AI agent provisioning system, AI agent provisioning method, and AI agent provisioning program

Abstract

This system provides an AI agent provision system that reflects the personality traits of specific individuals, allowing users to select and use AI agents with diverse personality traits. [Solution] The AI ​​agent provisioning system of the present invention comprises a personality model generation means that generates a personality model of an AI agent that imitates the thinking style and behavioral patterns of an individual based on data acquired regarding the individual's characteristics, and a marketplace management means that manages the AI ​​agent as a tradable digital asset.

Description

Technical Field

[0001] The present invention relates to an AI agent providing system, an AI agent providing method, and an AI agent providing program.

Background Art

[0002] Conventionally, in organizations such as companies, advanced expertise and excellent business know-how have tended to be personalized to specific individuals. As a result, there has been a problem that the knowledge of such individuals is not sufficiently shared and utilized within the organization, and the opportunity to improve the productivity of the entire organization is lost. In order to avoid such personalization of business skills, as shown in Non-Patent Document 1, an AI agent system aimed at automatic execution of tasks is known.

Prior Art Documents

Patent Documents

[0003]

Non-Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, these conventional AI agents generally have uniform behavioral characteristics with a focus on general-purpose task execution capabilities. Therefore, personal characteristics possessed by a specific individual, such as expertise and values, are not reflected, and AI agents with diverse personal characteristics have not been provided in a selectable manner.

[0005] Therefore, the present invention provides an AI agent provisioning system that reflects the personality characteristics of a specific individual and allows users to select and use AI agents with diverse personality characteristics. [Means for solving the problem]

[0006] To solve the above problems, for example, the configuration described in the claims may be adopted. [Effects of the Invention]

[0007] According to the present invention, it is possible for users to select and use AI agents that reflect the personality characteristics of a specific individual and possess diverse personality traits. Other issues, configurations, and effects not mentioned above will be clarified by the following description of the embodiments. [Brief explanation of the drawing]

[0008] [Figure 1] Figure 1 shows an example of the configuration of the AI ​​agent provisioning system according to the first embodiment. [Figure 2] Figure 2 shows an example of the configuration of an AI agent generation server. [Figure 3] Figure 3 shows an example of the information stored in the auxiliary storage device of the AI ​​agent generation server. [Figure 4] Figure 4 shows an example of the configuration of a marketplace management server. [Figure 5] Figure 5 shows an example of the information stored in the auxiliary storage device of the marketplace management server. [Figure 6] Figure 6 shows an example of the configuration of an administrator terminal. [Figure 7] Figure 7 shows an example of the configuration of a user terminal. [Figure 8] Figure 8 shows an example of an interview flow. [Figure 9] Figure 9 shows an example of a personality analysis flow chart. [Figure 10]FIG. 10 is a diagram showing an example of a feedback loop formation flow. [Figure 11] FIG. 11 is a diagram showing an example of a thinking process construction flow. [Figure 12] FIG. 12 is a diagram showing an example of an external collaboration flow. [Figure 13] FIG. 13 is a diagram showing an example of a reference vector generation flow. [Figure 14] FIG. 14 is a diagram showing an example of a deviation degree detection flow. [Figure 15] FIG. 15 is a diagram showing an example of an agent registration and disclosure flow. [Figure 16] FIG. 16 is a diagram showing an example of an agent matching and provision flow. [Figure 17] FIG. 17 is a diagram showing an example of a billing and revenue distribution flow. [Figure 18] FIG. 18 is a diagram showing an example of an activity log recording flow. [Figure 19] FIG. 19 is a diagram showing an example of a first screen. [Figure 20] FIG. 20 is a diagram showing an example of a second screen. [Figure 21] FIG. 21 is a diagram showing an example of the configuration of an AI agent providing system according to the second embodiment. [Figure 22] FIG. 22 is a diagram showing an example of the application of an AI agent providing system according to the second embodiment. [Figure 23] FIG. 23 is a diagram showing an example of the configuration of a marketplace management server according to the second embodiment. [Figure 24] FIG. 24 is a diagram showing an example of the information stored in the auxiliary storage device of a marketplace management server according to the second embodiment. [Figure 25] FIG. 25 is a diagram showing an example of a virtual interview flow. [Figure 26] FIG. 26 is a diagram showing an example of a team formation simulation flow. [Figure 27] FIG. 27 is a diagram showing an example of a decision-making support flow. [Figure 28] Figure 28 shows an example of the quality control flow for Ghost. [Figure 29] Figure 29 shows an example of the content delivery flow. [Figure 30] Figure 30 shows an example of the third screen. [Modes for carrying out the invention]

[0009] (1) Overview of the first embodiment Figure 1 is a conceptual diagram showing an example configuration of the AI ​​agent provisioning system 100 (hereinafter simply referred to as System 100) according to this embodiment. System 100 consists of an AI agent generation server 200, a marketplace management server 400, an administrator terminal 600, a user terminal 700, and an AI agent model DB 800, all connected to each other via a network for communication. The network may be wired or wireless, and may be the internet, an intranet, a LAN, a WAN, or a combination thereof. Each terminal can send and receive information via the network.

[0010] Each server and terminal in System 100 may be a mobile device such as a smartphone, tablet, mobile phone, or personal digital assistant (PDA), or a wearable device such as glasses, a wristwatch, or clothing. Alternatively, it may be a stationary or portable computer, or a server located in the cloud or on a network. Functionally, it may also be a VR (Virtual Reality) terminal, an AR (Augmented Reality) terminal, or an MR (Mixed Reality) terminal. Alternatively, it may be a combination of multiple such terminals. For example, a combination of one smartphone and one wearable device can logically function as a single terminal. Other types of information processing terminals may also be used.

[0011] Each server and terminal in System 100 is equipped with a processor that runs an operating system, applications, and programs; main memory such as RAM (Random Access Memory); auxiliary storage such as IC cards, hard disk drives, SSDs (SolID State Drives), and flash memory; a communication control unit such as a network card, wireless communication module, or mobile communication module; input devices such as a touch panel, keyboard, mouse, voice input, and camera; and output devices such as monitors and displays. The output devices may also be devices or terminals that transmit information for output to external monitors, displays, printers, or other equipment.

[0012] The main memory stores various programs and applications (also referred to as modules or processing units), and the processor executes these programs and applications to realize each functional element of the entire system 100. These modules (processing units) may be implemented in hardware, such as through integration. Furthermore, each module may be an independent program or application, or it may be implemented as a subprogram or function within a single integrated program or application.

[0013] In this specification, each module is described as the entity (subject) that performs the processing, but in reality, the processor that processes various programs and applications (modules) executes the processing. Various databases (DBs) are stored in the auxiliary storage device. A "database" is a functional element (storage unit) that stores a data set so that it can handle any data manipulation (e.g., extraction, addition, deletion, overwriting, etc.) from the processor or an external computer. The implementation method of the database is not limited; for example, it may be a database management system, spreadsheet software, or text files such as XML or JSON.

[0014] The AI ​​agent according to the present invention refers to an artificial intelligence software system that can interact with users and the environment while making autonomous decisions and taking action. Unlike conventional chatbots or single-function AI systems that have been widely used, it possesses characteristics such as purpose-orientedness, state retention, and response adaptability. The AI ​​agent in this embodiment is, for example, a generative AI. Such an AI agent has the function of automatically generating responses in natural language and action suggestions in response to user inquiries. In particular, a large-scale language model (LLM) with large-scale natural language processing capabilities is used at the core of response generation, which enables human-like contextual understanding and flexible responses. In this embodiment, a configuration in which the AI ​​agent is a large-scale language model will be described as an example.

[0015] Large-scale language models are generally constructed using text data ranging from tens of billions to trillions of words, collected from publicly available texts on the internet, books, papers, dialogue logs, etc. These models, for example, are based on the Transformer architecture and learn linguistic patterns, contexts, and logical structures through self-supervised learning, possessing the ability to generate appropriate output sentences even for unknown input sentences.

[0016] In this invention, by further training such a large-scale language model with prompts that reflect an individual's values ​​and way of thinking, as well as multifaceted information specific to the expert who serves as the model, an individualized and optimal output that mimics personality intelligence is achieved. In other words, system 100 provides an AI agent equipped with a personality model that reproduces the values ​​and thinking tendencies of an individual, such as an expert with specific high skills.

[0017] Here, when we say that an AI agent model "learns multifaceted information about a specific individual's values ​​and thought patterns," it means that, in response to input training data that expresses such values ​​and thought patterns, the parameters held within the model (such as weight matrices and bias terms) are statistically readjusted, and the model is modified to acquire the tendencies of language patterns, logical reasoning, and judgment criteria included in the input as its internal representation.

[0018] Specifically, taking the Transformer architecture as an example, the weight matrix in the Self-Attention layer is updated to emphasize the relationships between words and syntax that characterize values ​​and thought patterns. Furthermore, the weights and biases in the feedforward network layer change in a way that makes it easier to select specific interpretations and evaluation criteria for the input context. In addition, the embedding vectors are adjusted to reflect vocabulary and expression selection tendencies, creating a high-dimensional representation space that conforms to the speech style and linguistic sensibilities of a particular individual.

[0019] These parameter updates are performed using optimization algorithms such as gradient descent, and the tendencies of value judgments and decision-making in the training data are established as the weight distribution of Attention and the activation patterns of the hidden layers. As a result, the AI ​​agent model becomes able to generate responses to inputs (queries) that derive logical structures and conclusions similar to those of the individual being trained.

[0020] In this embodiment, the trained model (AI agent model) corresponding to each AI agent may be stored, for example, in an external AI agent model DB 800 (hereinafter simply referred to as the model DB 800). The model DB 800 holds the trained parameters of each model as binary data and manages metadata such as model ID, version, training date and time, and application constraints in a table format. When the AI ​​agent generation server 200 generates a predetermined AI agent using the functions described later, it records the new model data in the model DB 800. Furthermore, when the marketplace management server 400 operates a predetermined AI agent using the functions described later, it loads the corresponding model data from the model DB and uses it for inference processing.

[0021] This enables the replication of judgment criteria similar to those of specific experts, allowing for responsive personalization that mimics the style of those experts or individuals. In the following explanation, an AI agent personalized based on the personality of a particular expert will be referred to as a "ghost," and the person who served as the model for the AI ​​agent will be referred to as the original.

[0022] The AI ​​agent generation server 200 of system 100 is equipped with various modules that acquire multifaceted data from the user and build and update a personality model based on this data, and has the function of generating responses that reflect the characteristics of the expert themselves.

[0023] Meanwhile, the marketplace management server 400 registers the usage rights of the various AI agents generated as digital assets for trading, and provides functions such as anonymization, billing, revenue sharing, and external integration. The marketplace management server 400 also supports effective connections between users and AI agents through recommendation and matching processes based on the agent's usage history and attributes.

[0024] Furthermore, the AI ​​agent continuously incorporates thought logs, behavioral records, and external evaluations to achieve autonomous evolution in order to improve its self-consistency and the accuracy of its judgments based on its values. This allows it to function not merely as a response device, but as an intelligent agent that behaves like an extension of the user, fulfilling roles such as dialogue, judgment, and advice. The configuration of each server is described in detail below.

[0025] (2) Configuration of AI agent generation server 200 Figure 2 shows an example of the configuration of the AI ​​agent generation server 200 (hereinafter referred to as the generation server 200) according to this embodiment. The generation server 200 includes a main memory 201, an auxiliary memory 202, a processor 203, an input device 204, an output device 205, and a communication control unit 206, and is connected to a network via the communication control unit 206.

[0026] The generation server 200 is equipped with a processor and a memory device, and as the core of the system 100, it implements the functional modules detailed below. The generation server 200 constructs an AI agent model using multifaceted personal information received from the administrator terminal 600. The generation server 200 is an example of the AI ​​agent generation means of the present invention.

[0027] <Acquisition Module 210> The main memory 201 is loaded with 10 functional modules, which are called by the processor 203 during execution. The acquisition module 210 is responsible for receiving multifaceted personal information transmitted from, for example, an external terminal and pre-processing it to store it in the auxiliary memory 202.

[0028] Multifaceted information includes the following: • Documents, drawings, and other information transmission media created by the original author (including papers, daily reports, reports, work deliverables, and dialogue messages). • The original person's conversation history (including speech history, interview information, audio of lectures and presentations, etc.) • The original person's activity history (including career history, work history, position, title, affiliated organizations, etc.) • Information indicating the original person's other intellectual tendencies (information representing thoughts, beliefs, and feelings) • Information regarding how others perceive the original artist. The acquisition module 210 receives this information, performs preprocessing such as formatting and tagging, and stores it in the auxiliary storage device 202. It can also handle real-time information acquired from sensor devices, etc., as needed. The acquisition module 210 is an example of the acquisition means of the present invention.

[0029] <Prompt memory module 211> The prompt memory module 211 records system prompts in chronological order in response to user statements, in order to improve the reproducibility of context in the AI ​​agent's response generation. This enables dialogue generation that takes into account the continuity and consistency of past interactions.

[0030] <Reflection generation module 212> The reflection generation module 212 is responsible for extracting thought processes and values ​​that contributed to the construction of the personality model from statements and records made by the AI ​​agent, and generating textualized reflective data (also called expert reflection or analytical reflection). This analytical reflection is stored as text data and used to refine the personality model and generate value vectors.

[0031] <Personality Model Generation Module 213> The personality model generation module 213 models the user's personality tendencies, judgment axes, and value structure as a machine learning model based on the above-mentioned analysis reflection and thought log information. This model is designed with a structure based on, for example, the Big Five theory or similar psychological structure theories, and influences the agent's behavior. The personality model generation module 213 is an example of the personality model generation means of the present invention.

[0032] <Personality Model Management Module 214> The personality model management module 214 is responsible for continuously updating and verifying the consistency of the generated personality model. It incorporates new user data and evaluates its consistency with past personality structures and whether there are any deviations. Based on the reactions and evaluations obtained from the user or the system, the personality model management module 214 is responsible for updating the personality model and strengthening its memory structure. This function forms the core for realizing the autonomous evolution and learning of system 100. The personality model management module 214 is an example of the personality model management means of the present invention.

[0033] The personality model management module 214 further detects self-contradictions in past output using the values ​​emphasized in past thought processes within the output history of the generative AI, and corrects inconsistencies in thought processes based on the detected analysis results. The personality model management module 214 statistically analyzes the personality models of multiple high-performance ghosts in order to extract the ideal thinking and behavioral patterns required for a specific role. Specifically, it extracts common values ​​and judgment patterns as vectors from the basic memory of the target group of ghosts, and identifies densely packed trends through clustering. The central vectors of the obtained clusters are defined as core personality elements, and a role model ghost is generated as a new personality model.

[0034] The personality model management module 214 further generates a report summarizing the output of the AI ​​agent, notifies the original user, interprets the feedback from the original user, and updates the personality model based on that feedback.

[0035] <Question generation module 215> The question generation module 215 is responsible for automatically generating questions to be presented to the user (including the original user) or the AI ​​agent, based on acquired information and a personality model. These questions are designed to deepen memory and promote the acquisition of new insights, and the style and vocabulary are optimized according to the original user's attributes and past dialogue history. The question generation module 215 may utilize a generative AI provided as an external service. For example, the question generation module 215 can function as an AI interviewer that acquires multifaceted information from the original user.

