A human-machine collaborative AIAgent task management and seamless handover method and system

Abstract

The application provides a man-machine cooperative AI Agent task management and seamless switching method and system, and relates to the technical field of data processing.The method comprises the following steps: according to a manual intervention operation instruction issued by a manual operation end of a manual processing node, making an AI Agent exit a current follow-up conversation, and converting a task state from a manual state to a manual processing state; according to a result of converting the task state to the manual processing state, triggering an auxiliary capability module to provide real-time auxiliary capability to the manual processing node based on current conversation content; and according to a received conversation end request, converting the task state from the manual processing state to a completed state, and triggering an archiving module to collect communication records of the AI Agent stage and communication records of the manual processing node stage into a same task archive, and performing unified database storage.The application can obtain a communication context generated in the AI Agent stage and a manual summary, and avoids repeated expression of a business object.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for human-machine collaborative AI agent task management and seamless handover. Background Technology

[0002] In existing AI Agent task management systems, when an AI Agent encounters a complex scenario that it cannot handle during a conversation with a business object, it typically triggers manual intervention by directly terminating the conversation and generating a work order.

[0003] Specifically, when the AI ​​Agent engages in text conversations with business entities in scenarios such as credit collection, insurance claims, or customer service, it will terminate the conversation and generate a work order record in an independent work order system if it detects that the business entity expresses an intention to switch to human intervention, exhibits strong negative emotions, exceeds the AI ​​Agent's capability boundaries, or the business entity rejects multiple responses from the AI ​​Agent.

[0004] The work order record only includes a simple explanation of the transfer reason and does not carry the complete session context generated in the AI ​​Agent stage. After receiving the work order in the work order system, the manual processing node needs to re-initiate a session connection using the contact information of the business object recorded in the work order, and re-ask the business object for basic information, background of the request, and content previously communicated with the AI ​​Agent. This causes the business object to repeat the information already provided, resulting in low communication efficiency. At the same time, the session records of the AI ​​Agent stage are stored in the Agent system database, while the session records of the manual processing stage are stored in the human customer service system. The communication records of the two stages are stored separately, making it impossible to form a complete communication archive under the same task. This leads to the need to piece together data across systems for subsequent quality inspection, auditing, and dispute tracing, resulting in data gaps and difficulties in tracing. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method and system for AI Agent task management and seamless handover in human-machine collaboration, which can obtain the communication context generated in the AI ​​Agent stage and the summary transferred to human, so as to avoid repeated descriptions of business objects.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a method for AI agent task management and seamless handover in human-machine collaboration, the method comprising: Based on the task creation request received in the unified task management interface, obtain the business object information carried in the task creation request, create a task record with the task as the core management entity in the unified task management module, and bind a preset AI Agent to the task. Based on the task binding result, the AI ​​Agent is triggered to automatically start a follow-up session on the business object and set the task status to in progress. Based on the conversation context generated by the AI ​​Agent during the follow-up conversation, the human intervention decision model built into the Agent engine module performs multi-factor scoring on the conversation context to obtain the human intervention score result. Based on the comparison between the score of the transfer to human intervention and the preset threshold, when it is determined that a transfer to human intervention is required, the transfer to human intervention tool module is invoked to generate a structured transfer to human intervention notification, and the task status is changed from in progress to human intervention required. Based on the structured notification to be converted to manual processing, the status requiring manual processing is synchronized in real time to the manual operation terminal where the preset manual processing node is located through the status synchronization module. Based on the manual intervention operation command issued by the manual processing node at the manual operation terminal, the AI ​​Agent exits the current follow-up session and changes the task status from the manual intervention required state to the manual processing state. Based on the result of the task status transitioning to manual processing status, the auxiliary capability module is triggered to provide real-time auxiliary capabilities to the manual processing node based on the current session content; Based on the received session end request, the task status is changed from manual processing to completed, and the archiving module is triggered to collect the communication records of the AI ​​Agent stage and the communication records of the manual processing stage into the same task file and execute unified database storage.

