversations to enrich user profiles.

[0153] Data governance:

[0154] Data collection, cleaning, de-identification, standardization, and normalization: Preprocessing the massive amounts of collected data to remove noise, standardize the format, and ensure data quality and privacy security.

[0155] Example: Clean customer address information, standardize date formats, and de-identify sensitive data.

[0156] Step 2, Knowledge Construction and Integration (Ontology Knowledge Base & Vector Database):

[0157] Ontology knowledge base:

[0158] Insurance Product Contents: Stores detailed information on all insurance products, including coverage, terms and conditions, rates, and applicable groups.

[0159] Customer and Demand Ontology: Stores customer profile information such as age, occupation, family structure, income, risk preference, health status, and life stage (single, newly married, with children, retired, etc.), and establishes a mapping relationship with potential protection needs.

[0160] Business process entity: standard processes, key touchpoints, and communication norms for storage insurance marketing and after-sales service.

[0161] Risk and Compliance Ontology: Ensure the compliance of all marketing content and avoid misleading advertising.

[0162] Version control / manual correction: Continuously update and optimize ontology knowledge to ensure accuracy.

[0163] Vector database:

[0164] Vectorizing customer profiles, product information, and communication content facilitates efficient similarity matching and semantic search.

[0165] Step 3, Intelligent Analysis and Decision-Making (Large Model Service Gateway & Small Model & Atomic Capabilities & Intelligent Agent Framework):

[0166] Large-scale model service hub (AI-HUB):

[0167] Unified access / load balancing / intelligent routing: Receives requests from the data lake and business systems, and distributes the data to different AI models for processing.

[0168] Authentication, monitoring, and billing: ensuring data security and reasonable resource allocation.

[0169] Small model (fine-tuned small model / industry vertical model):

[0170] Customer Life Stage Identification Model: Based on data such as the customer's age, marital status, and children, the model automatically identifies the customer's current life stage (e.g., about to get married, becoming a parent for the first time, children's education stage, retirement planning, etc.).

[0171] Potential demand forecasting model: Combining customers' life stages, historical consumption behavior, risk preferences, etc., to predict their potential insurance protection needs (such as critical illness insurance, children's education funds, pension, etc.).

[0172] Content Preference Analysis Model: Analyzes customers' past feedback on different types of content to identify their preferred communication styles and information presentation formats (text, images, videos, etc.).

[0173] Atomic Power:

[0174] IDP Intelligent Document Analysis / Intent Recognition: Parses unstructured text provided by customers to understand their intent.

[0175] Knowledge Graph / Knowledge Recommendation: Utilizing knowledge graphs for complex relationship reasoning to make product recommendations.

[0176] Intent recognition / speech-to-text: Processing multimedia data.

[0177] Document vectorization: generating vector representations for text, images, etc.

[0178] Intelligent agent framework (brain):

[0179] Perception / Memory Management System: Continuously senses changes in customer status and updates customer memory accordingly.

[0180] Task decomposition / reasoning decision engine: Performs task decomposition and decision-making based on preset business rules and AI model output.

[0181] Example: Recognizing that a customer's birthday is approaching, a "birthday greeting" task is triggered. The inference decision engine will determine the personalized direction and recommended content for this greeting based on the customer's life stage, potential needs, content preferences, and other factors.

[0182] Multi-agent collaboration: coordinating multiple agents to jointly complete complex tasks.

[0183] Examples include collaborative work such as "customer profiling intelligence agent", "product recommendation intelligence agent", "content creation intelligence agent", and "channel reach intelligence agent".

[0184] Reflection and self-optimization: Continuously learn from user feedback and marketing effectiveness to optimize decision-making models.

[0185] Step 4, Application Orchestration and Content Generation (Application Orchestration & Intelligent Computing Workbench & Workflow Development):

[0186] Application orchestration (workflow orchestration / agent development / application deployment monitoring):

[0187] Workflow orchestration: Connecting the AI ​​capabilities and business processes of the above-mentioned stages to form an automated and intelligent marketing workflow.

[0188] Example: Define a "Customer Birthday Care Workflow":

[0189] Trigger: N days before the customer's birthday.

[0190] Customer profile update: Call the small model to update the latest customer profile.

[0191] Potential demand analysis: Use small models to predict customers' potential protection needs.

[0192] Personalized content generation: Generative AI is used to generate birthday wishes and product recommendations.

[0193] Distribution channels: Select channels such as SMS, WeChat, and APP push based on customer preferences.

[0194] Performance evaluation: Monitor customer feedback and interaction data.

[0195] Intelligent Computing Workbench (AI-H):

[0196] Workspace / Mirror Management / Data Source Management / Workflow Development: Provides visualization tools to facilitate business personnel and AI engineers in model training, workflow development, data management, etc.