[0036] <Thought Process Construction Module 216> The thought process construction module 216 is a module that contextually integrates multiple elements, including question statements and reflective information, to construct an inference structure. As a result, responses from users and AI agents become logical and transparent, improving explainability. The thought process construction module 216 is an example of the thought process construction means of the present invention.

[0037] <Vectorization Module 217> The vectorization module 217 has the function of transforming and embedding recorded information as semantic vectors. This makes it possible to computationally compare each stored data and model component, and is used for similarity estimation and deviation detection processing. The vectorization module 217 may also utilize embedded models provided as an external service.

[0038] <Deviation Detection Module 218> The deviance detection module 218 is a function that evaluates the extent to which newly acquired behaviors, statements, etc., deviate from the original person by comparing the constructed personality model with a role model further constructed from the personality model. It is also possible to design the module to trigger a warning or reconstruction process if a deviation is detected. The deviance detection module 218 is an example of the deviance detection means of the present invention.

[0039] <Output Module> Output module 219 is a function that provides the constructed agent's response sentences, questions, analysis reflections, reports, etc., as text output or speech synthesis output to the user terminal. Tagging and classification are also performed automatically as needed, and it supports feedback to other modules.

[0040] The auxiliary storage device 202 holds information such as expert information 221, agent information 222, basic memory information 223, prompt configuration information 224, thought log information 225, and activity log information 226. Each piece of information is linked to the others by common agent identifiers, user identifiers, etc., enabling each module of the main storage device 201 to access the information consistently.

[0041] <Expert Information 221> Expert information 221 includes attribute information of real experts who are the target of learning thought processes, and stores data such as the expert's field of expertise, position, age, years of experience, and speech patterns. This information is used for role model generation and reflection extraction processing.

[0042] Expert information 221 is composed of, for example, an expert profile table. In the expert profile table, for example, a unique identifier called the expert ID is associated with the expert's name or identification name, field of expertise, role definition, representative analytical perspective, analytical style, usage prompt configuration ID, remarks, etc.

[0043] The name or identifier is personal identification information such as a name or nickname that corresponds to the expert ID, and includes examples such as "Psychologist A". A field of expertise refers to the academic area to which the expert belongs or specializes, and examples include psychology, management engineering, and behavioral economics. The role definition indicates the main purpose of the analysis that the expert will undertake, and examples include "deepening understanding of values" and "analysis of judgment tendencies." A representative analytical perspective is the viewpoint used when observing and evaluating the subject of analysis, and examples include "leadership" and "problem-solving orientation." The analytical style indicates the tendency of the analytical method, and includes information such as whether it is "inductive" or "deductive."

[0044] The prompt configuration ID is an identifier used to identify the configuration of the question template used by the expert in question. Notes may include supplementary information such as references related to the expert profile and external links to definitions. Expert information 221 is referenced in processing such as the reflection generation module 212 and the question generation module 215.

[0045] <Agent Information 222> Agent information 222 consists of metadata including the configuration status, reference history, personality model ID, and update history of each AI agent. This information is referenced by the personality model management module 214 and the vectorization module 217, etc., and is used to maintain model consistency and manage evolution.

[0046] Agent information 222 is composed of, for example, an agent information table. In the agent information table, for example, an agent ID, which is a unique identifier, is associated with an expert ID, generation date and time, personality type, status flag, etc.

[0047] The expert ID refers to the ID of the expert who served as the model for the agent in question. The creation date and time refers to the date and time when the agent was first created. Personality type refers to the classification of the personality model set for the agent (e.g., empathetic type, logical type, etc.). This is a systematic classification based on the values, judgment criteria, and response style that the agent exhibits in dialogue, and is used as an indicator to evaluate psychological compatibility and dialogue tendencies with the user. Personality types may include the following types, for example:

[0048] Specifically, these include the "empathetic type" (highly empathetic and supportive), the "logical type" (analytical and reason-driven), the "intuitive type" (creativity-driven and highly flexible), the "cautious type" (risk-averse and highly planned), the "challenging type" (goal-oriented and action-oriented), and the "harmonious type" (cooperative and considerate of others). These can also be defined in correspondence with classification systems in psychology such as MBTI (Myers-Briggs Type Indicator) and the Big Five theory.

[0049] This invention features a function that automatically assigns a personality type to each AI agent by analyzing expert reflections and characteristic words / value vectors contained in behavioral patterns in the base memory during agent generation. This enables a highly personalized dialogue environment that recommends and matches users with compatible AI agents according to their cognitive style and dialogue tendencies. Status flags are control information that indicates the status of the AI ​​agent, such as enabled / disabled, public / private, etc.

[0050] The agent information 222 may also include other data items such as update date and time, related tag information, reference frequency, and past dialogue count, and is referenced by the agent provision module 413 and the personality model management module 214.

[0051] <Basic Memory Information 223> The basic memory information 223 is a record that stores information used for learning, such as interviews with users or experts, evaluation statements, dialogue records, and document data, in chronological order, and each record is uniquely identified by a record ID. These records are used as source data for vector transformation and expert reflection generation.

[0052] The basic memory information 223 is composed of, for example, a basic memory table. In the basic memory table, for example, a unique identifier called a record ID is associated with data type, data source, expert persona ID, analysis text, vector representation, importance score, timestamp, etc.

[0053] The data type refers to the type of information that the record deals with, and is distinguished by formats such as "interview," "evaluation," and "document." The data source is information indicating the origin of the original data analyzed, and specifically includes file names, conversation IDs, document titles, etc. The Expert Persona ID is an identifier for the expert used in the analysis and is associated with Expert Information 221.

[0054] The analysis text area is where the main text of expert reflections extracted and edited by the reflection generation module 212, etc., is stored. The vector representation is the result of representing text data, such as the reflection text, as numerical vectors, and is generated by the vectorization module 217. The importance score is an indicator of the usefulness and reference priority of a record, and is reflected in the system's search and output order. A timestamp records the date and time the record was created or retrieved.

[0055] Furthermore, the basic memory information 223 can be referenced by multiple processing modules, such as the thought process construction module 216 and the deviation detection module 218. Also, a single memory record may be associated with multiple agents or prompt templates.

[0056] <Prompt configuration information 224> The prompt configuration information 224 stores various templates used in the question generation module 215 and the thought process construction module 216. This allows prompts suitable for the target audience, such as writing style, difficulty level, and response format, to be selected and generated.

[0057] The prompt configuration information 224 is composed of, for example, a prompt configuration table. In the prompt configuration table, information such as persona name, role definition, analysis perspective, thinking constraints, target data items, and output format is associated with a unique identifier called a prompt ID.

[0058] The persona name indicates the name of the hypothetical expert or analyst associated with the prompt, and symbolizes the personality and perspective of the generated agent. The role definition describes the purpose and function that the persona will fulfill (e.g., decision support, deepening understanding of values, etc.). The analytical perspectives represent the main viewpoints that the persona focuses on (e.g., logic, emotional elements, motivations for action, etc.).

[0059] Thought constraints describe the rules and limitations that a persona must adhere to in their thought process (e.g., maintaining neutrality, objective description, elimination of specific biases, etc.). The target data item indicates the type of item (e.g., interview data, work log, etc.) of the basic memory information 223 that the prompt references or analyzes. The output format indicates the style and format of the output generated based on the prompt (e.g., text format, summary format, Q&A format, etc.).

[0060] The prompt configuration information 224 is referenced in the reflection generation module 212 and the personality model generation module 213, ensuring consistency and expertise in the output. Furthermore, templates can be rearranged and reused to accommodate multiple personas and output needs.

[0061] <Thought Log Information 225> The thought log information 225 holds historical information about the thought paths and intermediate reasoning generated by each agent. The log information includes the input information used, the selected reasoning route, and the generated intermediate representation, and is used for verification and updating by the personality model management module 214.

[0062] The thought log information 225 is composed of, for example, a thought log table. In the thought log table, information such as trigger input, thought process, reference memory ID, output result, self-assessment metadata, and timestamp is associated with a unique log ID.

[0063] The trigger input refers to the user utterance or input request that triggered the recording of the log entry. The thought process is a chronological record of the reasoning and decision-making processes referenced at each step in the AI ​​agent's output generation.

[0064] The reference memory ID indicates the identifier of one of the multiple memory records referenced in the log (such as the basic memory information 223). The output refers to the final response result of the thought process, the generated text, etc. Self-assessment metadata is auxiliary information that indicates the reliability, confidence level, and ambiguity of the output at the time of generation.

[0065] The timestamp indicates the date and time the thought log was generated. The thought log information 225 is recorded by the thought process construction module 216 and used for visualizing the AI's thought trajectory and for continuous improvement.

[0066] <Activity Log Information 226> Activity log information 226 is a time-series record of the actions taken by the AI ​​agent during interactions with users and external collaborations. This includes response content, conversation partners, topics, and evaluation scores, and is used as basic information for model evaluation and deviation detection.

[0067] Activity log information 226 is composed of, for example, an activity log table. The activity log table associates information such as the executed action, output content, deviation detection result, system response, execution date and time, and executing user with a unique log ID.

[0068] The execution action indicates the processing performed when the log was recorded (e.g., response generation, external API call). The output content shows the processing results and response content obtained from the aforementioned action. The deviation detection results include whether or not the output content deviated from the predefined criteria (yes / no), and a score indicating the degree of the deviation.

[0069] The system response shows the corresponding information issued by the system in response to the action taken or deviation detected, such as warnings, processing stoppages, and notes. The execution date and time represent the time when the process was performed, enabling time-series monitoring. The executing user indicates the entity that performed the process, allowing for distinctions such as whether it was an automated response by an agent (ghost) or intervention by an administrator.

[0070] Activity log information 226 ensures transparency in the processing history of the AI ​​system and is used for subsequent audits, troubleshooting, and quality improvement. This improves the reliability and manageability of the AI ​​agent in this invention.

[0071] With the above configuration, the generation server 200 can acquire diverse user information and expert knowledge, generate and manage personality models and reflection data, and execute processes to construct and output thought processes. Note that the various functions of the generation server 200 may be implemented in a distributed manner across multiple terminals.

[0072] (3) Configuration of Marketplace Management Server 400 Figure 4 shows an example of the configuration of the marketplace management server 400 (hereinafter referred to as the management server 400). The management server 400 comprises a main memory 401, an auxiliary memory 402, a processor 403, an input device 404, an output device 405, and a communication control unit 406, and is connected to the network via the communication control unit 406.

[0073] The management server 400 is a server that executes marketplace functions related to the registration, provision, evaluation, and collaboration of AI agents, and handles anonymization, billing, revenue distribution, and collaboration with external services for each AI agent. It also performs processing to present and recommend appropriate AI agents based on user information 421, registration information 422, evaluation information 423, matching information 424, usage history information 425, etc., stored in the auxiliary storage device 402. This makes it possible to efficiently match user needs with the expertise of AI agents. The marketplace management server 400 is an example of the marketplace management means of the present invention.

[0074] <Agent Registration Module 411> The main memory 401 is loaded with six functional modules, which are called by the processor 403 during execution. The agent registration module 411 is a module that receives information on AI agents requested by agent providers and performs registration processing such as identification information, area of ​​expertise, and usage conditions.

[0075] <Anonymization Module 412> The anonymization module 412 is a module that performs anonymization processing, such as masking and statistical analysis of identification information and historical information, with respect to registered AI agents in order to protect personally identifiable information and sensitive information. The anonymization module 412 is an example of the anonymization means of the present invention.

[0076] <Agent-provided module 413> The agent provision module 413 is a module that, in response to a user request, refers to matching information and evaluation information stored in the auxiliary storage device 402, performs matching processing that meets the conditions, and is responsible for recommending and presenting an AI agent. The agent provision module 413 is an example of the matching means of the present invention.

[0077] Furthermore, in matching AI agents, matching may also be performed considering the compatibility between the user and the AI ​​agent. In this case, the agent provision module 413 pre-classifies and analyzes the user's personality type and the AI ​​agent's personality type, and then performs a process to recommend an AI agent with high psychological compatibility based on these combinations.

[0078] The user's personality type is automatically estimated based on interview responses, personality assessment questionnaires, and past task tendencies, adhering to psychological models such as MBTI, Big Five, and DISC. Similarly, the AI ​​agent is categorized using a similar schema based on its output tendencies and response characteristics derived from its personality model, providing a mutually matching structure for compatibility evaluation.

[0079] This type of compatibility-based matching supports psychological receptivity and trust building between the user and the AI ​​agent, which cannot be fully compensated for by task suitability alone. For example, recommending an analytical agent to a user who prefers logically rigorous responses, and a collaborative agent to a user who values ​​emotional empathy, reduces the stress of the conversation and allows for the creation of a more natural and constructive relationship.

[0080] Furthermore, compatibility data is continuously updated based on post-use feedback and response logs, allowing for dynamic evolution of personality classification and recommendation accuracy. Through this process, the AI ​​agent provisioning system of the present invention can provide an intelligent support environment optimized for the individual characteristics of each user.

[0081] Furthermore, the agent provision module 413 forms a virtual team consisting of multiple AI agents selected by the user, or a combination of AI agents and humans. Specifically, the agent provision module 413 configures a virtual team consisting of multiple ghosts or humans suitable for the task, based on information about the user's target task and instructions for forming a virtual team. Each selected ghost (or human) cooperates based on their respective expertise and thinking styles, engaging in dialogue and collaboration via a shared memory space. This allows the user to receive advanced simulations and decision-making support that incorporate insights from diverse perspectives for complex problems. In this case, the agent provision module 413 is an example of the agent provision means of the present invention.

[0082] <Charging Module 414> The billing module 414 is a module that calculates usage fees based on the user's use of the agent, processes billing, and manages settlement records, and is processed in conjunction with user information and usage history information 425. The billing module 414 further has the function of selecting and recommending a billing model to be applied when calculating usage fees, based on at least one of the attributes of the AI ​​agent and the user's desired use. For example, the billing module 414 refers to the profile data of the expert who served as the model for the AI ​​agent and recommends a billing model for the AI ​​agent, such as a pay-per-use, flat-rate, or performance-based fee. The billing module 414 may also suggest a flat-rate fee if the user uses the agent as a consultant for daily work, and a performance-based fee if the user uses it to create presentation materials. The billing module 414 is an example of a billing means of the present invention.

[0083] <Revenue Sharing Module 415> The revenue distribution module 415 is a module that executes a logic for distributing revenue according to the agent's service provision status, usage frequency, evaluation results, etc., and calculates and records the amount of compensation for each provider. The revenue distribution module 415 is an example of the revenue distribution means of the present invention.

[0084] The auxiliary storage device 402 primarily stores user information 421, agent registration information 422, evaluation information 423, matching information 424, and usage history information 425. This information is used to enable appropriate relationship building and operational management between agents and users on the marketplace.

[0085] <User Information 421> User information 421 includes basic attribute information about the individual or organization using the marketplace. Based on this information, the system controls the presentation of suitable agents and the provision of services to the user.

[0086] User information 421 is composed of, for example, a user profile table. The user profile table associates, for example, a unique identifier called a user ID with information such as name, age, affiliated organization, start date of use, permission level, usage status, and personality type.

[0087] The name refers to the individual's name corresponding to the user ID. The age is the user's age information and is used for service control and analysis based on age group.