[0007] Secondly, a human-machine collaborative AI agent task management and seamless handover system includes: The task management module is used to receive task creation requests in a unified task management interface, create task records with tasks as the core management entity, bind preset AI agents to tasks, and maintain and display task status. The business object management module is used to maintain existing business object information and support the entry of new business objects; The Agent engine module is used to drive the AI ​​Agent to automatically start follow-up conversations with business objects on communication channels, and make a judgment on transferring to human based on the conversation context. When it is determined that a transfer to human is required, the human transfer tool module is called. The "Transfer to Human Agent" module receives the business instance identifier, channel identifier, and transfer to human agent summary from the Agent engine module, and generates a structured transfer to human agent notification. The status synchronization module is used to synchronize the task status in real time between the Agent engine module and the human operation terminal. When the task status changes to the human-required status, the human-required status and the human-required summary are broadcast to the human operation terminal where the preset human processing node is located. The manual intervention module is used to receive manual intervention operation instructions issued by the manual processing node at the manual operation terminal, send a termination instruction to the Agent engine module, cause the AI ​​Agent to exit the current follow-up session, and switch the task status to manual processing status. The auxiliary capability module is used to provide real-time auxiliary capabilities to the manual processing node based on the current session content during the session. The real-time auxiliary capabilities include at least one of the following: knowledge assistance, compliance verification, dynamic summary, and to-do list organization. The archiving module is used to switch the task status to completed after receiving a session end request, and to collect the communication records of the AIAgent stage and the communication records of the manual processing stage into the same task archive for unified storage.

[0008] The above-described solution of the present invention has at least the following beneficial effects: By unifying the AI ​​Agent phase and the manual processing phase into a framework centered on tasks, a conversation context-driven model for determining whether to switch to manual processing is used to score the conversation context using multiple factors, automatically triggering the manual switching process. The status synchronization module then synchronizes the required manual status and structured manual switching notifications to the manual operation terminal in real time. This allows the manual processing node to directly continue the conversation on the original communication channel when taking over the conversation without rebuilding the connection. Furthermore, it can obtain the communication context and manual switching summary generated in the AI ​​Agent phase, avoiding redundant descriptions of business objects.

[0009] Meanwhile, the archiving module collects communication records from the AI ​​Agent stage and the manual processing stage into the same task file and stores them in a unified database, achieving complete traceability of the entire task lifecycle. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating a human-machine collaborative AI Agent task management and seamless handover method provided by an embodiment of the present invention.

[0011] Figure 2 This is a schematic diagram of an AI Agent task management and seamless transfer system for human-machine collaboration provided by an embodiment of the present invention. Detailed Implementation

[0012] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough understanding of the present disclosure and to fully convey its scope to those skilled in the art.

[0013] like Figure 1 As shown, to enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention. In the present invention, AI Agent refers to an intelligent program entity driven by a large language model that can automatically interact with business objects in a communication channel; business object refers to the communication object targeted by the task, such as a borrower in a credit collection scenario, an insured in an insurance claims scenario, or a customer in a customer service scenario; human processing node refers to a human agent or human business person who takes over the conversation and continues to communicate with the business object; communication channel refers to the communication carrier through which the AI ​​Agent executes follow-up conversations and the human processing node conducts conversations, including but not limited to instant messaging interfaces, SMS gateways, and email systems.

[0014] Example 1 This embodiment provides a method for AI agent task management and seamless handover in human-machine collaboration, applied to a human-machine collaborative AI agent task management and seamless handover system. For example... Figure 1 As shown, Figure 1 This is a flowchart illustrating a human-machine collaborative AI Agent task management and seamless handover method provided in this embodiment. The method includes the following steps: Step S101: Based on the task creation request received in the unified task management interface, obtain the business object information carried in the task creation request, create a task record with the task as the core management entity in the unified task management module, and bind a preset AI Agent to the task.

[0015] Specifically, the unified task management interface is the unified operation entry point provided by the task management module. Task creation requests are initiated through this interface. Task creation requests can be triggered in two ways: selecting an existing business object and entering a new business object. When selecting an existing business object, the system selects the target business object from the list maintained by the business object management module. The system automatically imports the existing information of that business object and allows modification of the imported information, such as changing the channel number for testing. When entering a new business object, the system quickly fills in the fields according to a preset information template, which pre-sets the names and filling formats for each field. After successful task creation, the unified task management module generates a unique task identifier for the task and creates a task record with the task as the core management entity. The task record includes at least the task identifier, business object information, and task status. Simultaneously, a preset AI Agent is bound to the task according to preset binding rules. These rules can be: based on the correspondence between business object categories and AI Agent business capability tags, select an AI Agent whose business capability tag matches the business object category for binding.