[0197] Operator Development SDK: Provides development interfaces and supports customized feature development.

[0198] Content creation (using generative AI):

[0199] Generative AI: Based on customer profiles (life stage, interests and hobbies), potential needs, communication preferences, and birthday-triggered scenarios, it generates exclusive, warm birthday wishes and personalized insurance recommendations.

[0200] Example:

[0201] Customer A (30 years old, newly married, focused on family protection): "Dear [Customer Name], happy birthday! Newlyweds are at the sweetest time of your life. For your and your loved one's future, we especially recommend [XXXX Critical Illness Insurance] to safeguard your family. Wishing you happiness and fulfillment!"

[0202] Client B (45 years old, children's education stage, focusing on children's protection): "[Client Name], Happy Birthday! Thank you for your continued trust. Your child is about to face a crucial period in their education, and we have customized the [XXXX Education Savings Insurance] plan for you to help your child grow up without worry. We hope this care will add peace of mind to your birthday!"

[0203] Customer C (60 years old, retired, focused on retirement and health): "Dear [Customer Name], happy birthday! Wishing you good health and longevity! Retirement life should prioritize quality. We have carefully selected [XXXX High-End Medical Insurance] for you, allowing you to enjoy your later years without worries. We look forward to you experiencing our premium services!"

[0204] Intelligent agent development:

[0205] Content creation agent: responsible for calling generative AI and generating specific communication content based on context and objectives.

[0206] Recommendation agent: Based on customer needs and product knowledge base, recommend the most suitable product.

[0207] User profile intelligent agent: responsible for integrating and updating customer profiles.

[0208] Step 5, Result Output and Outreach (Data Layer):

[0209] Multimodal data lake: Stores generated personalized communication content (text, images, and even short video scripts) and synchronizes it to different outreach channels.

[0210] API: Through the API interface, personalized content can be pushed to the customer's preferred channels.

[0211] SMS / WeChat / APP push: The generated content will be accurately pushed to customers through channels such as their pre-registered mobile phone number, WeChat, and insurance company APP.

[0212] Customer service chatbots / online customer service: If a customer interacts, the chatbot can provide intelligent responses and further guidance based on the generated content.

[0213] Step 6, Effect Evaluation and Iteration (Reflection and Self-Optimization):

[0214] Data collection / cleaning / data quality inspection: Collect customer interaction data (click-through rate, open rate, response rate, inquiry rate, conversion rate, etc.).

[0215] Model training / model adjustment / model maintenance: Based on customer feedback and marketing performance data, the AI ​​factory's brain (intelligent framework) will reflect and self-optimize.

[0216] Example: If a certain communication style is found to have a low click-through rate, generative AI will learn and adjust its content generation strategy. If a certain type of product has a high recommendation conversion rate, the recommendation model will increase the recommendation weight for this type of product.

[0217] Advantages of this invention:

[0218] This invention leverages the powerful capabilities of an AI factory to upgrade traditional standardized marketing processes into a highly personalized end-to-end experience. From in-depth customer data insights and accurate prediction of potential needs to intelligent generation of exclusive content and precise multi-channel outreach, it ultimately enhances customer loyalty and drives business growth. This personalized marketing approach not only improves customer satisfaction but also significantly increases the efficiency and conversion rate of insurance marketing.

[0219] This invention's data layer supports unified processing of multiple data sources such as documents, web pages, APIs, and audio, achieving multimodal data fusion. By constructing an ontology knowledge base specifically for the insurance industry, it enables accurate business understanding. An AI core layer is set up, atomically decomposing AI capabilities and arranging and combining them to form a complete business process. Multiple professional intelligent agents work collaboratively to provide comprehensive services. Based on customer profiles, it provides personalized insurance content generation services, achieving a personalized customer experience.

[0220] This invention introduces dynamic planning and adaptive scheduling capabilities based on AI large models to intelligently select, combine, and adjust execution paths in traditional workflow orchestration, making it more adaptable to changing and personalized inputs. This allows the results to meet the requirements of insurance customers who seek personalization and emotional resonance, and who need exclusive and warm communication content.

[0221] This invention provides an intelligent model routing method that automatically selects the optimal model based on task complexity and domain characteristics. Since different choices such as large and small models may occur during business design, we have designed a mechanism to automatically select the most suitable model based on task type and domain characteristics.