[0088] The "Affiliated Organization" information indicates the organizational unit to which the user belongs, such as a company, group, or team. This information is used to understand usage patterns and manage operations for each organizational unit.

[0089] The start date of use indicates the date the user began using the system and is used for usage history analysis and license management. The permission level indicates the access restrictions and scope of operations granted to the user.

[0090] The usage status indicates the validity (e.g., active / inactive) and status (e.g., active / deactivated) of the user account. Note that user information 421 may include information items other than those listed above.

[0091] Personality type refers to a classification of the user's personality (e.g., empathetic type, logical type, etc.). This is a systematic classification based on the values, judgment criteria, and response style that the user demonstrates in conversations with others, and is used as an indicator to evaluate psychological compatibility and conversational tendencies with the AI ​​agent. Personality types may include the following types, for example:

[0092] Specifically, these include the "empathetic type" (highly empathetic and supportive), the "logical type" (analytical and reason-driven), the "intuitive type" (creativity-driven and highly flexible), the "cautious type" (risk-averse and highly planned), the "challenging type" (goal-oriented and action-oriented), and the "harmonious type" (cooperative and considerate of others). These can also be defined in correspondence with classification systems in psychology such as MBTI (Myers-Briggs Type Indicator) and the Big Five theory.

[0093] This invention includes a mechanism that, upon new user registration, acquires information from responses to an optional questionnaire, obtains information that can classify the user's personality, and automatically assigns a personality type to each user. This enables the AI ​​agent to recommend and match the user with an AI agent that is a good match for the user, based on the cognitive style and conversational tendencies acquired by the AI ​​agent.

[0094] <Agent Registration Information 422> Agent registration information 422 is information used for registering and managing AI agents provided on the marketplace. It is used to understand the characteristics and usage conditions of registered agents, and contributes to improving agent visibility and searchability.

[0095] Registration information 422 consists of, for example, a market registration agent information table. This table records information such as agent ID, origin, publication status, tag information, intended use, output format, usage conditions, billing model, and cumulative performance, linked to a unique identifier called the registration agent ID.

[0096] The Agent ID is the identifier in Agent Information 222 for the registered agent in question.

[0097] The origin is information that identifies whether the agent was generated using a real-world expert or user as a model, or whether it is based on an existing role model.

[0098] The public status indicates the agent's disclosure status on the market (public / private / limited access, etc.).

[0099] Tag information consists of keywords that represent the characteristics and tendencies of agents (e.g., leader type, conservative orientation, etc.), and serves as a clue for users when searching or selecting agents.

[0100] The intended use represents the main usage scenario for the agent (e.g., business support, consulting, educational purposes, etc.).

[0101] The usage conditions include restrictions on the terms of service, such as the available time slots and API call limits.

[0102] The billing model refers to the fee structure for using the agent (e.g., flat rate, usage-based, performance-based).

[0103] Cumulative performance includes information about the agent's operational history, such as the number of times it has been used in the past, the total number of responses, and the last update date. Note that registration information 422 may also include other items.

[0104] <Evaluation Information 423> Evaluation information 423 includes subjective feedback and satisfaction levels from users when using the agent. This information is used to improve agent quality and recommendation accuracy.

[0105] The evaluation information 423 is composed of, for example, a review information table. This table records items such as the user log ID, reviewer ID, evaluation score, review text, and review date and time, linked to a unique identifier called the review ID.

[0106] The usage log ID is an identifier used to identify the agent usage history targeted by the review. The reviewer ID is information that uniquely identifies the user who posted the review.

[0107] The evaluation score is a numerical representation, for example, on a 5-point scale (1-5), of the overall assessment of the agent being reviewed and the content of their response. The review text is a free-form section where reviewers can describe the reasons for their evaluation, specific opinions, suggestions, etc.

[0108] The review date and time indicates the date and time the review was recorded. Note that evaluation information 423 may include other evaluation metrics, tags, administrator notes, etc.

[0109] <Matching Information 424> Matching Information 424 is a record of how user needs match the services offered by agents. It is used to support agent recommendations that align with user intentions.

[0110] Matching information 424 is composed of, for example, a matching information table. This table uses a unique matching ID as the primary key and records information such as agent ID, skill / domain tag, intended use, price range, available time slots, evaluation score / number of reviews, and visibility range in association with it.

[0111] The Agent ID is information that uniquely identifies the agent corresponding to the matching criteria. Skill / Domain tags are tag information that represents areas of expertise or areas of knowledge the agent can handle.

[0112] The intended use is information that represents the purpose for which matching is expected (e.g., consultation, design, business support, etc.) and is used to match it with the user's search needs. The price range and available time slots are parameters that indicate the agent's usage conditions.

[0113] The evaluation score / number of reviews represents an evaluation value based on the agent's past usage history and is used as information to improve matching accuracy and assess reliability. The visibility range is an attribute used to manage the public status of the matching information (e.g., public / limited access / private).

[0114] <Usage History Information 425> Usage history information 425 is a record of the actual interactions and usage history between users and agents. It is used to understand the overall operational status of the marketplace, to handle individual problems, and to analyze usage trends.

[0115] Usage history information 425 is composed of, for example, a usage log table. This table uses a unique usage log ID as its primary key and stores associated information such as user ID, agent ID, start date / end date / time of use, billing type, usage fee, and payment status.

[0116] The User ID is an identifier used to uniquely identify the user who used the AI ​​agent in the relevant usage history. The Agent ID is information that identifies the AI ​​agent that was used.

[0117] The start and end dates represent the service provision period for the AI ​​agent and are used to analyze actual usage time and duration. The billing type is an attribute that identifies the fee structure for the usage (e.g., flat rate, performance-based).

[0118] The usage fee records the amount charged for the usage in question and is used as basic data for billing and revenue sharing processing. The payment status indicates the status of the billing process (e.g., completed, pending) and is used for billing / payment management.

[0119] As described above, the management server 400 can comprehensively perform a series of management functions, including agent registration, provision, evaluation, matching, and billing. This enables the provision of an efficient and fair trading environment for users and agents, and promotes the optimal use of AI agents tailored to individual needs.

[0120] (4) Configuration of Administrator Terminal 600 Figure 6 shows an example of the configuration of the administrator terminal 600 according to this embodiment. The administrator terminal 600 communicates with the generation server 200 or the management server 400 via the network and provides an operating environment for operating and maintaining the overall system.

[0121] The administrator terminal 600 is an information processing device used by an administrator who operates and maintains the system 100. The administrator terminal 600 comprises a processor 603, main memory 601, auxiliary memory 602, input device 604, output device 605, camera 606, and communication control unit 607. The administrator terminal 600 runs a management application that remotely controls the generation server 200, and can comprehensively manage the entire AI agent generation process by selecting the target expert, specifying prompt configurations, registering training data, confirming and correcting the content of the generated AI agent, and providing feedback on the output results.

[0122] The main memory 601 is loaded with the information acquisition module 611, the server communication module 612, and the display module 613, which are then executed by the processor 603. The information acquisition module 611 acquires various types of information necessary for AI agent generation, such as expert profiles, references, business interview results, and report contents, and stores them in the auxiliary memory 602 after performing preprocessing such as format conversion and tagging as needed. This enables integrated and efficient data utilization in subsequent processing.

[0123] The server integration module 612 communicates with the generation server 200 and the management server 400 via API, and causes the generation server 200 and the management server 400 to execute their respective processes, as described later. The server integration module 612 maintains bidirectional communication at all times and can immediately notify the administrator terminal 600 of events occurring on each server.

[0124] The display module 613 visualizes acquired information as a GUI and provides an intuitive setting change interface using click operations and drag-and-drop. It also overlays the video feed from camera 606 as needed.

[0125] The auxiliary storage device 602 stores the administrator client application 621. The application 621 includes user interface resources, theme files, cache data, and configuration files.

[0126] With the above configuration, the administrator can operate the administrator terminal 600 to integrally control the system 100, which consists of multiple server devices, and perform tasks such as generating, registering, setting the scope of publication, monitoring usage, and adjusting output quality of AI agents that reflect the expertise and thought processes of specialists.

[0127] (5) Configuration of user terminal 700 Figure 7 shows an example of the configuration of a user terminal 700 according to this embodiment. The user terminal 700 is mainly used when communicating with the generation server 200 or the management server 400 via the network and using the AI ​​agent provided by the management server 400.

[0128] The user terminal 700 is, for example, a general-purpose information processing device and can take various forms such as desktop, notebook, tablet, or smartphone. In the system 100 according to this embodiment, the user terminal 700 is a terminal operated by a user who uses an AI agent registered with the management server 400. The user terminal 700 includes a main memory 701, an auxiliary memory 702, a processor 703, an input device 704 (keyboard, touch panel, microphone, etc.), an output device 705 (display, speaker, etc.), a camera 706, and a communication control unit 707. The communication control unit 707 can connect to any network, such as a wireless / wired LAN. It is configured to connect to other devices via the network. Through the user terminal 700, the user can use an AI agent registered in the marketplace to interact with the AI ​​agent and obtain information. The user can also have the AI ​​agent perform various information processing tasks through the user terminal 700.

[0129] The main memory 701 is loaded with the information acquisition module 711, the server cooperation module 712, and the display module 713, which are then executed by the processor 703. The information acquisition module 711 is responsible for acquiring interaction information such as user speech and input operations at the user terminal 700. For example, it acquires language information, facial expressions, and gestures spoken by the user using the input device 704 or camera 706, converts them into an appropriate format, and stores them in the auxiliary memory 702. This enables the AI ​​agent to control its response according to the user's requests and state.

[0130] The server integration module 712 is responsible for sending and receiving information between the user terminal 700 and the generation server 200 and management server 400. This module transmits interaction information and usage requests obtained from the user to the server side in a predetermined format, and also passes response results and agent information received from each server to other modules on the main memory 701. This allows the user to smoothly utilize the AI ​​agents registered on the server.

[0131] The display module 713 is responsible for displaying the AI ​​agent's response results, interview content, dialogue logs, recommendation information, and other information in a visually accessible format on the user terminal 700. The display module 713 works in conjunction with the output device 705 to appropriately lay out and present text information, charts, and highlights on the screen, thereby supporting user understanding and operation. It can also dynamically switch the displayed content according to the user's operation history and browsing status.

[0132] The auxiliary storage device 702 stores the client application 721. The application 721 includes user interface resources, theme files, cache data, and user configuration files.

[0133] In this way, the user terminal 700 provides an operating environment for the AI ​​agent's user to receive dialogue and business support, and through cooperation with various server devices, it can smoothly receive response results, display recommendations, and input evaluations. The user terminal 700 functions as a central interface that realizes a dialogue experience utilizing the AI ​​agent's knowledge, while coordinating with the generation server 200 and the management server 400.

[0134] (6) Regarding the processing of system 100 Next, referring to the diagram, the processes performed in system 100 will be described in order.

[0135] (6-1) Interview Processing Figure 8 is a flowchart showing an example of the interview flow 800 according to this embodiment. In this flow 800, the generation server 200 presents questions based on the interview script and obtains multifaceted information from the original interviewee. This process is mainly performed by the AI ​​interviewer, which is the question generation module 215 of the generation server 200. The details of each step will be explained below.

[0136] First, in step S810, the generation server 200 loads the interview script. Specifically, the question generation module 215 retrieves a question script corresponding to the model from the auxiliary storage device 202 and initializes the question structure and follow-up branching conditions. This script is structured based on the prompt configuration information 224 and includes the design of how the AI ​​agent asks questions, their order, and how to guide the discussion in a way that suits the purpose.

[0137] Next, in step S820, the generation server 200 presents the questions. Based on the loaded script, the question generation module 215 presents the interview questions visually or audibly on the user terminal 700 used by the original person. In the case of audible presentation, it is also possible to configure the system to cooperate with an external API that performs TTS processing.

[0138] Next, in step S830, the generation server 200 analyzes the response content. The reflection generation module 212 performs natural language analysis on the user response obtained via the acquisition module 210 using LLM or similar methods to extract intent, perspective, and topic. If necessary, it refers to past response history and thought log information 225 to supplement the consistency and logic of the response context.

[0139] Next, in step S840, the generation server 200 determines whether the question's objective has been achieved. The question generation module 215 compares the response analysis results with the script structure to determine whether the current question has achieved its intended objective. For example, in the case of a question about values, it detects whether it includes sufficient introspection or the presentation of decision-making criteria.

[0140] Next, in step S850, the flow is branched based on whether the objective has been achieved. If it has not been achieved, the process proceeds to step S860, where additional questions are generated. If it has been achieved, the process proceeds to step S870, where the next question script is introduced.

[0141] In step S860, the generation server 200 dynamically generates follow-up questions. The question generation module 215 immediately generates supplementary or in-depth questions based on the response context and script structure and presents them to the user. This enables flexible interview development that is not dependent on a fixed script.

[0142] Next, in step S870, the generation server 200 transitions to the next question. The question generation module 215 calls the next question node according to a predefined script sequence and proceeds to a similar presentation process (S820). The consistency of the main script structure is maintained even in branching via follow-up.

[0143] Next, in step S880, the interview completion is determined. If all question nodes have been answered, or if the termination conditions in the script are met, the process proceeds to step S890. If not completed, the process returns to step S820.

[0144] Finally, in step S890, the generation server 200 records the response log and interview transcript. The acquisition module 210 records a complete transcript, including all of the original person's utterances and their presentation order, as well as the system's presentation log, and stores it in the basic memory information 223. This makes it available for use as the AI ​​agent's thought log information 225 and as basic information for memory reuse. Then, based on the acquired multifaceted information, a personality model modeled after the original person is constructed and stored in the model DB 800. This executes a personality model generation step that generates an AI agent personality model that mimics the thought patterns and behavioral patterns of the original person, based on the multifaceted data acquired regarding the individual's characteristics.

[0145] (6-2) First prompt creation process Figure 9 is a flowchart showing an example of the personality analysis flow 900 according to this embodiment. In this flow 900, the generation server 200 acquires self-perception and evaluation data from the constructed AI, quantitatively analyzes the differences (gaps) between them, and generates and stores introspective reflections of personality characteristics. This process is mainly performed by the acquisition module 210, the vectorization module 217, the reflection generation module 212, and the personality model management module 214. The details of each step will be described below.

[0146] First, in step S910, the generation server 200 acquires self-recognition data and peer evaluation data. Specifically, the acquisition module 210 identifies user input (e.g., self-evaluation, questionnaires) and evaluation data from others (e.g., reviews, feedback logs), determines their source and time series, and then separates and acquires them. The data to be acquired is linked to user information and usage history information, and the auxiliary storage device 202 is used as the storage source.

[0147] Next, in step S920, the generation server 200 performs vectorization of each data point and generates centroid vectors. Specifically, the vectorization module 217 encodes the self-recognition text and the other-evaluation text using a natural language processing engine, representing them as vectors that retain contextual meaning. Subsequently, the centroid vectors of the self-vector and the other-vector are calculated using a weighted average or geometric center across multiple data points.

[0148] Next, in step S930, the generation server 200 calculates the difference vector (gap vector). The vectorization module 217 calculates the difference vector between the obtained self-centroid vector and the other-centroid vector, and generates a gap vector that characterizes the direction and deviation. The difference vector serves as the basis for interpretation and reflection generation in the next step.