[0016] Step S102: Based on the task binding result, trigger the AI ​​Agent to automatically start a follow-up session on the business object and set the task status to in progress.

[0017] Specifically, after successful task binding, the Agent engine module automatically establishes a session connection with the business object and sends the first round of session messages according to the communication channels and channel numbers in the task record, thus initiating a follow-up session. Once the follow-up session is initiated, the unified task management module sets the task status to "in progress," indicating that the AI ​​Agent is actively following up on the task. It should be noted that the task status in this invention is managed and transitioned using a finite state machine. A finite state machine is a state management mechanism consisting of a finite number of states and transition rules between states; the specific transition rules are detailed in Embodiment Seven.

[0018] Step S103: Based on the conversation context generated by the AI ​​Agent during the follow-up conversation, the conversation context is scored by a multi-factor scoring model built into the Agent engine module to obtain the human scoring result.

[0019] The session context refers to all session messages and their corresponding business information accumulated during the follow-up session, including the message content sent by the business object, the message content replied by the AI ​​Agent, and the business elements extracted during the session. The specific implementation process of multi-factor scoring is described in detail in Example 3.

[0020] Step S104: Based on the comparison between the score of the transfer to human intervention and the preset threshold, when it is determined that a transfer to human intervention is required, the transfer to human intervention tool module is invoked to generate a structured transfer to human intervention notification, and the task status is changed from in progress to human intervention required.

[0021] Specifically, the preset threshold is a comparison benchmark pre-set according to the sensitivity requirements of the business scenario for transferring to human intervention. Its value ranges from 0 to 1; the higher the sensitivity requirements of the business scenario for transferring to human intervention, the smaller the preset threshold is set. The human intervention score is compared with the preset threshold: when the human intervention score is greater than or equal to the preset threshold, it is determined that human intervention is required. The Agent engine module calls the human intervention tool module and passes the business instance identifier of the current task, the channel identifier of the communication channel, and the human intervention summary to the human intervention tool module. The human intervention tool module generates a structured human intervention notification and simultaneously changes the task status from in progress to human intervention required, indicating that the task requires human intervention. When the human intervention score is less than the preset threshold, it is determined that human intervention is not required. The AI ​​Agent continues to follow up on the conversation and returns to step S103 to continue multi-factor scoring. The business instance identifier is a unique business identifier for the current task, used to uniquely identify a task instance within the system. The specific generation process of the structured human intervention notification is described in detail in Example 4.

[0022] Step S105: Based on the structured notification to switch to manual processing, the status requiring manual processing is synchronized in real time to the manual operation terminal where the preset manual processing node is located through the status synchronization module.

[0023] Specifically, after detecting a change in task status to require manual intervention, the status synchronization module broadcasts the manual intervention status and the structured notification to the preset manual processing node in real time. The status synchronization module can use message queues or long connections to achieve real-time broadcasting, which are conventional network communication technologies in this field and will not be elaborated upon here.

[0024] Step S106: Based on the manual intervention operation instruction issued by the manual processing node at the manual operation terminal, AIAgent is made to exit the current follow-up session, and the task status is changed from the manual intervention state to the manual processing state.

[0025] Specifically, after receiving the manual intervention instruction from the manual processing node at the manual operation terminal, the manual intervention module sends a termination instruction to the Agent engine module. Upon receiving the termination instruction, the Agent engine module controls the AI ​​Agent to stop generating response messages and exit the current follow-up session, returning an exit flag to the manual intervention module. Upon receiving the exit flag, the manual intervention module changes the task status from "requires manual intervention" to "manual processing," indicating that the manual processing node has taken over the session. The manual processing node directly communicates with the business object through the same communication channel, and the business object perceives it as a continuation of the same dialogue, without needing to switch session windows or re-establish session connections, achieving seamless takeover. The specific process of generating the manual intervention instruction is described in detail in Example 5.

