Multi-Agent Artificial Intelligence System Based on Contextual Analysis of Organizational Processes

Abstract

The invention relates to a multi-agent based artificial intelligence system with agentic decision-making and dynamic task orchestration capabilities, running entirely on-premise on servers, and used in comprehensive corporate processes (including pensions, contracts, health, claims, policy management, digital channels, legislation, information security, human resources, campaigns, and workflows) within the insurance and pension industry.

Description

1 TARIFF MULTIPLE-AGENT APPROACH BASED ON CONTEXTUAL ANALYSIS OF ORGANIZATIONAL PROCESSES ARTIFICIAL INTELLIGENCE SYSTEM Technical Area 5 The invention relates to the comprehensive corporate processes of the insurance and pension sector (pension, contract, health, claims, policy management, digital channels, legislation, information security, human resources (including resources, campaigns and workflows) used multi-expert agent-based artificial intelligence. It is related to the intelligence system. 10 The invention specifically enables agent-based decision-making that runs entirely on on-premise servers. multi-specialist agent with mission delivery and dynamic task orchestration capabilities It is related to agent-based artificial intelligence systems. State of the Art Today, the existing artificial intelligence solutions developed for corporate structures are mostly... single-task, fixed-rules-based, or general-purpose model integration focused on These are structures. At the heart of these systems are those that directly answer the user's query or 20 single-function artificial intelligence applications or rules that automate a specific process These include chat systems based on security principles. However, such approaches, especially those involving high security, are problematic. and responding to complex needs arising in regulated sectors It is insufficient. Especially existing systems that are operated entirely on on-premise servers. Artificial intelligence solutions generally serve only a single functional area. For example, specific operations such as human resources operations or help desk support. Applications developed for processes will only perform defined tasks. It is structured in this way. However, there are multiple artificial 30s with different areas of expertise. The intelligence agent interacts by analyzing the context and sequentially transferring tasks. The process involves automatically transferring the outputs of these agents to the next process and capabilities such as having all of this happen under a single central decision-making mechanism, Existing solutions either don't have it at all or have very limited features. is being implemented. 35 2 A significant drawback of current solutions is that different areas of expertise are isolated from each other. and is designed in a non-interactive way. Human resources operations, institution internal process management, information security support units, platform management and operational Agents developed in areas such as support often operate independently of each other. not sharing data and coordinating under a common decision-making mechanism 5 It is unable to undertake the task in this way. This situation reflects multifaceted institutional needs. This prevents the development of integrated solutions. Another fundamental shortcoming in current approaches is the lack of dynamic contextual information. The inability to include them in the execution processes. RAG-like methods are mostly model 10 It is implemented in a centralized and static manner. These methods involve agents during execution. They are unable to request the information they need or to improve the quality of their decisions with up-to-date data. It is unable to process the data in a way that will improve it. This situation affects the accuracy and reliability of the responses generated. This limits its reliability. Furthermore, it creates a new business need in traditional architectures. Integrating new agents into the system to be able to respond requires high technical expertise. 15 and requires a long development process. This situation necessitates the agility of the corporate structure. reduces and enables the artificial intelligence infrastructure to adapt to changing business needs. This makes things more difficult. There are also significant limitations in terms of data security and regulatory compliance. Many solutions are available, including role-based access control or auditable transaction logs. institutionally mandated security layers, such as archiving in this format. It does not include this. This also applies to highly regulated sectors such as insurance and pensions. This prevents scaled-up use. Solutions in current technology; • Context-sensitive automation of agent coordination, • Integrated with sequential task transfer of multiple expert AI agents his work, • Dynamically incorporating contextual information into processes, • Low-code based rapid expansion and scalability, 30 • and meeting corporate security and auditability standards It is unable to bring together core competencies such as these within a single architecture. This situation affects insurance. and complex processes specific to the retirement sector are integrated with artificial intelligence, secure and limiting its scalable implementation and the need for innovative solutions 35 It reveals. 3 In conclusion, the existence of the above problems and the inadequacy of the current solutions, This has made it necessary to make improvements in the technical field. Purpose of the Invention The present invention eliminates the aforementioned disadvantages and contributes to the relevant technical field. Multi-agent analysis of organizational processes that brings new advantages It is related to artificial intelligence systems. The main purpose of the invention is to create an agentic 10 that runs entirely on on-premise servers. multi-specialist agents with decision-making and dynamic task orchestration capabilities The goal is to develop an artificial intelligence system based on this principle. The purpose of the invention is to address the complex corporate needs specific to the insurance and pension sector. 