[0222] This invention covers multiple aspects, including data collection, decision logic, and system architecture. Task and domain feature identification involves performing Natural Language Processing (NLP) on user-input request text to extract key information, establishing a predefined list of task types, and then performing entity matching with an insurance domain knowledge base to accurately identify the insurance domain. Model capability and cost-effectiveness require detailed evaluation of each integrated large model to quantify its performance across different task types and domains, particularly evaluating performance metrics, cost metrics, and latency metrics. Dynamic decision engine selection involves real-time decision-making based on identified task features and model evaluation data, using methods such as rule engines, learning-based recommendation systems, multi-objective optimization, and fallback mechanisms. System architecture integration integrates components such as request interception and parsing, decision invocation, model routing, feedback loops, and configuration management into the service gateway. Continuous optimization and iteration enable continuous monitoring and optimization of the dynamic model.

[0223] This invention utilizes a large model to achieve a certain degree of semantic understanding of document content. This means the system can not only recognize text but also understand the relationships between text, the context, and the overall intent of the document. Technically, a pre-trained large model is fine-tuned on a large amount of document data, enabling it to learn domain-specific document patterns and semantic relationships. Features such as text, layout information, and fonts extracted from the WPS engine are used as input to the large model to enhance its understanding capabilities.

[0224] The advantages of this invention's application of orchestration and intelligent agent frameworks include: introducing dynamic planning and adaptive scheduling capabilities based on a large AI model. The orchestration framework is no longer a simple set of preset steps, but rather intelligently selects, combines, and adjusts execution paths based on real-time input, task objectives, and environmental changes. Several key mechanisms ensure this: First, the large AI model provides reasoning capabilities to interpret high-level user requests, decompose them into smaller subtasks, and generate necessary logic or code in real time. This goes beyond predefined functions or static architectures, allowing for more flexible and autonomous behavior. Second, dynamic orchestration and sandbox execution generate and execute code or logic instantly in a secure sandbox environment, allowing for real-time problem-solving and adaptation to unforeseen circumstances. Intelligent agents can "think, act, observe results, and iterate" within the same loop, continuously adjusting their methods based on observations, achieving dynamic and adaptive workflow management.

[0225] This invention's orchestration framework is a "model router" and a "tool coordinator." It determines which large AI model is best suited for the current task through task query analysis. This analysis uses natural language processing to understand the intent, keywords, and context of the task and query, extracting relevant features from the input to aid model selection. Furthermore, it optimizes the switching and data flow between different models and tools, reducing latency and resource waste, thus achieving multi-model / tool ​​collaboration and optimization.

[0226] This invention provides a high-level interpretability and visualization interface, allowing users to clearly see how the AI ​​system understands tasks, breaks down problems, calls different components, and ultimately arrives at results. This not only enhances user trust but also facilitates troubleshooting and performance tuning for developers, achieving interpretable and visual orchestration.

[0227] This invention uses a large AI model as the "brain" of an intelligent agent, endowing it with powerful autonomous planning and multi-step reasoning capabilities. The intelligent agent no longer merely executes instructions, but can understand high-level goals and autonomously decompose them into a series of executable sub-tasks, and select appropriate tools to complete these sub-tasks, thus realizing autonomous planning and reasoning based on a large model.

[0228] This invention's intelligent agent framework supports collaborative work among multiple intelligent agents, forming an "intelligent agent team." Each agent can focus on a specific domain or task and work together to solve complex problems through efficient communication and collaboration mechanisms (e.g., shared knowledge bases, task distribution, and result integration). This collaboration may generate "emergent intelligence" that a single agent cannot achieve, thus realizing multi-agent collaboration.

[0229] The above are merely preferred embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A system for improving business through multiple agents, characterized in that, include: The system comprises a data layer, an AI core layer, and an application layer. The data layer includes modules such as a multimodal data lake, data governance, an ontology knowledge base, and a vector database. The AI ​​core layer includes modules such as a large model service gateway, small models, and atomic capabilities. The application layer includes modules such as application orchestration, an intelligent agent framework, and an intelligent computing workbench. The data layer is used for: Data collection is achieved by inputting historical customer data into the multimodal data lake, capturing publicly available customer information, and converting customer service phone conversations into text to obtain the communication content. The data governance process preprocesses the collected data to create customer profiles. The insurance product ontology, customer and demand ontology, business process ontology, and risk and compliance ontology are constructed in the ontology knowledge base. The customer profile, product information in the insurance product entity, and communication content data are vectorized using a vector database. The personalized communication content generated by the application layer is stored, and the personalized communication content is pushed to the customer's preferred reach channels through the API interface; The AI ​​core layer is used for: A unified access service is provided through a large model service gateway, which receives requests from different applications, users or systems and distributes the requests to various model instances through load balancing. By fine-tuning pre-trained large or general models using small models, optimized models for specific industries or vertical fields can be constructed. IDP provides intelligent document analysis, knowledge graph, speech-to-text, intent recognition, document vectorization, and knowledge recommendation capabilities through atomic capabilities. The application layer is used for: By applying orchestration to define the sequence, dependencies, and execution logic of tasks, we can develop and build intelligent software entities that can perceive the environment, make decisions, and take actions. The intelligent agent framework provides the intelligent software entities with the ability to acquire information, data, and signals from the environment; the ability to identify and invoke external tools, APIs, or services; the ability to arrange the order and dependencies of tasks performed by intelligent software entities; the ability to communicate and collaborate among multiple intelligent software entities; the ability for intelligent software entities to store, retrieve, and manage past experiences, knowledge, and information; and the ability to perform logical reasoning, analysis, and judgment based on acquired information, data, signals, memorized knowledge, and set goals. The intelligent computing workbench provides a virtual environment for development, experimentation, and task execution, and offers the ability to store, version control, distribute, and deploy images of computing environments, software configurations, and dependencies. It also provides tools and interface design for building and orchestrating computing or data processing workflows. Execute the workflow to generate personalized communication content and send it to customers.