[0026] Step S107: Based on the result of the task status transitioning to manual processing status, the auxiliary capability module is triggered to provide real-time auxiliary capabilities to the manual processing node based on the current session content.

[0027] Specifically, after the task status transitions to manual processing status, the auxiliary capability module runs in parallel with the conversation stream, performing real-time analysis of the current conversation content and providing the manual processing node with at least one real-time auxiliary capability among knowledge assistance, compliance verification, dynamic summarization, and task organization. The specific implementation process is detailed in Example Six. It should be noted that the above-mentioned real-time auxiliary capabilities are implemented based on the same large language model and knowledge base as the AI ​​Agent stage, ensuring the consistency of capabilities between the AI ​​Agent stage and the manual processing stage.

[0028] Step S108: Based on the received session end request, change the task status from manual processing status to completed status, and trigger the archiving module to collect the communication records of the AI ​​Agent stage and the communication records of the manual processing node stage into the same task file and perform unified database storage.

[0029] Specifically, after receiving a session end request, the archiving module first determines the processing entity of the current task: if the processing entity is an AI Agent, the AI ​​Agent performs the end operation; if the processing entity is a human processing node, the human processing node triggers the end and archive button on the human operation terminal, generating a session end request. The task status is changed to the completed state, which indicates that the task has ended and archiving has been completed. The specific implementation process of unified database storage is described in detail in Example 8.

[0030] Example 2 This embodiment provides a detailed description of the specific composition of the business object information in Embodiment 1 above. The business object information includes at least one of the following: communication channel, channel number, business object category, associated product type, pending processing time, and pending processing amount.

[0031] Among them, the channel number refers to the contact identifier of the business object on the communication channel, such as a mobile phone number, instant messaging account, or email address. The channel number is anonymized during storage and display, and in actual implementation, the corresponding communication channel is accessed using a standard channel access protocol; the business object classification is used to identify the business category to which the business object belongs, such as classifying business objects into credit business objects, insurance business objects, or customer service business objects according to business type; the associated product type refers to the type of business products associated with the business object; the pending processing time refers to the duration of the pending matters corresponding to the task; and the pending processing amount refers to the amount involved in the pending matters corresponding to the task.

[0032] The communication channel serves as the communication carrier for the AI ​​Agent to execute follow-up conversations and for the manual processing node to conduct conversations. During the AI ​​Agent phase, the AI ​​Agent conducts follow-up conversations with the business object through the communication channel; after the manual processing node takes over the conversation, the manual processing node continues the conversation with the business object using the same communication channel, thereby achieving channel continuity and eliminating the need for the business object to switch communication channels.

[0033] Example 3 This embodiment provides a detailed explanation of the specific implementation process of the manual judgment model in Embodiment 1 above for multi-factor scoring of the session context, including the following steps: Step S301: Based on the conversation context, extract the score for identifying the intention of the human request, the score for the intensity of negative emotions, the value for exceeding the capability boundary, and the count of repeated rejections.

[0034] Specifically, the large language model performs semantic understanding of the conversation context through system prompts, extracting four factors: First, it determines whether the business object explicitly requests human assistance. When the business object explicitly requests human assistance, the human request intent recognition score is 1; otherwise, it is 0. Second, it analyzes the emotional tendency of the message sent by the business object, obtaining a negative emotion intensity score. The negative emotion intensity score ranges from 0 to 1; the more negative the business object's emotion, the closer the negative emotion intensity score is to 1. Third, it determines whether the business object's current request exceeds the preset capability boundary of the AI ​​Agent. If it does, the exceeding capability boundary marker value is 1; otherwise, it is 0. Here, the preset capability boundary refers to the range of business that can be handled by the AI ​​Agent in advance. Fourth, it counts the number of times the business object continuously rejects the AI ​​Agent's response, obtaining the repeated rejection count, and then normalizes the repeated rejection count to obtain the repeated rejection count value. The specific rule for the normalization process is: divide the repeated rejection count by the preset maximum repeated rejection count, and use the quotient as the repeated rejection count value; when the repeated rejection count is greater than the preset maximum repeated rejection count, the repeated rejection count value is 1. The large language model achieves the above-mentioned intent recognition and semantic understanding through system prompts and tool invocation mechanisms, which is a conventional technical means in this field.