15 agentic artificial intelligence-based, multi-specialist agents used throughout the processes The goal is to create an organizational intelligence infrastructure that is constantly evolving and expanding. The system, traditional chatbots that serve only a single function or have fixed rules Going beyond employee automation solutions, we specialize in multiple verticals. a number of artificial intelligence agents under a single roof, on-premise servers, It coordinates in a safe and autonomous manner. 20 Another purpose of the invention is to eliminate the need for users to know which agent to choose. without delay, analyzing queries in natural language through contextual analysis and optimizing them The specialist agent automatically activates and works dynamically with other agents. 25 that enables its operation, having an agentic decision-making and task-direction architecture. The goal is to develop a multi-expert agent-based artificial intelligence system and method. This will enable... In a query regarding policy regulations, a regulatory agent; in a query regarding claims analysis... The damage agent is activated; the outputs of the agents are transferred to each other as needed, and The task chain is managed automatically by the system. Another aim of the invention is to enable each agent to work not only in their own area of ​​responsibility but also with other specialists. By enabling continuous interaction with agents, the quality and speed of the solution are improved. It enhances; and at the same time, it provides technical solutions to new process, business area or department needs. This allows for the rapid development of new agents without requiring specialized expertise, and also benefits existing ones. enabling integration into the ecosystem; thus allowing for both horizontal integration (agent 35 4 multiple scalable both vertically (increase in number) and vertically (increase in depth of expertise) The goal is to develop an expert agent-based artificial intelligence system and method. Another aim of the invention is to create a system that runs entirely on on-premise servers. This prevents corporate data from leaving the organization, and controls all data flow. 5 Keeping it under control in accordance with safety and regulatory standards; also LDAP Data security and access through role-based authentication and role-based agent authorization. Multi-expert agent-based artificial intelligence that ensures control and regulatory compliance. The goal is to establish the system and method. Another aim of the invention is to replace classic systems that serve only a single department. on the contrary, it offers a wide set of functions that can be used by the entire organization; human from resource processes to information security, from OKR management to internal communication support, It specializes in numerous functions, from policy analysis to health data interpretation. unifying them into a single structure through agents; thus enabling employees to perform operational tasks. to achieve high efficiency in their tasks and to meet the needs of the customer profile. multiple specialist agents that enable the company to acquire corporate insights quickly and accurately. The goal is to develop an artificial intelligence system and method based on this technology. Another aim of the invention is to create a 20-year-old insurance and pension company that focuses on the needs of the insurance and pension sector. a constantly growing and corporate entity that dynamically brings together a large number of specialist agents. fully integrated into the infrastructure; automating complex processes, human- reducing the need for intervention, providing high accuracy and speed, and being corporate. A multi-expert agent-based artificial intelligence system that sets a new standard in the use of AI. The goal is to reveal the intelligence system and method. 25 The structural and characteristic features and all the advantages of the invention are given in the figures below. And thanks to the detailed explanation written with references to these figures, it becomes clearer. This will be understood as such. Therefore, the evaluation should also include these forms and detailed explanations. This should be done taking this into consideration. 30 Figures that will help understand the invention. Figure 1: Schematic and workflow diagram of the system described in the invention. 35 Explanation of Part References 10. Agentic orchestration engine 20. Task chain manager 30. Fine-tuned model layer 5 40. Open source LLM layer 50. Commercial LLM layer Layer 60 RAG 70. Agent development interface 80. Output integration and response component 10 90. Auditable record component 100. Security and compliance layer A. Agent Detailed Description of the Invention This detailed explanation describes the invention as a multi-expert agent-based artificial intelligence system. preferred alternatives, solely for the purpose of better understanding the subject and It is explained in a way that will not create any limiting effects. 20 Schematic diagram of the multi-expert agent-based artificial intelligence system that is the subject of the invention. This information is provided. Accordingly, a multi-expert agent-based artificial intelligence system in its most basic form; The system's starting point and decision-making center, and the user interface. By analyzing user queries transmitted through the system at the semantic and contextual level, 25 Agentic orchestration that selects suitable specialist agents (A) and establishes the chain of duty. engine (10) manages the cooperation between agents (A), optimizes the task sequence and The task chain manager (20) who combines the results, insurance, pension, policy, legislation and fine-tuned model containing domain-specific models trained with damage data layer (30), which integrates open source models hosted within the organization and context 30 open source LLM layer (40) which performs the determination tasks, API-based external commercial models that enable integration with other models and include them in the task chain when needed. LLM layer (50), internal document archives, policy records, legislation databases RAG verifies the accuracy of answers by drawing contextual information from sources such as these. layer (60), agent 35 which enables the definition of new agent (A) and workflow design. The development interface (70) combines all the outputs produced in the system and the user interface. 