2. The system for improving business through multiple agents according to claim 1, characterized in that: It also includes an infrastructure layer for providing basic cloud computing capabilities.

3. The system for improving business through multiple agents according to claim 1, characterized in that: The multimodal data lake inputs structured and unstructured data such as customer insurance records, historical interaction data, policy information, past product purchases, service preferences, and communication records through documents and APIs. The policy information includes policy anniversary and birthday. Information of customers on public social media is crawled from web pages, and customer service telephone communication voices are obtained through audio and converted into text to obtain the communication content.

4. The system for improving business through multiple agents according to claim 1, characterized in that: The insurance product itself stores detailed information on all insurance products, including coverage, terms, rates, and applicable groups. The customer and demand ontology stores profile information of customers, including their age, occupation, family structure, income, risk preference, health status, and life stage, and establishes a mapping relationship with potential protection needs. The business process entity stores the standard processes, key touchpoints, and communication specifications for insurance marketing and after-sales service. The risk and compliance ontology ensures the compliance of all marketing content.

5. The system for improving business through multiple agents according to claim 1, characterized in that: The data governance process involves preprocessing the massive amounts of collected data to remove noise and standardize the format. The preprocessing includes data cleaning, date format standardization, and sensitive data anonymization.

6. The system for improving business through multiple agents according to claim 3, characterized in that: By fine-tuning pre-trained large or general models using smaller models, optimized models can be built for specific industries or vertical fields; including: Based on the customer life stage identification model, the system automatically identifies the customer's life stage according to data on age, marital status, and children. Based on the potential demand forecasting model, combined with the customer's life stage, historical consumption behavior, and risk preference, we predict the customer's potential insurance protection needs. Based on the content preference analysis model, we analyze customers' past feedback on different types of content to identify their preferred communication styles and information presentation formats.

7. The system for improving business through multiple agents according to claim 1, characterized in that: IDP provides intelligent document analysis, knowledge graph, speech-to-text, intent recognition, document vectorization, and knowledge recommendation capabilities through atomic capabilities, for the following purposes: Based on IDP's intelligent classification and analysis capabilities, unstructured documents provided by customers are analyzed, and customer intent is understood based on intent recognition capabilities. Leveraging knowledge graph capabilities for complex relationship reasoning, and utilizing knowledge recommendation capabilities to recommend products to customers based on the reasoning results; Multimedia data is processed based on speech-to-text capabilities, and vector representations are generated for text and images based on document vectorization capabilities.

8. The system for improving business through multiple agents according to claim 1, characterized in that: The intelligent framework continuously senses changes in customer status through a perception system, updates customer memory through a memory management system, decomposes and makes decisions based on preset business rules and the outputs of various models in the AI ​​core layer, and coordinates multiple intelligent agents to complete complex tasks through multi-intelligent collaboration.

9. The system for improving business through multiple agents according to claim 1, characterized in that: The application orchestration connects the AI ​​capabilities and business processes of the AI ​​core layer and the application layer to form an automated and intelligent marketing workflow. The intelligent computing workbench provides visualization tools, which business personnel and engineers use for model training, marketing workflow development, and data source management. During the execution of the marketing campaign, generative AI is used to create content based on customer profiles, potential needs, communication preferences, and triggering scenarios, generating exclusive and personalized communication content, including insurance advice.

10. A method for improving services through multiple agents, employing a system for improving services through multiple agents as described in any one of claims 1 to 9, characterized in that: The large model service gateway receives requests from business systems and distributes the data to different large models for processing. Based on the requested business content, the large model retrieves the business process from the ontology knowledge base and calls the application orchestration to connect the business process with AI capabilities to form a marketing workflow. The marketing workflow is executed according to the preset date, and the small model is called to update the latest customer profile and predict the customer's potential protection needs. Use generative AI to generate marketing messages and product recommendations; Choose distribution channels based on customer preferences and send marketing messages and product recommendations to customers.

Images