[0149] Next, in step S940, the generation server 200 generates a prompt containing a gap summary and analysis instructions. The reflection generation module 212 automatically classifies the direction and domain of the divergence vector (e.g., assertiveness, cooperativeness, etc.) and matches it with the corresponding template to dynamically construct a prompt for reflection generation. The prompt is assigned numerical information and semantic labels.

[0150] Next, in step S950, the generation server 200 generates gap analysis reflections using the LLM. The reflection generation module 212 inputs prompts to the LLM and generates reflection statements aimed at interpreting the differences between oneself and others, analyzing factors, and eliciting awareness. These results are available as user feedback and are also passed to the subsequent memory structuring process.

[0151] Finally, in step S960, the generation server 200 structures the analysis results and supporting data and stores them as basic memory information 223. The personality model management module 214 generates and stores a personality analysis record that combines reflection text, difference vectors, original data IDs, etc. This enables personalized support based on the user's personality structure and tracking of changes over time.

[0152] (6-3) Formation of the feedback loop Figure 10 is a flowchart showing an example of the feedback loop formation flow 1000 according to this embodiment. In this flow 1000, the generation server 200 generates a summary report based on past activity records (thought logs and response history) by the AI ​​agent. Then, after collecting evaluations and feedback from the original user, the process updates the corresponding personality model or base memory. Each step is executed collaboratively by the reflection generation module 212, the personality model management module 214, etc. The details of each step will be explained below.

[0153] First, in step S1010, the generation server 200 performs trigger detection. Triggers include not only periodic execution after a predetermined period (e.g., one week), but also event-driven triggers that are triggered by the occurrence of specific events (e.g., continuous use, changes in evaluation score, external actions). This allows for flexible adjustment of the feedback generation timing according to the user's usage.

[0154] Next, in step S1020, the generation server 200 collects thought logs and memory information to be used for feedback. Specifically, it extracts response history, decision reasons, and meta-evaluations (confidence level, style, etc.) recorded over a certain period of time from the thought log information 225 and the basic memory information 223.

[0155] Furthermore, the generation server 200 may perform screening of the log memory to be collected based on importance scores. This ensures that records of judgments and thoughts essential to the user are prioritized for feedback. The score is determined based on importance labels and evaluation scores assigned during past conversations.

[0156] Next, in step S1030, the generation server 200 uses an LLM (Large-Scale Language Model) to generate a summary report based on the activity log. The generated report describes recent trends, highlights to emphasize, and characteristic decision points in natural language. The report may be adjusted in style and vocabulary to aid in the user's understanding.

[0157] Next, in step S1040, the generation server 200 presents the generated report to the original user and accepts feedback input in either a free-text or selection format. This process may be performed asynchronously in conjunction with the UI mechanism (such as a management app or web screen), and the user's response is returned to the server with their ID.

[0158] Next, in step S1050, the generation server 200 interprets and structures the response (feedback input) from the original person and generates update commands for the base memory or personality model. Specifically, the reflection generation module 212 determines the intent of the text, the tendency towards affirmation / negation, and the points that need revision, and extracts update options such as resetting importance, requesting deletion, and adding new perspectives.

[0159] Next, in step S1060, the generation server 200 updates the contents of the base memory based on the update command. The personality model management module 214 adjusts the importance score of existing records and adds new reflection logs as new records. Version control information (updator, timestamp, operation history, etc.) is also added as needed.

[0160] Furthermore, the generation server 200 may perform consistency checks on the updated personality model or base memory information 223. Specifically, it can detect unintended inconsistencies, extreme biases, excesses or deficiencies in skill tags, etc., and generate warning logs to ensure the consistency of sustained self-growth.

[0161] Through the above process, the feedback loop formation flow 1000 can understand the discrepancy between the AI ​​agent's activities (responses) and thoughts and its self-assessment, and reflect this in the recorded information, thereby improving the AI ​​agent's adaptability to the original user. In particular, the proactive intervention of the generation server 200 establishes a continuous and individually optimized growth support cycle.

[0162] Thus, the personality model management module 214 according to the present invention periodically summarizes the ghost's activities, notifies the original person of its output and thought process, and accepts responses, thereby forming a feedback loop. This process not only contributes to updating the model, but also provides the original person with an opportunity to objectively observe their own thinking style and judgment tendencies. In particular, by confirming the logical development and the basis for value judgments output by the ghost, the original person can see a structured reproduction of their own thinking.

[0163] The formation of such a feedback loop has the secondary effect of prompting new insights and introspection in the original individual. For example, the judgment patterns and value priorities that the original individual unconsciously employed can be visualized through the output of the ghost, leading to discoveries such as, "I was prioritizing this perspective," or "I was reaching conclusions through this thought process." This kind of intellectual feedback also contributes to the personal growth and evolution of the original individual's thinking.

[0164] Furthermore, through feedback from the original individual, the personality model is updated, allowing the ghost not only to maintain loyalty but also to build a relationship with the original individual where their thinking style and values ​​evolve together. In other words, the AI ​​agent functions not only as a mirror image of the original individual but also as an intellectual partner that deepens self-understanding. In these multifaceted aspects, the formation of a feedback loop provides an extremely meaningful intellectual support interface.

[0165] (6-4) Thought process construction process Figure 11 is a flowchart showing an example of the thought process construction flow 1100 according to this embodiment. This flow 1100 is a process that generates a response based on step-by-step thinking by utilizing relevant knowledge stored in the base memory in response to a question from a user using the AI ​​agent. This process is mainly performed by the vectorization module 217, the thought process construction module 216, and the personality model management module 214 of the generation server 200. The details of each step are described below.

[0166] First, in step S1110, the generation server 200 vectorizes the question text input by the user. Specifically, the vectorization module 217 acquires the question text and applies a natural language embedding model (e.g., text-embedding-3) to generate an internal representation as a semantic vector. The generated vector is stored in temporary memory for use in subsequent search processing.

[0167] Next, in step S1120, the generation server 200 retrieves relevant information from the base memory table by vector search. Specifically, the personality model management module 214 evaluates the similarity between the question vector and the vector of each record in the base memory and extracts the top candidate records. Subsequently, a re-ranking process organizes the information considering contextual consistency and redundancy, etc., to form the final set of contextual candidates.

[0168] Next, in step S1130, the generation server 200 constructs the final prompt. Specifically, the thinking process construction module 216 formats the set of candidate contexts as context information and generates a prompt to which system instructions, role definitions, thinking instructions, etc., are added. This prompt has an explicit thinking framework embedded in it for performing step-by-step reasoning in response to the question.

[0169] Next, in step S1140, the generation server 200 performs response generation by the AI ​​agent. Specifically, the thinking process construction module 216 inputs the final prompt to the AI ​​agent to execute the response, generating a natural language response that includes the thought process. This response includes the reasoning process and basis for judgment in response to the user's question, which enhances transparency and credibility.

[0170] Finally, in step S1150, the generation server 200 presents the response generated by the AI ​​agent to the user. The response can include, if necessary, the thought process (intermediate reasoning and cited evidence). At this time, the display module 713 of the user terminal 700 outputs the response and its background information in a highly visible format. This makes it easier for the user to understand not only the conclusion but also the derivation process. Through the above processing, the present invention realizes highly transparent decision support.

[0171] (6-5) External integration, action execution process Figure 12 is a flowchart showing an example of an external collaboration flow 1200 according to this embodiment. This flow 1200 is a process in which an AI agent formulates an execution plan based on relevant knowledge, using a goal presented by the user in natural language, and performs actions in collaboration with an external system. This process is executed by an AI agent that has been instructed by the user to perform a task. The details of each step are described below.

[0172] First, in step S1210, the AI ​​agent receives a goal statement in natural language from the user (e.g., "Schedule a team meeting for next week"). The input is received via a GUI or voice interface, converted to text, and then the AI ​​agent analyzes its intent and content. This allows for goal-oriented semantic interpretation, while supplementing ambiguous expressions and abbreviations.

[0173] Next, in step S1220, the AI ​​agent interprets the content of the acquired goal statement and searches and extracts related knowledge from the underlying memory database. Specifically, the AI ​​agent searches for similar past goals, business documents, operation history, etc., and performs a re-ranking based on a relevance score, taking into account user characteristics and frequency of use.

[0174] Next, in step S1230, the AI ​​agent generates an execution plan consisting of a series of subtasks necessary to achieve the goal, based on the search results from the previous step. The AI ​​agent creates a structured plan that takes into account the dependencies between tasks, referring to typical patterns and predefined templates. The plan includes information such as the required action types, execution order, and candidate tools to be used.

[0175] Next, in step S1240, the AI ​​agent determines the next action to take based on the current task and selects the corresponding external tool. The AI ​​agent refers to the open API specifications and tool metadata (argument format, limitations, authentication requirements, etc.) to verify the suitability and availability of the tool, and then determines which service to use (e.g., calendar API, messaging bot, etc.).

[0176] Next, in step S1250, the AI ​​agent invokes the tool selected in the previous step to perform an action. The AI ​​agent converts the input parameters (date and time, destination, body, etc.) constructed by LLM into the tool's API format and securely performs the integration process. The processing results are received in the form of a status code, response body, etc.

[0177] Next, in step S1260, the AI ​​agent observes the execution results and evaluates any deviations from the plan. The LLM compares the response content with the goal statement and makes a comprehensive determination of whether the task was completed, whether it differed from expectations, and whether there were any error messages. If an anomaly is detected, the cause (e.g., authentication error, scheduling conflict) is extracted and used for subsequent actions.

[0178] Next, in step S1270, if there is a discrepancy, the AI ​​agent dynamically modifies the execution plan. The AI ​​agent breaks down the goal again as needed and performs replanning, which may involve selecting new tools or changing their order. After replanning, the process returns to steps S1240 and beyond.

[0179] Finally, in step S1280, if it is determined that the goal has been achieved, the AI ​​agent notifies the user accordingly and terminates the process. If the action repeatedly fails or is interrupted, the system may be configured to assign a status code of "abnormal termination" and return a message prompting the user to take action.

[0180] (6-6) Reference vector generation flow Figure 13 is a flowchart showing an example of the reference vector generation flow 1300 according to this embodiment. This flow 1300 is a process in the generation server 200 that generates vector information (hereinafter referred to as the reference vector) that will serve as the baseline for a personality model that reflects the values ​​of the original person. This process aims to capture multifaceted personality characteristics by combining the text contained in the question answer log with multiple expert reflections (reflection statements). The details of each step will be explained below.

[0181] First, in step S1310, the generation server 200 extracts text related to values. Specifically, the reflection generation module 212 extracts utterances (values ​​text) related to an individual's value judgments and decision-making tendencies from the response sentences corresponding to the interview question log. The selection of text to be extracted can be based on whether it contains words related to a pre-designed evaluation perspective or label set (e.g., integrity, cooperativeness, etc.).

[0182] Next, in step S1320, the generation server 200 vectorizes the extracted text and generates a baseline vector V_baseline. Specifically, the content of the text is input into a natural language processing model and converted into an embedding vector that reflects the context. At this time, the reflection generation module 212 obtains highly accurate representation vectors by using a pre-trained model (e.g., a domain-specific LLM) that is tailored to the individual's speech style and term preferences. It is also desirable to perform preprocessing such as removing unnecessary words and formatting (normalizing) the sentences before vectorization.

[0183] Next, in step S1330, the generation server 200 obtains multiple expert reflection vectors. Each expert persona has a different analytical perspective (e.g., values ​​analysis, leadership characteristics analysis, etc.), and the resulting reflection vectors reflect a variety of aspects. These vectors are automatically generated according to the prompt configuration information 224 and are stored along with an identifiable ID for each expert.

[0184] Finally, in step S1340, the generation server 200 calculates a weighted average to correct and save V_baseline. Specifically, it takes into account the confidence and expertise scores (meta-information) of each reflection vector as weights and integrates them into a single representative vector using a weighted average. The resulting baseline vector V_baseline is saved in the ghost profile as the initial state of the personality model and is used as a reference point for subsequent thought log correction and agent generation. Furthermore, explainability can be ensured by adding the original question item ID and extracted label as metadata during saving.

[0185] (6-7) Deviance detection process Figure 14 is a flowchart showing an example of the deviation rights flow 1400 according to this embodiment. This flow 1400 is a process in the generation server 200 that evaluates whether the content of the responses and actions presented by the personality model (hereinafter referred to as "ghost") deviates from the ghost's baseline vector (V_baseline), and records and notifies as necessary. This process is an auxiliary feedback mechanism to ensure the stability and consistency of the ghost's responses, and aims to improve the reliability and explainability of its operation. The details of each step will be explained below.

[0186] First, in step S1410, the generation server 200 receives the output results from the ghost and performs a process to convert the response sentences and actions into vector format. Specifically, it converts natural language text or metadata obtained from the dialogue log or action log into an embedded vector using a predetermined natural language processing engine to obtain the output vector V_generated. This vector reflects the contextual meaning of the utterance and the intent of the action.

[0187] Next, in step S1420, the deviation detection module 218 of the generation server 200 calculates the distance between the output vector V_generated and the reference vector V_baseline held by the ghost. Indicators such as cosine similarity or Euclidean distance are used to calculate the distance, quantitatively measuring the semantic discrepancy between the two vectors. If the output is action metadata, the calculation is performed via a transformation model in which the correspondence between action attributes and semantic vectors is defined in advance.

[0188] Next, in steps S1430 and S1440, the generation server 200 determines whether the distance value exceeds a set threshold. The threshold is personalized based on each ghost's personality traits, purpose, user settings, etc. For example, a lower threshold may be applied to an agent for medical use that should avoid emotional speech. In this step, if the deviation is less than the threshold, it is considered a normal output and the process proceeds to the next step (step S1450).

[0189] On the other hand, if it is determined that the deviation exceeds a threshold, in step S1460, the generation server 200 identifies the output as a "deviation action" and takes corrective action. Specifically, it can take selective actions such as outputting a warning, pausing the utterance, or summarizing and re-presenting the content. In step S1470, the deviation is recorded in a log and used in subsequent relearning processes and ghost evaluation processes. The record includes metadata such as the output vector, distance value, corresponding label, and utterance ID.

[0190] Furthermore, this process helps the ghost's personality model maintain consistent behavior over the long term by detecting and addressing deviations. The recorded deviation logs are also viewable from the administrator terminal 600, allowing for annotation and correction instructions to be given to the entire dialogue history as needed. This serves as a supporting foundation for enhancing the explainability and transparency of the ghost.

[0191] (6-8) Agent registration and publication process Figure 15 is a flowchart showing an example of the agent registration and publication flow 1500 according to this embodiment. This flow 1500 is a series of processes that registers the personality model (agent) generated in the generation server 200 on the publication platform (marketplace that provides AI agents) after undergoing predetermined review and anonymization processing. This makes the generated agent available on the marketplace for use by other users. The details of each step will be explained below.

[0192] First, in step S1510, the administrator applies to the generation server 200 for agent registration via the administrator terminal 600. The application is made after the agent is completed and has undergone preliminary review and self-evaluation. Meta information such as the agent ID, purpose of creation, and usage restrictions are attached to the registration application. Administrator privileges are required to apply, and the system is designed so that individual users cannot publicly register agents. In the second embodiment described later, applications for AI agent registration may also be made by individual users.