6 Output integration and response component (80), which presents the output to the user through, all in the system The auditable record component (90) that archives the steps of the process, allows system access to data. controlling, ensuring auditability and compliance checks with regulations It includes the implementing security and control compliance layer (100). The invention is a multi-expert agent-based artificial intelligence system for insurance and pensions. in all of the highly complex corporate processes specific to the sector (retirement, contract, health, claims, policy management, digital channels, legislation, information security, human resources developed for use (including resources, campaigns and workflows), entirely running on an on-premise server infrastructure and with an agentic artificial intelligence architecture 10 a system that automatically coordinates numerous agents with different areas of expertise It is a platform. The invention is a multi-expert agent-based artificial intelligence system that utilizes natural elements from the user. It analyzes the queries in the language through the agentic orchestration engine (10) and the appropriate 15 Specialist agents (A) are automatically activated. The agents involved in the meeting (A) are each working in their own area of ​​expertise. They are autonomous software components. These agents (A) are functionally legislative agents. (A), damage analysis agent (A), policy management agent (A), health data agent (A), human 20 resources agent (A), information security agent (A), campaign and customer communications agent (A), business The flow and process agent can be (A). There can be “n” number of agents (A) in the system, and additional experts can be included. Agents (A) can be added via the system's agent development interface (70). For example; regulatory agent (A) interprets insurance legislation, claims analysis agent (A) calculates claims 25 evaluates the files, policy management agent (A) determines policy coverage, health Data agent (A) analyzes medical data, human resources agent (A) analyzes internal processes. manages, information security agent (A) oversees access controls, campaign and customer Communications agent (A) provides campaign management and customer communication. These agents (A) They form task chains by constantly interacting with each other. Every 30 interim results, analyses, data or decisions produced by agent (A) in its own area of ​​expertise The outputs that form the recommendations are transferred from one agent (A) to another as needed, and The entire process is completed autonomously from end to end. The system goes beyond classic solutions that serve only a single function, 35 (A) numerous agents specializing in different verticals are secured under one roof. 7 It enables autonomous coordination. Fine-tuned model. High-level information supported by domain-specific information thanks to layer (30) and RAG layer (60). Outputs are produced with accuracy. In addition, new agents can be developed with the agent development interface (70). (A) It can be added quickly without requiring technical expertise, making the system both horizontal and horizontal. (increase in the number of agent (A)) as well as vertically (increase in depth of expertise) 5 This makes it possible to scale up. The system described in the invention is designed entirely on-premise (IP) infrastructure. Its operation prevents corporate data from leaving the organization; it is LDAP-based. Data security, access control, and 10 through authentication and role-based authorization. Regulatory compliance is ensured. The starting point of the multi-expert agent-based artificial intelligence system that is the subject of the invention is and Agentic orchestration engine (10), which forms the decision center, by the user 15 queries or requests conveyed in natural language are analyzed at semantic, contextual and meaning levels. It is analyzing. Agentic orchestration engine (10) not only at the keyword level but also at the user level the intent behind the query, the context, and the expertise required for the solution It also identifies the types. As a result of this analysis, which specialist agents (A) are deployed 20 the order in which they should enter, the order in which they will work, and how the chain of command will be structured It is determined automatically. The Agentic Orchestration Engine (10) in the system the current agents' (A) capabilities matrix, job descriptions and previous It creates the optimal task chain based on patterns learned from interactions. This The structure relieves the user of the need to make manual selections or call specific agents. 25 It completely saves. The task chain manager (20) is determined by the agentic orchestration engine (10). It is responsible for executing the task flow and managing the interaction between agents. Each An agent (A) is positioned sequentially in the task chain, and the output of an agent (A) is 30 It is automatically converted into the input of the next agent (A). In this way, the singular processes combine to perform multi-step, complex tasks without human intervention. completion is ensured. The task chain manager (20) executes a steady stream and also stream execution 35 It also dynamically adapts to new data needs or changes in conditions that arise during the process. 8 It manages in this way. The said task chain manager (20), when necessary restructuring the mission sequence, deploying additional agents (A) or making them unnecessary It optimizes the flow by skipping steps. This feature makes the system context-sensitive. its adaptability and the key difference that sets it apart from classic automation solutions. It constitutes one of the points. 5 The fine-tuned model layer (30) includes domain-specific and task chain elements. It includes models that operate with high accuracy. These models are used in insurance, pension, Corporate data related to specific areas such as health, claims, policy analytics, or legislation. They were trained with sets of models. Each of these models, specialized in specific tasks, made decisions. It is automatically activated by the motor at the required points. For example, in a task chain requiring legislative analysis, with the relevant legislative dataset. The trained model comes into play and is solely responsible for generating solutions to this problem. This approach ensures that the responses are highly accurate, domain-specific, and... It ensures that it is produced in a manner consistent with corporate terminology. 15 The open source LLM layer (40) is a general-purpose natural resource hosted on on-premises infrastructure. large languages ​​that handle tasks such as language comprehension, information extraction, or classification. It includes models. The open source LLM layer (40), in particular, the user query In the initial stages of analysis, in context determination, classification and information extraction, 20 It is involved in the task. Open source LLM layer (40), fine-tuned model It works with layer (30). Open source LLM layer (40), general context solving and fine-tuned model layer (30) also domain-specific expertise It provides the open source LLM layer (40), fine-tuned model layer. (30) Thanks to its ability to work together, the system can handle both the 25 requiring deep expertise. It can provide high accuracy in various fields, as well as general-purpose information processing. It offers a wide range of analytical capabilities in its tasks. This hybrid approach, This prevents the system from being limited to the capabilities of only a particular model. The commercial LLM tier (50) supports licensed major language models offered by external providers. It is the component that integrates into the system via API-based integration. Commercial LLM layer (50), high language production capacity, access to different data sources or advanced understanding Their skills are automatically incorporated into the task chain in scenarios where they are required. Thus, in addition to its basic structure that operates with internal resources, the system accommodates specific usage scenarios. In these cases, it expands with external LLM services. This is especially true for large 35s. 9 Strategic flexibility in processes requiring large-scale knowledge production or linguistic creativity. It provides. RAG layer (60), information retrieval-augmented RAG (Retrieval-Augmented Generation is the production layer, and contextual information is needed during task execution. RAG is a dynamic information enrichment mechanism that is activated at birth. layer (60), secure data such as internal document archives, product and legislation records By scanning its resources, it finds, processes, and retrieves relevant information in real time. It is included in the chain of duties. The RAG layer (60) performs static data querying as well as; by specific agents (A) It is actively used. For example, a tool that provides information about "internal processes". Agent (A) automatically retrieves the relevant documents based on the query received from the user. It identifies, interprets the content, and generates a response based on this context. Similarly, RAG Tier 15 is used for tasks such as regulatory compliance or policy coverage. (60), by providing details on the correct articles or provisions, the accuracy of the outputs and This significantly increases its reliability. Thanks to this, the system only uses the model's reliability. It is not based on previously acquired knowledge; it is current, institution-specific, and when necessary. By integrating verifiable information into the context, highly accurate, auditable and It produces contextually rich outputs. 20 Agent development interface (70), in terms of the sustainability and scalability of the system It is a component of strategic importance. Thanks to the agent development interface (70), technical New agents (A) are easily created and configured without requiring expertise. and is integrated into the existing task chain. When a new agent (A) is added, agentic 25 The decision engine (10) automatically includes this agent (A) in the evaluation scope and It activates the system in appropriate queries. Thus, the system responds to changing business needs. an agile and rapidly adaptable structure that is constantly evolving It has been transformed. In addition, the functional diversity of agents (A) is both horizontal (agent (A)) (increase in number) and vertical (development of depth of expertise) growth 30 This approach makes it possible for the system to grow according to user needs. It is transforming into an evolving and constantly expanding ecosystem of agents. Security and control compliance layer (100) ensures the system's data security, access control and It is the most critical component in terms of regulatory compliance. Security and control compliance layer 35 (100), by integrating with corporate directory services, only corporate employees It provides access to the system. User identities are verified upon login, and each agent... Separate access policies are defined for (A) based on roles and responsibilities. This structure Thanks to this, each user interacts only with agents (A) for whom they are authorized and Agents (A) can only access authorized data sources. The security and compliance layer (100) also protects the data integrity of the system, personal to fulfill obligations regarding data protection and relevant laws It is structured to ensure compliance with regulations. This concerns security and... Compliance layer (80), KVKK, SEDDK and applicable in the insurance and pension sector It is designed to be fully compliant with similar regulations, thus enabling 10 ensuring the system has a high level of reliability in terms of regulatory requirements is doing. All steps of the system can be monitored via the auditable log component (90). They are archived in the form of records, and thus internal audit units or 15 Traceability is ensured by regulatory authorities when required. Output The integration and recording component (80) combines all outputs produced in the system and It is presented to the user through a user interface. The invention, a multi-expert agent-based artificial intelligence system, utilizes 20 natural elements from the user. from interpreting the language query to the automatic chain of tasks from the creation and integration of contextual information to the sequential deployment of expert agents. to autonomously carry out the entire process, including its coordination. It is structured. Each agent (A) is limited to performing operations only within its own mission area. not only that, but also by collaborating with other agents through sequential task transfer. 