[0193] Next, in step S1520, the generation server 200 automatically extracts personally identifiable information from the base memory and profile information. The base memory stores question-answer logs, behavioral patterns, and preference vectors used during agent generation, and attributes that constitute personal information, such as name, date of birth, affiliated organization, and usage history, are extracted from these. The extraction results are used as basic data for anonymization processing.

[0194] Next, in step S1530, the generation server 200 performs anonymization processing on the extracted identification information. This processing includes masking information, grouping attributes (e.g., age range notation), and conversion to unique identifiers. Furthermore, profile information containing highly personally identifiable utterances or extreme preferences is either set to private or adjusted according to editing guidelines. This reduces the risk of identifying the original person from the personality model while enabling the provision of practical information.

[0195] Finally, in step S1540, as part of the marketplace management process, agents are reviewed for publication eligibility, and approved agents are registered on the marketplace. During registration, evaluation information, expected usage scenarios, dialogue samples, pricing, and usage conditions are added, making them searchable, comparable, and usable by users. The registration information is synchronized with the database on the management server 400, and usage history and reviews are continuously accumulated.

[0196] (6-9) Agent search, contract, and service provision flow processing Figure 16 is a flowchart showing an example of the agent search and provision flow 1600 according to this embodiment. This flow 1600 is a series of processes in which the management server 400 searches for and selects an appropriate personality model (agent) based on the user's request and initiates the provision of a session for dialogue or business collaboration. The purpose of this process is to achieve optimal matching between the user and the agent and to improve utilization efficiency and satisfaction. The details of each step will be explained below.

[0197] First, in step S1610, the user operating the user terminal 700 enters search criteria such as purpose, skills, target domain, and desired style, and performs an agent search. The agent provision module 413 of the management server 400 extracts a list of candidates from the published model DB 800 based on the specified search criteria. Search criteria include free word searches, tag specifications, and the use of purpose-specific templates.

[0198] Next, in step S1620, the agent provision module 413 of the management server 400 ranks the extracted candidates based on tag information, skill descriptions, review content, evaluation scores, etc., and presents a group of preferred candidates. At this time, past usage history and review trends recorded in evaluation information 423 and matching information 424 are referenced, and scoring is performed based on quantitative reliability indicators. If a recommendation mechanism based on the user's usage history is incorporated, individual optimization is taken into consideration. In addition, the agent provision module 413 may set the recommendation ranking based on the compatibility of the user's personality with that of the AI ​​agent, based on the personality types that are automatically or manually set for the user and the AI ​​agent, respectively.

[0199] Next, in step S1630, the agent provision module 413 of the management server 400 displays matching candidate information on the screen based on the candidate list above, and accepts the user's selection and usage agreement. The agreement includes confirmation of the usage conditions (fee structure, usage time, scope, etc.) defined for each agent, and session preparation begins after agreement. The contract information is recorded as usage history information 425 and is used for subsequent billing processing and reward distribution.

[0200] Finally, in step S1640, the agent provision module 413 of the management server 400 initiates a dialogue session or business processing session between the selected agent and the user. During the session, real-time dialogue takes place between the user terminal 700 and the agent processing module, and auxiliary processes such as logging, additional agent recommendation, and prompt regeneration are performed as needed. At the end of the session, the user is prompted to input an evaluation of their usage, which will be used to improve matching accuracy in the future.

[0201] (6-10) Billing and revenue sharing processing Figure 17 is a flowchart showing an example of the usage status recording, billing, and revenue distribution flow 1700 according to this embodiment. This flow 1700 is a series of processes in which the management server 400 records the usage status of agents, performs a predetermined billing process for users, and then distributes the revenue to the agent owners, etc. This process aims to ensure a sustainable economic cycle based on the circulation of agents and to enhance the value of utilizing personality models as intellectual assets. The details of each step will be explained below.

[0202] First, in step S1710, the agent provision module 413 of the management server 400 monitors and records the agent's usage history. Specifically, it acquires usage metrics such as the number of API calls by users, the number of interactions, the number of actions performed, and the time spent, and records them in the usage history information 425. This process is performed in real time or in batch format, and the recorded data is used for subsequent billing and agent evaluation processes.

[0203] Next, in step S1720, the billing module 414 of the management server 400 calculates the billing amount according to the contract plan based on the recorded usage history. Multiple billing models are assumed for the contract plan, such as pay-as-you-go, monthly flat rate, and initial free + additional charges, and the conditions of the fee structure, such as price, unit, and upper limit, are set for each AI agent. The billing calculation process is automated via the billing module 414 and also supports the management of usage limits for each user.

[0204] Next, in step S1730, the management server 400 performs payment processing based on the billing amount through an external service. Various payment processing services are envisioned as payment methods, and billing is performed based on the user's registered payment information. The success or failure of the payment is recorded along with the transaction ID, and a notification is sent to the user.

[0205] Finally, in step S1740, the management server 400 distributes the earned revenue to the agent's owner or participants. The distribution ratio is apportioned to the owner, contributing users, expert personas, etc., based on the rules established via the administrator terminal 600 during AI agent registration. The distribution processing results are recorded in the ledger information via the management server 400 and can be used later for review and tax processing by the administrator.

[0206] (6-11) Deviation detection and activity log recording process Figure 18 is a flowchart showing an example of the deviation detection and activity log recording flow 1800 according to this embodiment. This flow 1800 is a process in which the management server 400 sequentially monitors semantic deviations for each response or action output by the AI ​​agent and saves the detection results as a log. The purpose of this process is to secure an activity history that serves as the basis for evaluating the stability and reliability of the personality model and to utilize it in subsequent learning, correction, and explanation processes. The details of each step are described below.

[0207] First, in step S1810, the personality model management module 214 of the generation server 200 initiates monitoring processing targeting the output whenever the AI ​​agent (personality model) generates a response or action to the user. This processing may be automatically initiated based on output triggers and may be designed to apply regardless of the output form, such as interactive responses, decision-making actions, or notification messages.

[0208] Next, in step S1820, the deviance detection module 218 of the generation server 200 vectorizes the output response and action content and calculates the semantic distance by comparing it with the reference vector V_baseline held by the ghost. A natural language vector embedding model is used for the vectorization process, and an evaluation function such as cosine similarity or Euclidean distance is applied to calculate the distance. This evaluation is performed using the same mechanism as the deviance interpretation process in Figure 14.

[0209] Next, in step S1830, if the distance evaluation results in a deviation exceeding a predetermined threshold, the generation server 200 executes a warning notification and control processing for the output. The control processing includes interrupting the response, re-presenting it, and switching the dialogue style, which contributes to preventing malfunctions and inappropriate expressions of the AI ​​agent. The notification is also sent to the user terminal 700 or the administrator terminal 600 to visualize the deviation detection.

[0210] Finally, in step S1840, the personality model management module 214 of the generation server 200 records and saves all responses generated by the AI ​​agent and their deviation detection results as an activity log. The log includes metadata such as output text, vector distance values, thresholds, presence or absence of deviations, and type of action, which are sent to the management server 400 for centralized management. The accumulated activity logs are widely used as supplementary information for subsequent personalized learning, deviation tendency analysis, and user explanations.

[0211] (7) Screen example Next, we will describe an example screen for System 100 with reference to the drawings.

[0212] (7-1) User Interface Screen Figure 19 shows an example of the user interface screen of the AI ​​interview tool according to this embodiment. This screen is displayed during an interview session with the original person to collect dialogue data necessary for generating a personality model, and is configured to provide integrated visualization of the progress, navigation support, and answer input operations.

[0213] On the left side of the screen, a section is displayed showing the overall structure and progress of the interview. In this example, the interview is structured into multiple points, such as "Icebreaker" and "Part 1: Career Foundation and Professional Philosophy," making it easier for the user to understand the purpose of each phase. The current question number and estimated time are also displayed, allowing the user to check the overall progress at any time.

[0214] In the center, the AI ​​interviewer's responses are displayed in a chat format, and the conversational guidelines are presented at the beginning of the interview, such as "There are no right or wrong answers" and "Please give us your honest opinion." This is designed to reduce the psychological burden on the original interviewee and allow them to focus on disclosing their thought process.

[0215] At the bottom of the screen are input fields and a submit button for the user to enter their answers. Users can also switch to voice input by using the "High-Quality Voice Input" button or the "Voice ON / OFF" toggle, as needed. Additionally, hints such as "Include specific anecdotes or experiences" are displayed at the bottom of the screen, designed to encourage more in-depth responses.

[0216] This screen is an interface displayed on the user terminal 700, and is configured to synchronously advance the interview session while communicating with the generation server 200. The generation server 200 is responsible for controlling the presentation of the questions displayed on the screen, collecting answers from the original interviewee, and generating answer vectors. The collected data is also sent to the management server 400 as needed and used for centralized management of progress and answer history.

[0217] User responses recorded through the screen are stored in the log structure data as a dialogue log. The log data consists of spoken text, input method (text / voice), timestamp, corresponding question ID, and response metadata. Furthermore, if voice input is used, both the hash of the audio file and the denoised text conversion result are linked. This prevents misinterpretation and information loss due to audio.

[0218] Furthermore, this screen incorporates an answer assistance function to support the user's thought process. For example, depending on the question, prompts such as "Please include specific personal experiences" or "Please reflect on your decision-making criteria" are automatically displayed at the bottom of the screen. These assistance sentences are obtained from the generation server 200 based on prompt configuration information 224 defined for each question. The system is designed to dynamically switch and highlight assistance sentences based on the user's input tendencies and past answers.

[0219] (7-2) Team chat screen Figure 20 shows an example of a team chat screen according to this embodiment. This screen is a UI displayed on the user terminal 700 and is an interaction interface that enables real-time dialogue with a personality model (ghost) via the generation server 200. The user can call a specific ghost by mentioning it within the team space and input questions, requests, consultations, etc.

[0220] In this example, the user asks "Tanaka Taro Ghost," an AI agent modeled after the highly skilled salesperson "Tanaka Taro," about "the secrets to improving sales performance," and the display shows Ghost's response. Ghost's utterances are the output of the personality model that responds as an AI agent, and the explanations are presented in text format based on past experience, values, and the set persona. In this example, the response provides specific methods regarding in-depth analysis of customer issues and interviewing techniques from the perspective of someone with 20 years of sales experience.

[0221] At the bottom of the screen, there is a free-form text field and a send button, allowing users to enter any message and continue communicating asynchronously or synchronously with ghosts or other team members. Furthermore, hints below the message field encourage the use of mentions, making the system adaptable to complex conversations with multiple ghosts or teams.

[0222] Furthermore, the content of the conversation obtained through the chat screen is recorded in log format as needed and can be reflected in activity log information or personality model growth processing (Figure 14). In addition, during the response generation process, the user's question is vectorized and then compared with the knowledge graph and stored memory to achieve individually optimized output.

[0223] Thus, in this embodiment, the generation server 200 has the function of simultaneously registering and managing multiple personality models (ghosts), and each ghost is associated with a unique agent ID along with attribute information such as area of ​​expertise, personality traits, and usage restrictions. The user terminal 700 can call up multiple ghosts with which it has a usage agreement at any time by mentioning them in the format "@(agent name)". This enables smooth switching of knowledge and perspectives across multiple fields.

[0224] Furthermore, a chat permission control mechanism is implemented to control speech in the team chat function. Each ghost is configured to grant or restrict access rights based on the user's group affiliation, authentication token, and purpose of use, and it is also possible to manage private ghosts and internal-only agents. For example, a manager-level ghost may be accessible to a section manager, but only users with a manager-level position or higher may be able to access a director-level ghost. This permission information is stored in user information 421 and is automatically verified at the start of a session. This ensures secure knowledge sharing while minimizing the risk of leakage of internal skills and confidential information, as well as unauthorized access.

[0225] Furthermore, the generation server 200 performs response optimization processing in conjunction with the user's profile information in response to input from the user terminal 700. Specifically, it refers to the user's area of ​​expertise, conversation style preferences, past usage history, etc., stored in the user information 421, and incorporates them as dynamic instruction sentences (input prompts) when generating prompts. As a result, even with the same ghost, adaptive response expressions are realized for different users, providing a personalized conversational experience.

[0226] In the business world, such AI agents can be used to create a "ghost" of a highly skilled project manager (the original person) and have it function as an autonomous project manager substitute. For example, the ghost can be given the task of "monitoring the progress of assigned projects daily, checking the project management tool ticket if a delay risk is detected, sending a reminder to the person in charge via a chat tool to check the status, and escalating to me if there is no response within 24 hours." In this case, the AI ​​agent autonomously interacts with external APIs such as project management tools and chat tools to perform the task. Furthermore, the thinking process construction module 216 functions to clearly explain the reason why a reminder was deemed necessary (e.g., delay data from the project management tool ticket), so that the person in charge can understand and respond accordingly. This enables smooth project management that goes beyond mere automated notifications.

[0227] Another more advanced business application is the simulation of new business development. In this application, the role model generation function and virtual team formation function are first executed, and a virtual team is formed by gathering ghosts of excellent marketers, engineers, sales representatives, etc., from both inside and outside the company. When the formed virtual team is made to execute the business plan, each ghost rapidly simulates the results and potential challenges after market launch, in cooperation with each other, based on the expertise and thinking styles stored in their own base memory. This enables advanced decision support through the combination of diverse expertise, which would be difficult for humans alone due to time and cost constraints. These ghosts can operate 24 hours a day, 365 days a year, realizing the full automation of intellectual work.

[0228] Furthermore, in the area of ​​talent development, a "ghost" of a busy leader (the original leader) can be generated and provided to subordinates and junior colleagues as an in-house mentoring platform. Subordinates can consult with the leader's ghost 24 hours a day, 365 days a year, regarding work-related matters. In this case, the ghost's base memory stores the leader's experience, abstracted through anonymization so that personal names and specific project names are not violated, allowing for the secure sharing of only essential insights without infringing on individual confidentiality agreements. This enables the knowledge of one leader to be scaled across the entire organization, contributing to the improvement of the organization's overall capabilities.

[0229] Furthermore, in the lifestyle domain, it can be applied to a life matching service that analyzes essential compatibility before actually meeting by generating a ghost representing the user as the original person and conducting communication simulations with the ghosts of potential partners on the platform. In this service, the personality model generation module 213 extracts the user's deep-seated values ​​through dialogue with an AI interviewer and questionnaires, and generates a highly detailed personality model. In the simulation between the ghosts, the deviation detection module 218 functions to suppress responses that deviate from the original person's values ​​(for example, a normally mild-mannered person becoming extremely aggressive), thereby ensuring a highly reliable compatibility analysis. This application can be applied, for example, to the marriage market.

[0230] Furthermore, in the social and public sphere, by generating "ghosts" from the teaching know-how of top-class instructors nationwide and utilizing them as individually optimized educational platforms, it is possible to contribute to correcting educational disparities. In this case, students can ask questions to the ghosts as individualized instruction at any time. At this time, the personality model generation module 213 stores expert reflections in its base memory, which analyze the instructor's educational philosophy and teaching methods from the perspectives of multiple expert personas such as "psychologists" and "educational technologists." As a result, the ghosts not only provide knowledge-based answers, but also infer the background of "why this student is struggling here" and provide multifaceted instruction tailored to the student's level of understanding. In addition, beyond mere knowledge transfer, it is possible to provide instruction rooted in the student's (user's) inner self, such as motivational coaching and mental care.

[0231] (8) Summary As described above, according to the system 100 of the present invention, an AI agent that mimics an individual is generated based on multifaceted data concerning the individual's characteristics, and the generated AI agent is provided as a tradable digital asset. Therefore, it becomes possible to comprehensively manage everything from the generation to the distribution of AI agents, and it is possible to promote the formalization and utilization of intellectual assets.

[0232] Furthermore, according to system 100, when an AI agent is registered, it includes an anonymization module 412 that detects and anonymizes the personally identifiable information associated with the agent. This makes it possible to protect the privacy of experts while enabling their use as digital assets.

[0233] Furthermore, according to system 100, the billing module 414 of the management server 400 has the function of processing charges based on user usage and distributing the revenue to the owners of the AI ​​agents. This makes it possible to provide continuous economic incentives to the creators of AI agents. This makes it possible to promote the improvement of AI agent quality and the formation of a sustainable ecosystem.

[0234] Furthermore, according to System 100, since it automatically selects an appropriate billing model according to the attributes of the AI ​​agent, service providers can implement flexible pricing strategies, and users can achieve reasonable cost burdens according to their usage and frequency.

[0235] Furthermore, according to System 100, users can effectively search for AI agents by specifying their desired conditions (expertise of the modeled expert, output format, application, etc.) in natural language. This improves the accuracy of matching user needs with agent capabilities and efficiently provides the most suitable agent for achieving the objective.

[0236] Furthermore, according to System 100, users can simply input abstract goals in natural language, and the AI ​​agent will translate these into concrete execution plans, enabling external integration to execute tasks in cooperation with external systems. This allows for highly autonomous execution processing that understands the user's intentions, and the use of the AI ​​agent can support the automation of daily tasks and creative activities.

[0237] Furthermore, according to system 100, it is equipped with a personality model generation module 213 that generates personality models based on expert reflection, allowing the AI ​​to incorporate an individual's deep-seated values ​​and thought patterns. Therefore, it is possible to construct a highly accurate AI agent that goes beyond superficial linguistic imitation and reflects even internal characteristics.

[0238] Furthermore, System 100 includes a deviation detection module 218 that vectorizes the output of the AI ​​agent in real time and detects the degree of deviation by comparing it with a reference vector. This allows for the detection and suppression of responses that deviate from individual characteristics, thereby improving the stability and reliability of the agent's output.

[0239] Furthermore, according to system 100, it includes a thought process construction module 216 that extracts contextual information from the base memory in response to an input question and generates a response that visualizes the thought process. As a result, users can more easily understand the reasons and rationale behind the AI ​​agent's responses and enjoy highly explainable intellectual support.

[0240] Furthermore, according to system 100, it includes a personality model management module 214 that updates the personality model by utilizing output history and feedback from the user. As a result, the AI ​​agent can adapt to the passage of time and environmental changes, and achieve continuous evolution that reflects the user's latest intentions and judgment tendencies.

[0241] Furthermore, the personality model management module 214 checks the consistency between past thought processes and values, and performs a self-consistency verification process to detect and correct inconsistencies in thinking. This reduces internal contradictions in the personality model and maintains reliability as a more consistent personality intelligence.

[0242] Furthermore, the personality model management module 214 summarizes the output of the AI ​​agent and notifies the person who served as the model for the AI ​​agent, and then processes the received feedback to reflect it in the model. In this way, a cycle of co-evolution between the AI ​​agent and the person can be realized while deepening mutual understanding between them.

[0243] Furthermore, the personality model management module 214 analyzes and integrates multiple personality models to generate role model ghosts with thought and behavior patterns optimized for specific roles. This makes it possible to create ideal AI agents that can be used for talent development, team design, and other applications.

[0244] Furthermore, the agent provisioning module 413 executes the process of forming a virtual team that includes multiple AI agents and humans selected by the user. This enables advanced intellectual support in which members with different areas of expertise collaborate to address multifaceted challenges.

[0245] (9) Second Embodiment Next, the second embodiment of the present invention will be described. FIG. 21 is a diagram showing an example of the configuration of an AI agent providing system 100B (hereinafter, system 100B) according to the second embodiment. In the above-described system 100, an AI agent imitating mainly highly skilled experts was provided to others via a marketplace, whereas in system 100B, it is assumed that a ghost that is a clone of the user himself / herself is generated and provided to others.

[0246] Specifically, as shown in FIG. 21, a user (job seeker A, job seeker B, etc.) generates an AI agent (ghost A, ghost B, etc.) modeled on himself / herself via an AI agent generation server 200. The generated ghosts are provided to an employment staff C of a recruiting company and a section chief D of the department where the user is to be assigned via a marketplace management server 400. In this case, the marketplace management server 400 may be referred to as an AI agent providing server.

[0247] In the illustrated example, an employment staff C and a section chief D belonging to a recruiting company can conduct various types of questions and task presentations in advance for ghosts A and ghosts B modeled on job seekers. As a result, it becomes possible to objectively and efficiently evaluate the thinking tendencies, values, problem-solving approaches, etc. of job seekers, which were difficult to grasp in conventional document screening and interviews, through conversations with the AI agents.

[0248] For example, in order to grasp the adaptability to corporate culture and communication style, the employment staff C inputs questions such as "How do you handle it when a conflict occurs at work?" and "What do you do when you cannot accept the instructions of your supervisor?" to ghost B and obtains the response as the answer. Also, in order to judge the practical performance ability after assignment, the section chief D inputs a task such as "Propose the initial response tasks when a product trouble occurs in the quality assurance department" to ghost A and obtains the response as the answer.

[0249] These dialogues and task-response processes are logged as output from the AI ​​agent, and information such as thought processes, logical structure, and priority of value judgments is automatically analyzed. The analysis results are scored based on multiple perspectives such as "task comprehension," "logical consistency," and "signs of cooperativeness," and output as a comprehensive evaluation report of the job seeker ghost. This kind of pre-evaluation makes it possible to grasp the suitability of a candidate from multiple perspectives before proceeding to the actual interview. In other words, it becomes possible to obtain more multifaceted information that could not be obtained from the information contained in application documents submitted by job seekers as part of the document screening process, and to use it in the selection process.

[0250] Pre-interviews using this type of "ghost" candidate do not require the candidate's physical presence, eliminating the need for traditional, cumbersome scheduling and securing interview dates. Especially in the initial phases when comparing a significant number of candidates, this enables a flexible selection process that is not dependent on time or location, significantly improving the efficiency of recruiters' work.

[0251] Furthermore, because the ghost profile is generated based on the candidate's personality model, it is less likely to be affected by nervousness that can prevent candidates from demonstrating their true abilities, as is often seen in traditional interviews, and less likely to involve superficial statements made for the sake of the interview. Therefore, it is possible to more accurately extract the candidate's true feelings that form the core of their values ​​and way of thinking. This makes it possible to grasp deeper information that is difficult to obtain from resumes or face-to-face interviews, such as the candidate's inner motivations and attitude towards problem-solving.

[0252] Furthermore, because the "ghost" profile is constructed based on statistics of the individual's past experiences, skills, and behavioral history, it is difficult to rely on superficial responses such as temporary memorization or model answers. This characteristic helps to suppress exaggeration of skills that deviate from actual abilities, and false or embellished self-promotion, allowing for a fairer and more realistic evaluation. In addition, by conducting virtual interviews and simulated work performance before actual interviews, it is possible to evaluate aspects that are difficult to judge with traditional resumes and skill sheets, such as cultural fit and compatibility of values.

[0253] Furthermore, a virtual interview environment with an AI agent allows for challenging questions and adverse situations that would be difficult to implement in a face-to-face setting, while maintaining considerable consideration. As a result, it becomes possible to observe candidates' stress tolerance, emotional control, and ability to cope with unexpected situations in a gentler format than in a real interview. These characteristics are also beneficial from the perspective of mental health and adaptability, which can become issues when actually performing work.

[0254] In addition, all speech logs and responses during interviews are recorded and analyzed, which can be used for comparing multiple candidates and evaluating compatibility. This is also useful from the perspective of accountability and record keeping regarding selection results, and has the effect of increasing the transparency and reproducibility of the entire recruitment process. As described above, pre-interviews using "ghost job seekers" can provide a new selection approach that balances efficiency, fairness, and deep understanding.

[0255] Furthermore, by using ghosts (AI agents that mimic users on the recruiting company side) of recruiters and department heads, this system can also conduct virtual interviews and work simulations using only ghosts. Figure 22 shows an example of how system 100B can be used.

[0256] As shown in Figure 22, an example of a virtual interview using ghosts is described, specifically a virtual group interview conducted by an AI agent. In this example, ghosts of recruiter C and department head D from the recruiting company (Ghost C, Ghost D) and ghosts of job seekers A and B (Ghost A, Ghost B) autonomously interact based on a pre-set interview agenda to conduct a virtual group interview.

[0257] The virtual group interview may be conducted with only multiple job seeker ghosts, and the dialogue logs may be evaluated by recruiter C and department head D, both affiliated with the hiring company, or each of the recruiter C and department head D's ghosts may evaluate them separately. Alternatively, a virtual recruitment meeting may be held where the recruiter C and department head D's ghosts discuss the candidates based on the dialogue logs from the virtual group interview conducted with only multiple job seeker ghosts.

[0258] Based on the dialogue logs obtained from such virtual group interviews, factors such as teamwork, alignment of values, consistency of thinking, skills, and compatibility scores are quantified. Based on these scores, recruiters can select actual interviewees and make hiring decisions, thereby streamlining the selection process and reducing mismatches. It also improves the transparency and consistency of the selection process. For job seekers, knowing how their virtual interviewee was evaluated beforehand allows them to receive feedback and opportunities for growth. Thus, this invention achieves a high level of mutual understanding and efficiency in recruitment activities.

[0259] (9-1) System Configuration Next, the system configuration of system 100B according to the second embodiment will be described. In the following description, only the parts that differ from the first embodiment will be described, the same components will be denoted by the same reference numerals, and repeated explanations will be omitted. Figure 23 is a diagram showing an example of the configuration of the marketplace management server 400B (hereinafter simply referred to as management server 400B) according to the second embodiment.

[0260] In addition to the modules 411-415 shown in Figure 4, the main memory 401 of the management server 400B is also loaded with the following functional modules: dialogue control module 416, dialogue evaluation module 417, report generation module 418, simulation module 419, and activity recording module 420. Each of these modules is invoked by the processor 203 during execution.

[0261] The dialogue control module 416 has the function of starting, progressing, and ending autonomous dialogue sessions based on multiple ghosts and topics specified by the user. It manages the turn (speaking turn) in the discussion and, if the discussion stalls, presents new questions or supplementary information to promote the smooth progress of the dialogue. It also maintains balance among participating ghosts by controlling the order of utterances and time allocation.

[0262] The dialogue evaluation module 417 analyzes dialogue logs between multiple ghosts using natural language processing technology and calculates evaluation indicators such as the emotional polarity of utterances, the proportion of cooperative or confrontational utterances, contribution, knowledge validity, and topic diversity. Based on these indicators, it generates an overall dialogue score and compatibility score and records them as evaluation information.

[0263] The report generation module 418 generates evaluation reports and activity reports based on the analysis results of the dialogue evaluation module 417 and the historical information obtained by the activity record module 420. The reports include evaluation scores, analysis comments, and improvement suggestions, and are output in a format that users can intuitively understand.

[0264] The simulation module 419 conducts a virtual discussion or meeting based on a specified combination of ghosts and agenda, taking into account the roles of each ghost. During execution, it works in cooperation with the dialogue control module 416 to manage the order of utterances and progress, and after completion, it hands over the dialogue log to the dialogue evaluation module 417. In other words, the simulation module 419 performs the function of managing the smooth progress of discussions among multiple AI agents.

[0265] The activity recording module 420 acquires the entire activity history of interactions and simulations between ghosts and records it in the activity history information 428 and interaction log information 427 of the auxiliary storage device 402. The recorded content is used for subsequent evaluation and analysis, and for future performance improvements.

[0266] In addition to the respective pieces of information 421 to 425 shown in FIG. 4, the auxiliary storage device 402 further stores role setting information 426, dialogue log information 427, activity history information 428, report information 429, and evaluation information 430.

[0267] The role setting information 426 is information regarding the roles assigned to each ghost participating in a simulation or dialogue session. For example, it includes role types such as "job seeker", "department head", "recruitment staff", etc., and control parameters such as the speaking policy and speaking frequency associated with that role. This enables ensuring diverse perspectives and speech balance in the dialogue between ghosts.

[0268] The role setting information 426 is constituted by a role setting table shown in FIG. 24. The role setting table records, for a uniquely set role ID, a session ID, a ghost ID, an assigned role, a setting date and time, and a setting user ID. Here, the session ID is an identifier set for each simulation or dialogue session. The ghost ID is an identifier for identifying the ghost to which the role is assigned, and the assigned role indicates the role type that the ghost undertakes. The setting date and time is the date and time when the role was assigned, and the setting user ID indicates the user who performed the setting.

[0269] The dialogue log information 427 is information that records the content of the dialogue between ghosts executed by the dialogue control module 416 in chronological order. It includes a speaker ID, the content of the speech, a timestamp, a dialogue session ID, etc., and is used for subsequent compatibility evaluation and report generation.

[0270] The dialogue log information 427 is constituted by a dialogue log table shown in FIG. 24. The dialogue log table records, based on a dialogue ID, the order of speech, the speaker ID, the content of the speech, the timestamp, and the session ID. The speaker ID is information for identifying the ghost or user who made the speech, and the content of the speech is stored as natural language text or the like. The timestamp indicates the time when the speech was made, and the session ID identifies the dialogue session to which the speech belongs.

[0271] Activity history information 428 is information that records the usage history and simulation participation history of each ghost, and is composed of the activity history table shown in Figure 24. The activity history table records the activity history of each ghost identified by the history ID, and includes the ghost ID, related session ID, update event, update date and time, and update content summary. The update event indicates the type of event, such as a change in ghost settings or an update of the personality model, and the update content summary briefly shows an overview of the changes.

[0272] Report information 429 is information that stores evaluation reports and activity reports created by the report generation module 418, and consists of the report information table shown in Figure 24. The report information table records the session ID, generation date and time, summary content, and evaluation score for reports identified by the report ID. The evaluation score is a quantitative evaluation of the session or ghost response content that is the subject of the report.

[0273] The evaluation information 430 stores numerical data of the evaluation score, compatibility score, and the evaluation indicators that form the basis of the calculation, which are calculated by the dialogue evaluation module 417 between ghosts. This information will be used for future matching and recommendation processes.

[0274] The evaluation information 430 consists of the evaluation information table shown in Figure 24. The evaluation information table records evaluations of ghosts identified by evaluation IDs, or evaluations of compatibility between ghosts, and includes ghost IDs (including pair values), evaluation scores (including compatibility scores), each evaluation metric, evaluation date and time, and related session IDs. Each evaluation metric includes evaluation values ​​for each element such as cooperativeness, degree of agreement on issues, emotional stability, and validity of knowledge, and the evaluation score and compatibility score are calculated as overall indicators of these. In other words, the evaluation information table records evaluation scores for individual ghosts and compatibility scores for multiple ghosts.