25 This structure enables processes to become multi-step and human- It is completed with high accuracy without the need for any intervention. Furthermore, the new agents (A) can be added to the system within minutes via the agent development interface (70) The possibility of adding to it makes the system a constantly expanding and evolving ecosystem. It enables him to win. 30 The implementation method for the multi-expert agent-based artificial intelligence system that is the subject of this invention is as follows: It is as follows; a) Agentic decision 35 of a query received in natural language via the user interface semantic and contextual level analysis by the engine (10), 11 b) Determining the expert agents (A) required for the solution according to the analysis result and these agents (A) are automatically assigned by the task chain manager (20) structuring, c) The output of each agent (A) in the task chain sequentially follows the next one. The agent (A) is made into input and the task flow is dynamically 5 execution, d) Fine-tuned model at points needed in the task chain. layer (30), open source LLM layer (40) and commercial LLM as needed commissioning of layer (50), e) Need for contextual information during task execution RAG layer (60) 10 corporate document archives, policy records and regulatory data being met in real time from their bases, f) New specialist agents (A) agent development without requiring technical expertise By defining it through the interface (70) and integrating it into the system, g) Identity 15 by the security and compliance layer (100) of all process steps with verification, role-based access control, and auditable transaction logs. to be secured, h) User interface via output integration and recording component (80) of the final output presented to the user via its face, i) Archiving of the final output by the auditable record component (90). 20

Claims

12 REQUESTS 1. In the comprehensive corporate processes of the insurance and pension sector (pension, contract, health, claims, policy management, digital channels, legislation, information security, human resources, campaigns and workflows) multiple specialists used 5 It is an agent-based artificial intelligence system; its feature is: - semantically analyzes user queries transmitted through the user interface. by analyzing at the contextual level, each in their own field of expertise Expert agents (A) that make up the autonomous software components performing the task The agentic orchestration engine (10) that selects and creates the task chain, 10 - Managing cooperation between agents (A), optimizing task sequence and the task chain manager that combines the results (20), - trained with insurance, pension, policy, legislation and claims data. a fine-tuned model layer containing domain-specific models (30), 15 - integrating open source models hosted within the organization and an open-source LLM layer that performs context-setting tasks (40), - Enables integration with API-based external models, when needed. commercial LLM layer (50) which includes in the task chain, 20 - internal document archives, policy records, legislation databases verifying the accuracy of answers by extracting contextual information from their sources. RAG layer (60) that provides - Agent that enables the definition of new agent (A) and workflow design. development interface (70), 25 - controls data access to the system, ensures auditability, and Security and control compliance that implements regulatory compliance checks It contains layer (80).

2. In the comprehensive corporate processes of the insurance and pension sector (pension, 30 contract, health, claims, policy management, digital channels, legislation, information (security, human resources, campaigns and workflows) multiple specialists are used It is an agent-based artificial intelligence system application method, the characteristic of which is; a) Agentic response to a query received in natural language via the user interface semantic and contextual level analysis by decision engine (10) 35 being done, 13 b) Determining the expert agents (A) required for the solution according to the analysis result. and these agents (A) are automatically ordered by the task chain manager (20). structuring as such, c) The output of each agent (A) in the task chain is sequentially one the next agent (A) is made the input and the task flow is dynamic 5 carried out in this manner, d) The field is fine-tuned at points needed in the task chain. model layer (30), open source LLM layer (40) and commercial if necessary Enabling the LLM layer (50), e) Need for contextual information during task execution RAG layer (60) 10 corporate document archives, policy records and regulatory data being met in real time from their bases, f) New specialist agents (A) agent development without requiring technical expertise By defining it through the interface (70) and integrating it into the system, g) Identity of all process steps by the security and compliance layer (80) 15 with verification, role-based access control, and auditable transaction logs. to be secured, h) The final output is made available to the user via the output integration and recording component (80) presented to the user via the interface i) Archiving of the final output by the auditable record component (90) 20 It includes the steps of the process.

Images