[0275] (9-2) Regarding virtual interview processing Figure 25 is a flowchart showing an example of the virtual interview flow 2500 according to this embodiment. This flow 2500 is a process in which the management server 400B conducts virtual interviews between a recruiter ghost and a job seeker ghost for candidate matching in recruitment activities, analyzes the results of the dialogue, and supports candidate selection. This process is mainly executed by the dialogue control module 416, dialogue evaluation module 417, report generation module 418, simulation module 419, and activity recording module 420 of the management server 400B. The details of each step will be described below.

[0276] First, in step S2510, the generation server 200 generates a recruiter ghost modeled after the recruiter based on their attributes and recruitment policies. During generation, personality types and speaking styles reflecting the company culture and desired candidate profile are set, and interview question tendencies and evaluation criteria are incorporated. At this time, a URL for taking an AI interview is sent to the recruiter's user terminal 700. The acquisition module 210 of the generation server 200 acquires multifaceted information about job seekers through various questionnaires, including AI interviews with recruiters. The same process is followed for generating ghosts of department heads.

[0277] Next, in step S2520, the generation server 200 generates a job seeker ghost modeled after the job seeker based on their career information, skill set, and values ​​data. This ghost is given business knowledge and self-promotion strategies related to the applied position, and its abilities and suitability are evaluated through dialogue with a recruiter ghost. This process is initiated by a job application from the user who is a job seeker. At this time, a URL for taking an AI interview is sent to the job seeker's user terminal 700. The acquisition module 210 of the generation server 200 acquires multifaceted information about the job seeker through various questionnaires, including an AI interview with the job seeker.

[0278] Next, in step S2530, the management server 400B conducts a virtual interview between the recruiter ghost and the job seeker ghost. In this process, the HR ghost acts as the moderator, managing the order of speaking and the progress of the agenda, and generating additional and follow-up questions as needed, thereby creating an interview session that can evaluate not only skills but also cultural fit and interpersonal aptitude.

[0279] Next, in step S2540, the management server 400B analyzes the dialogue logs obtained from the virtual interview. Specifically, the dialogue evaluation module 417 uses natural language processing technology to analyze the emotional polarity, logic, consistency, reaction speed, and cooperativeness of the utterances, and calculates the degree of match between the skills and the job requirements, the affinity of values, and the suitability of the communication style.

[0280] Next, in step S2550, the management server 400B generates a matching score and a detailed analysis report based on the analysis results. The report generation module 418 creates a report that includes scores for each evaluation item, example utterances, strengths, and concerns, and outputs it in a format that recruiters can use to evaluate candidates.

[0281] Finally, in step S2560, recruiters and department heads evaluate candidates based on the matching scores and analysis reports. If necessary, they compare and rank multiple candidates to identify those who should proceed to the next interview stage. These results are stored in the auxiliary storage device 402 via the activity log module 420 and used to improve future recruitment strategies and models.

[0282] Furthermore, System 100B is not limited to recruitment activities; it can be used for a variety of other purposes. For example, in supporting the formation of project teams, it can be used to have ghosts of potential employee members interact with each other, and by analyzing the logs of these interactions, it is possible to evaluate their collaborative abilities and synergistic thinking, and design the optimal team structure.

[0283] (9-3) Team Formation Simulation Process Figure 26 is a flowchart showing an example of the team formation simulation flow 2600 according to this embodiment. This flow 2600 is a process in which the management server 400B conducts a dialogue simulation between candidate ghosts who possess the skills and aptitudes required for the project, and determines the optimal team composition based on the results. This process is mainly executed by the simulation module 419, dialogue control module 416, dialogue evaluation module 417, report generation module 418, and activity recording module 420 of the management server 400B. The details of each step will be described below.

[0284] First, in step S2610, the management server 400B selects multiple candidate ghosts with the required skills and experience from the marketplace based on the input from the project manager (PM). At this time, it refers to the ghost's skill tags, past evaluation scores, activity history, etc., to confirm their suitability as candidates.

[0285] Next, in step S2620, the management server 400B starts a dialogue simulation between the selected ghosts. The dialogue control module 416 manages the agenda setting and the order of utterances, while the simulation module 419 controls the progress of the dialogue and the introduction of additional agenda items. The agenda can be specific content aligned with the project theme (for example, "Developing a sales strategy for a new product").

[0286] Next, in step S2630, the management server 400B displays the conversations between the ghosts in real time on the user terminal 700 through a user interface such as the team chat screen. This allows the PM and other stakeholders to keep track of the progress of the discussion and the interactions between members, and to modify the agenda or insert questions as needed.

[0287] Next, in step S2640, the management server 400B analyzes the dialogue log and generates an evaluation report. The dialogue evaluation module 417 calculates evaluation indicators such as cooperativeness score, contribution of speech, diversity of issues, and degree of conflict of opinions, and the report generation module 418 creates an analysis report that combines these indicators.

[0288] Finally, in step S2650, the management server 400B assembles a team with the most productive and suitable member configuration for the project objectives, based on the analysis report and evaluation score. This result is saved to the auxiliary storage device 402 via the activity log module 420 and used for future project organization and ghost model improvement.

[0289] Furthermore, virtual brainstorming using ghosts can be used to combine the insights of members with diverse backgrounds and generate innovative ideas and strategies. In this case, an idea generation team can be formed by combining ghosts of the company's own development staff with ghosts of external experts.

[0290] Furthermore, in organizational development and consulting applications, it can be used to check compatibility with staff in departments to which employees will be transferred during personnel changes. In other words, by having multiple "ghosts" modeled after real employees and stakeholders interact with each other, it is possible to visualize in advance potential communication challenges arising from role-based tensions and differences in values ​​within the organization. For example, by using ghosts of people in different positions, such as senior management, middle management, and junior employees, and simulating their views on specific business issues or policies, the causes of interdepartmental conflicts and misunderstandings can be highlighted, and it functions as a support tool to facilitate the consideration of constructive improvement measures and adjustment strategies.

[0291] (9-4) Decision support processing Figure 27 is a flowchart showing an example of the decision support flow 2700 according to this embodiment. This flow 2700 is a process in which the management server 400B conducts discussion simulations between ghosts modeled after each business unit manager in order to support company-wide policy decisions by management, and formulates a meeting strategy based on the results. This process is mainly executed by the simulation module 419, dialogue control module 416, dialogue evaluation module 417, and report generation module 418 of the management server 400B. The details of each step will be explained below.

[0292] First, in step S2710, the management server 400B selects a ghost modeled after each business unit manager based on input from the business planning staff. This makes it possible to reproduce the speaking tendencies and values ​​of key people involved in actual decision-making in a virtual space.

[0293] Next, in step S2720, the management server 400B sets company-wide issues. For example, themes requiring cross-departmental discussions are set, such as "Company-wide DX Promotion Policy" or "New Market Entry Strategy." These issues are registered as discussion scenarios by the simulation module 419.

[0294] Next, in step S2730, the management server 400B executes a discussion simulation between ghosts based on the configured tasks. The dialogue control module 416 manages the order of utterances and the progress of the agenda, and the simulation module 419 inserts additional agenda items and conflict structures as needed to reproduce a more realistic discussion.

[0295] Next, in step S2740, the management server 400B predicts interdepartmental conflicts of interest and divergences in arguments based on the results of the discussion simulation. The dialogue evaluation module 417 analyzes the content of utterances and response tendencies using natural language processing and quantifies the degree of cooperation, ease of consensus building, and degree of disagreement.

[0296] Finally, in step S2750, the management server 400B formulates the meeting agenda and facilitation strategy based on the prediction results. The report generation module 418 creates a meeting management report including these and presents it to management, thereby facilitating constructive and efficient decision-making in the actual meeting.

[0297] Furthermore, ghosts may be assigned to maintain and operate the AI ​​agent marketplace. In this case, as a quality control measure for the AI ​​agents (ghosts) registered in the marketplace, newly generated and registered ghosts can be made to interact with existing high-quality ghosts that serve as a benchmark (for example, pre-generated and registered role model ghosts). This allows for verification of whether the responses are logical, consistent, and cooperative. The dialogue logs are analyzed based on evaluation metrics using natural language processing to quantitatively identify issues such as the consistency of responses, knowledge biases, overreactions, and contradictions. This ensures that the quality of ghosts is maintained above a certain level, enabling them to be operated as safe and reliable AI agents.

[0298] (9-5) Regarding the quality control process for ghosts Figure 28 is a flowchart showing an example of the ghost quality control flow 2800 according to this embodiment. This flow 2800 is a process in which the management server 400B verifies the response quality of newly generated and registered ghosts and starts operation on the marketplace only if it meets predetermined criteria. This process is mainly executed by the simulation module 419, dialogue control module 416, dialogue evaluation module 417, and report generation module 418 of the management server 400B. The details of each step are described below.

[0299] First, in step S2810, the management server 400B registers the new ghost generated by the generation server 200. This ghost is based on a personality model built by the user using the generation server 200 and is stored on the marketplace as registration information 422.

[0300] Next, in step S2820, the management server 400B initiates a dialogue between the new ghost and a standard ghost (role model ghost) that has been previously certified as meeting high-quality standards. The dialogue control module 416 controls the agenda setting and utterance order, and obtains responses from both ghosts.

[0301] Next, in step S2830, the management server 400B analyzes the generated dialogue log. The dialogue evaluation module 417 uses natural language processing technology to analyze the logic, consistency, cooperativeness, and response speed of the utterances and extracts evaluation indicators.

[0302] Next, in step S2840, the management server 400B performs a quality assessment based on evaluation indicators. The evaluation includes logical consistency, consistency of thought processes, accuracy of knowledge, presence or absence of overreactions or contradictions, and collaborative dialogue attitude.

[0303] Next, in step S2850, the management server 400B quantitatively identifies problems from the analysis results. For example, it detects a lack of knowledge or bias regarding a specific topic, or inappropriate speech tendencies.

[0304] Finally, in step S2860, the management server 400B publishes the ghost for use on the marketplace only if the evaluation results meet the quality standards. If the standards are not met, it prompts for relearning or parameter adjustments for improvement. This ensures that the system always provides users with only highly reliable and secure ghosts.

[0305] Furthermore, System 100B can also be used to provide new digital content. Specifically, it can generate and deliver content such as virtual debates or commentaries on real-world events, using ghosts modeled after historical figures and celebrities. Such functionality is expected to have applications in fields such as education, research, and entertainment.

[0306] (9-6) Regarding the processing of distribution content Figure 29 is a flowchart showing an example of the content delivery flow 2900 according to this embodiment. This flow 2900 is a process in which the management server 400B registers ghosts generated using historical figures and famous people as models, and generates and delivers educational, research, and entertainment content through virtual discussions, commentaries, etc. This process is mainly executed by the simulation module 419, dialogue control module 416, and report generation module 418 of the management server 400B. The details of each step will be explained below.

[0307] First, in step S2910, the management server 400B registers ghosts generated by the administrator based on historical figures and celebrities. During generation, a personality model is constructed based on information such as the person's thoughts, statements, writings, and historical background. For example, if the person is a scholar, academic works such as research papers, books, and lecture records written by that person are referenced to reflect their theoretical background and research stance. If the person is a politician, for example, their political speeches, parliamentary responses, statements in diplomatic negotiations, and policy-making processes are analyzed to model their political ideals and diplomatic stance. If the person is a famous military commander, for example, their battle record, military strategies, interactions with vassals and allies, anecdotes and correspondence, and significant actions that influenced the society and political situation of the time are extracted and reflected in the personality model. The model constructed in this way is said to be able to faithfully reproduce the person's way of thinking and decision-making patterns.

[0308] Next, in step S2920, the management server 400B causes the ghost to conduct a virtual discussion or comment on real-world events. The topics can be broadly set to include educational themes, social issues, historical events, etc., and the dialogue control module 416 controls the order of utterances and the progression of the discussion.

[0309] Next, in step S2930, the management server 400B collects the dialogue logs generated during the discussion and commentary process. The collected logs are used for subsequent content editing, quality control, and analysis. For example, by having the generated ghosts of historical figures engage in discussions and commentaries, it is possible to obtain statements from multiple perspectives based on the ideological background and values ​​of those individuals. For instance, it becomes possible to analyze the circumstances of historical events, conduct virtual exchanges of opinions between people from different eras and fields, or present historical lessons and unique solutions to contemporary social issues. This creates intellectual dialogues that transcend time and fields that cannot be realized in the real world, providing high added value in educational, research, and entertainment applications.

[0310] Finally, in step S2940, the management server 400B generates and distributes educational, research, and entertainment content based on the collected dialogue logs and analysis results. Distribution can also be done in conjunction with external services such as online platforms and learning management systems. This allows users to experience discussions with combinations that would be impossible in reality, as well as commentary from diverse perspectives.

[0311] (9-7) Screen example Figure 30 shows an example of a third screen, illustrating a screen for selecting multiple ghosts registered in the marketplace. This screen is designed to allow users to efficiently search and select ghosts that match their desired skill set and budget requirements. A filter panel is located on the left side of the screen, providing keyword input fields, skill specification fields, a budget slider, and filtering options based on ratings (star ratings). Combining these conditions can improve search accuracy.

[0312] On the left side of the screen, a list of ghosts that match the filter criteria is displayed in card format as "Ghost Search Results." Each ghost card displays the ghost name, area of ​​expertise and skills, technology stack in tag format, and "AI Match Score." The AI ​​Match Score is a numerical representation of the degree of suitability to the conditions and objectives set by the user, and is used to determine the priority of candidates.

[0313] Each ghost card has a "View Profile" button, which, when clicked, allows you to view the ghost's detailed profile (background, areas of expertise, past usage history, review ratings, etc.). This allows for selection based not only on skills but also on conversational style and past performance information.

[0314] This screen plays a crucial role in usage scenarios such as selecting job candidates and project members. For example, by selecting an appropriate "ghost" on this screen as a preliminary step before starting a virtual interview flow (see Figure 25) or a team formation simulation flow (see Figure 26), the accuracy and efficiency of subsequent processing can be improved. Furthermore, the displayed AI match score is calculated by the dialogue evaluation module 417 of system 100B and can be used as an objective and quantitative selection criterion.

[0315] Furthermore, the filtering function allows for flexible selection tailored to user needs. For example, if a project needs to be completed within a short timeframe, the budget slider can be set higher, limiting the search to ghosts with high star ratings. On the other hand, if a new project requires diverse ideas, it's possible to prioritize skill tags and keyword searches to select a wide range of candidates and combine ghosts with different match scores.

[0316] As described above, System 100B can flexibly accommodate not only proxy work for itself, but also advanced usage forms such as intelligent simulation and organizational design support.

[0317] (10) Variations As a variation of the present invention, the auxiliary storage device 402 may store the following information. • Model version management information • Dialogue context template information • Role model • Ghost information This information will be used to improve the operational quality of Ghost, ensure its reusability, and stabilize its conversational performance.

[0318] Model version control information manages the version history of the underlying model, training dataset, and parameter settings used by Ghost. This allows for performance comparisons between different model generations, rollback to previous versions, and measurement of the effectiveness of improvement measures. In particular, understanding the differences between versions is effective in maintaining reliability for Ghosts that are operated over a long period of time.

[0319] Dialogue context template information stores predefined dialogue structures, utterance patterns, and question sequences according to specific usage scenarios and business objectives. This makes it possible to perform evaluations and responses according to the same dialogue framework even with different ghosts, ensuring fairness and reproducibility in comparative verification.

[0320] Role model ghost information stores attribute information and response examples of high-quality ghosts that serve as reference standards in the generation and adjustment of other ghosts. This information is used as a comparison point during quality verification and adjustment of new ghosts, contributing to the maintenance of standard levels such as logic, consistency, and collaborative speech style.

[0321] Furthermore, the construction of a personality model in this invention is achieved by reflecting a specific individual's thinking tendencies, values, and specialized knowledge in a general-purpose large-scale language model (hereinafter referred to as the general-purpose model). Specifically, a general-purpose model is prepared in advance, and multifaceted data about the individual (interview responses, behavioral history, evaluation data, etc.) is input as training data, and additional learning (fine-tuning) is performed to obtain a personalized personality model. In this fine-tuning type, the parameters of the entire model are updated through learning, and the general-purpose model part and the personality model part are represented as a single, integrated entity.

[0322] The advantage of the fine-tuning approach lies in its ability to highly integrate the existing knowledge and reasoning capabilities of a general-purpose model with the individual's unique thinking style, ensuring a high level of consistency and naturalness in responses. Furthermore, because it operates as a single model during inference, it enables fast and efficient response generation. On the other hand, it is difficult to isolate the personality model portion, and replacing the personality requires retraining, which may increase capacity and computational resources.

[0323] On the other hand, the present invention may employ a configuration in which the general-purpose model and the personality model are physically or logically separated and managed. In this case, the general-purpose model holds parameters as the basis for contextual understanding and language generation, while the personality model holds individualized features and knowledge vectors as separate modules. For example, this could involve using low-rank adaptation (LoRA) or an adapter layer to hold additional parameters for the personality model while keeping the weights of the general-purpose model fixed, or storing embedding vectors and knowledge graphs in a memory area dedicated to the personality model.

[0324] In this segmented configuration, the parameters of the general-purpose model and the personality model are dynamically combined during inference to generate an AI agent with the desired personality. This makes it easy to replace personality models and operate multiple personalities simultaneously, enabling the immediate provision of AI agents optimized for each user. Furthermore, by independently updating the personality model, it is possible to quickly reflect individual growth and changes in circumstances without having to retrain the entire general-purpose model.

[0325] In other words, the personality model in this invention refers to a set of parameters, or equivalent feature sets, that are personalized to reflect a specific individual's thought process, values, decision-making tendencies, expression style, etc., based on a general-purpose large-scale language model (LLM). Technically, the personality model consists of the portion of the entire set of parameters constituting the large-scale language model that has been updated by using data specific to the individual (speech history, behavioral logs, evaluation data, etc.) for learning or tuning, or a set of personalized parameters added to the general-purpose model. This set of parameters may exist in a form that is mixed into the entire model (integrated fine-tuning), or it may exist in a form that is separable from the general-purpose model (adapter layer, additional embedding vectors, external knowledge database, etc.).

[0326] Furthermore, while the above embodiments have described examples where the AI ​​agent is a pre-trained model (large-scale language model), this is not the only way to describe it. The AI ​​agent may consist of a pre-trained model (large-scale language model) and a knowledge database. Specifically, the pre-trained model is responsible for understanding the context of natural language and for inference, while the knowledge database functions as a source of factual information and up-to-date data. The knowledge database may include a structured knowledge graph or a vector database that performs similarity searches based on semantic vectors. The pre-trained model queries the knowledge database in response to user questions and instructions, and incorporates the search results into response generation. This allows for the incorporation of the latest information and specialized knowledge that the model alone cannot hold, while achieving contextually natural output.

[0327] (11) Others It should be noted that the present invention is not limited to the embodiments described above, and various modifications are included. For example, the embodiments described above are described in detail to make the present invention easier to understand, and are not necessarily limited to those having all the configurations described. Furthermore, it is possible to replace parts of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add configurations from other embodiments to the configuration of one embodiment. In addition, it is possible to add, delete, replace, or combine parts of the configuration of each embodiment with other configurations.

[0328] Furthermore, each of the above configurations, functions, processing units, and processing means may be implemented in hardware, in whole or in part, for example, by designing them as integrated circuits. Alternatively, each of the above configurations and functions may be implemented in software by a processor interpreting and executing programs that implement each function. Information such as programs, tables, and files that implement each function can be stored in memory, a recording device such as a hard disk or SSD (SolID State Drive), or a recording medium such as an IC card, SD card, or DVD. In addition, the processing performed by each server may be implemented by the processing of one or more servers having configurations different from those described above, or the functional modules that perform each processing may be implemented on servers different from those described above.

[0329] Furthermore, the control lines and information lines shown are those deemed necessary for explanatory purposes, and not all control lines and information lines are necessarily shown in the actual product. In reality, it is safe to assume that almost all components are interconnected. Furthermore, the above-described embodiments disclose at least the configuration described in the claims. Furthermore, the above-mentioned embodiments disclose at least the following: (1) AI agent provisioning system, An AI agent generation means that generates an AI agent that mimics the thinking style and behavioral patterns of the individual based on multifaceted data obtained regarding the individual's characteristics, An AI agent provisioning system comprising: a marketplace management means for managing a marketplace that provides the aforementioned AI agent to users as a tradable digital asset; and a marketplace management means for managing a marketplace that provides the aforementioned AI agent to users as a tradable digital asset. (2) The AI ​​agent provisioning system according to (1), further comprising an anonymization means for detecting personally identifiable information that can identify an individual linked to the AI ​​agent and anonymizing such personally identifiable information when the AI ​​agent is registered as a digital asset in the marketplace. (3) The aforementioned marketplace management means includes a billing means for calculating the usage fee for the AI ​​agent based on the user's usage history of the AI ​​agent, The AI ​​agent provision system according to (1) or (2), further comprising a revenue distribution means for distributing the revenue obtained by deducting a prescribed fee from the calculated usage fee to the owner of the AI ​​agent. (4) The AI ​​agent provisioning system according to (3), further comprising: the billing means selecting a billing model to be applied when calculating the usage fee based on the attributes of the AI ​​agent. (5) The marketplace management means further comprises a matching means for matching and searching for the AI ​​agent desired by the user. The matching means accepts the user's desired attributes of the AI ​​agent, including the user's personal profile and at least one of the intended use and output format of the AI ​​agent, and matches the user with the most suitable AI agent, as described in (1) or (2). (6) The AI ​​agent provision system according to (1) or (2), wherein the marketplace management means further comprises an agent provision means for forming a virtual team consisting of a plurality of AI agents selected by the user, or the AI ​​agents and humans. (7) The AI ​​agent providing system according to (1) or (2), wherein the AI ​​agent interprets an abstract goal set by the user in natural language, generates a concrete execution plan, and performs external cooperation processing to execute tasks in cooperation with an external system based on the plan. (8) The AI ​​agent generation means further comprises a personality model generation means that generates the AI ​​agent based on the multifaceted data, The AI ​​agent providing system according to (1) or (2), wherein the personality model generation means generates the personality model based on expert reflection, which is an analytical text that is generated from the multifaceted data relating to the acquired individual's characteristics and adds meaning and interpretation to the individual's statements and actions. (9) The AI ​​agent providing system according to (1) or (2), further comprising: an AI agent generation means that generates and maintains a reference vector relating to the characteristics of the individual; a deviation detection means that vectorizes the output of the AI ​​agent; calculates the distance from the reference vector; and outputs a warning when the calculated distance exceeds a predetermined threshold. (10) The AI ​​agent provisioning system according to (1) or (2), wherein the AI ​​agent generation means further comprises a thought process construction means that, when a question is input to the AI ​​agent, extracts information related to the question from a base memory associated with the personality model of the AI ​​agent, and uses the extracted information as a context to generate a response including a thought process. (11) The system further includes a personality model management means for continuously updating and managing the personality model of the AI ​​agent based on the output from the AI ​​agent or feedback from the original person who served as the model. The AI ​​agent providing system according to (1) or (2), wherein the personality model management means generates a gap analysis reflection that quantitatively shows the gap between data relating to the original person's self-perception and evaluation data from others around the person, and updates the personality model based on the gap analysis reflection. (12) The AI ​​agent providing system according to (11), wherein the personality model management means detects self-contradictions in past outputs using values ​​that were emphasized in past thought processes in the output history of the AI ​​agent, and updates the personality model by correcting deviations in thought processes based on the detected analysis results. (13) The AI ​​agent provisioning system according to (11), wherein the personality model management means generates a report summarizing the output content of the AI ​​agent, notifies the original person, interprets the feedback from the original person, and updates the personality model based on the feedback. (14) The AI ​​agent providing system according to (11), wherein the personality model management means integrates the personality models of the AI ​​agents registered in the marketplace and generates a role model ghost that shows an ideal thinking and behavioral pattern in a specific role. (15) An AI agent provisioning system in which the computer processor A personality model generation step that generates a personality model of the AI ​​agent that imitates the thinking style and behavioral patterns of the individual based on multifaceted data obtained regarding the individual's characteristics, A method for providing an AI agent, comprising: a marketplace management step of providing the aforementioned AI agent to a user as a tradable digital asset; and a method for providing an AI agent. (16) An AI agent providing program that is executed on a computer having a processor, wherein the AI ​​agent providing program causes the processor to execute the method described in (15). [Explanation of Symbols]

[0330] 100, 100B…AI agent provisioning system, 200…AI agent generation server, 201…Main memory, 210…Acquisition module, 210…Prompt memory module, 212…Reflection generation module, 212…Personality model generation module, 214…Personality model management module, 215…Question generation module, 216…Thought process construction module, 217…Vectorization module, 218…Deviation detection module, 219…Output module, 202…Auxiliary memory, 221…Expert information, 222…Agent Information, 223...Basic memory information, 224...Prompt configuration information, 225...Thought log information, 226...Activity log information, 400...Marketplace management server, 401...Main memory, 411...Agent registration module, 412...Anonymization module, 413...Agent provision module, 414...Billing module, 415...Revenue distribution module, 402...Secondary memory, 421...User information, 422...Registration information, 423...Evaluation information, 424...Matching information, 425...Usage history information, 600...Administrator terminal, 700...User terminal

Claims

1. This is an AI agent provisioning system, An AI agent generation means that generates an AI agent including a personality model that mimics the individual's thought patterns and behavioral patterns based on data acquired regarding the individual's characteristics, A marketplace management means for managing the marketplace that registers the aforementioned AI agent as a tradable digital asset and provides it to users, The system includes a personality model management means that updates and manages the personality model of the AI ​​agent based on the output from the AI ​​agent or feedback received from the individual who served as the model, The personality model management means is an AI agent provisioning system that integrates the personality models of multiple AI agents registered in the marketplace and generates a new personality model that shows thought and behavior patterns in a specific role.

2. The AI ​​agent provisioning system according to claim 1, further comprising an anonymization means for detecting personally identifiable information that can identify an individual linked to the AI ​​agent and anonymizing said personally identifiable information when the AI ​​agent is registered in the marketplace as a digital asset.

3. The aforementioned marketplace management means includes a billing means for calculating the usage fee for the AI ​​agent based on the user's usage history of the AI ​​agent, The AI ​​agent provision system according to claim 1 or 2, further comprising a revenue distribution means for distributing the revenue obtained by deducting a predetermined fee from the calculated usage fee to the owner of the AI ​​agent.

4. The AI ​​agent provision system according to claim 3, wherein the billing means further selects a billing model to be applied when calculating the usage fee based on the attributes of the AI ​​agent.

5. The marketplace management means further comprises a matching means for matching and searching for the AI ​​agent desired by the user. The AI ​​agent provisioning system according to claim 1 or 2, wherein the matching means accepts the user's desired attributes of the AI ​​agent, including the individual's profile and at least one of the intended use and output format of the AI ​​agent, and matches the user with a corresponding AI agent.

6. The AI ​​agent provision system according to claim 1 or 2, wherein the marketplace management means further comprises agent provision means for forming a virtual team consisting of a plurality of AI agents selected by a user, or the AI ​​agents and humans.

7. The AI ​​agent providing system according to claim 1 or 2, wherein the AI ​​agent generates an execution plan from a goal set by a user in natural language, and performs external cooperation processing to execute tasks in cooperation with an external system based on the said plan.

8. The AI ​​agent generation means generates the personality model based on expert reflection, which is an analytical text that is generated from the data acquired regarding the characteristics of the individual and adds meaning and interpretation to the individual's statements and actions, according to the AI ​​agent provisioning system according to claim 1 or 2.

9. The AI ​​agent providing system according to claim 1 or 2, further comprising: an AI agent generation means that generates and maintains a reference vector relating to the characteristics of the individual; a deviation detection means that vectorizes the output of the AI ​​agent; calculates the distance from the reference vector; and outputs a warning when the calculated distance exceeds a predetermined threshold.

10. The AI ​​agent provisioning system according to claim 1 or 2, wherein the AI ​​agent generation means further comprises a thought process construction means that, when a question is input to the AI ​​agent, extracts information related to the question associated with the personality model of the AI ​​agent, and generates a response from the extracted information that includes the thought process used to generate the answer.

11. The AI ​​agent providing system according to claim 1 or 2, wherein the personality model management means generates a gap analysis reflection showing the gap between data relating to the individual's self-perception and evaluation data from others around the individual, and updates the personality model based on the gap analysis reflection.

12. The AI ​​agent providing system according to claim 11, wherein the personality model management means updates the personality model by using information on values ​​that were emphasized in past thought processes in the output history of the AI ​​agent to detect inconsistencies in past outputs and correcting deviations in thought processes based on the detected analysis results.

13. The AI ​​agent provision system according to claim 11, wherein the personality model management means generates a report summarizing the output content of the AI ​​agent, notifies the individual concerned, obtains feedback from the individual concerned, and updates the personality model based on the feedback.

14. A method for providing an AI agent, wherein a computer, A personality model generation step that generates a personality model for an AI agent, which includes a personality model that mimics the thinking style and behavioral patterns of the individual, based on data obtained regarding the individual's characteristics. The aforementioned AI agent is registered as a tradable digital asset on the marketplace and provided to users as part of a marketplace management step, A personality model management step is performed in which the personality model of the AI ​​agent is updated and managed based on the output from the AI ​​agent or feedback received from the individual who served as the model. In the personality model management step, the computer integrates the personality models of multiple AI agents registered in the marketplace to generate a new personality model that shows thought and behavior patterns in a specific role, in an AI agent provision method.

15. The method for providing an AI agent according to claim 14, wherein when the computer registers the AI ​​agent as a digital asset in the marketplace, it further performs an anonymization step of detecting personally identifiable information that can identify an individual linked to the AI ​​agent and anonymizing the personally identifiable information.

16. The computer includes a billing step in which, in the marketplace management step, the computer calculates the usage fee for the AI ​​agent based on the user's usage record of the AI ​​agent, The method for providing an AI agent according to claim 14 or 15, further comprising a revenue distribution step of distributing the revenue obtained by deducting a predetermined fee from the calculated usage fee to the owner of the AI ​​agent.

17. The method for providing an AI agent according to claim 16, wherein the computer further selects a billing model to be applied when calculating the usage fee based on the attributes of the AI ​​agent in the billing step.

18. An AI agent providing program that is executed on a computer having a processor, wherein the AI ​​agent providing program causes the processor to execute the method according to claim 14 or 15.

Images