[0035] Step S302: Multiply the score for recognizing the intention of the human request by its corresponding first preset weight coefficient to obtain a first weighted value; multiply the score for the intensity of negative emotion by its corresponding second preset weight coefficient to obtain a second weighted value; multiply the value for exceeding the capability boundary by its corresponding third preset weight coefficient to obtain a third weighted value; multiply the count of repeated rejections by its corresponding fourth preset weight coefficient to obtain a fourth weighted value; sum the first, second, third, and fourth weighted values ​​to obtain the result of transferring the score to human review.

[0036] The sum of the first preset weight coefficient, the second preset weight coefficient, the third preset weight coefficient, and the fourth preset weight coefficient is 1. Each preset weight coefficient is pre-configured according to the sensitivity of the business scenario to the corresponding factor. The more sensitive the business scenario is to a certain factor, the larger the preset weight coefficient corresponding to that factor is configured.

[0037] By using the above multi-factor scoring method, the four dimensions of human request intent, negative emotions, capability boundaries, and repeated rejections are uniformly quantified into a human referral score. Combined with preset thresholds, this enables accurate judgment of the timing of human referral, avoiding premature referrals that waste human resources and delayed referrals that lead to a decline in the user experience.

[0038] Example 4 This embodiment provides a detailed description of the specific generation process of the structured notification to manual notification in Embodiment 1 above, including the following steps: Step S401: Based on the session context, the Agent engine module automatically identifies the core requirements of the business object and obtains a description of the core requirements. Specifically, the large language model performs semantic summarization on the messages sent by the business object in the session context, extracts the core problem that the business object needs to solve in this session, and generates a description of the core requirements according to a preset description format.

[0039] Step S402: Based on the session context, the Agent engine module determines the reason why the AI ​​Agent cannot handle the situation and obtains a description of the reason for the inability to handle it. Specifically, the reason why the AI ​​Agent cannot handle the situation is determined according to the type of factor that triggered the transfer to human intervention: when the score for identifying the intent to request human intervention is 1, the reason for the inability to handle the situation is that the business object explicitly requests human service; when the value of the "exceeds capability boundary" flag is 1, the reason for the inability to handle the situation is that the business object's request exceeds the scope of business that the AI ​​Agent can handle; when the score for negative emotion intensity or the count of repeated rejections exceeds their respective warning limits, the reason for the inability to handle the situation is that the business object's emotions have intensified or that they continue to disagree with the AI ​​Agent's response.

[0040] Step S403: Based on the session context, the Agent engine module extracts key business information to obtain a key business information summary. This key business information includes product information, amount information, time information, and commitment information mentioned during the session. Specifically, the large language model extracts elements from the session context and organizes the extracted elements according to element type to form a key business information summary.

[0041] Step S404: Generate a summary for transferring to human intervention based on the core request description, the reason for the inability to process the request, and the key business information summary. Specifically, according to the preset summary template, fill in the corresponding field positions in the preset summary template with the core request description, the reason for the inability to process the request, and the key business information summary to generate the summary for transferring to human intervention.

[0042] Step S405: Based on the human intervention summary, the business instance identifier of the current task, and the channel identifier of the communication channel, a structured human intervention notification is generated. Specifically, the structured human intervention notification includes a notification identifier, a business instance identifier, a channel identifier, a human intervention summary, and a notification generation time. The human intervention tool module assembles the above content according to a preset structured data format to generate the structured human intervention notification. Through the structured notification content, the human processing node can quickly understand the overall task after receiving the information.

[0043] Example 5 This embodiment provides a detailed description of the specific process for generating the manual intervention operation command in Embodiment 1 above, including the following steps: Step S501: Based on the "Need Human Intervention" status synchronized to the human intervention terminal by the status synchronization module, a summary of the transfer to human intervention and past communication content from the AI ​​Agent stage are displayed in the unified task management interface on the human intervention terminal. Specifically, after receiving the "Need Human Intervention" status, the human intervention terminal displays a "Need Human Intervention" tag for the task in the task list of the unified task management interface; after the human processing node clicks on the task, the unified task management interface displays a summary of the transfer to human intervention and past communication content from the AI ​​Agent stage. The human processing node can quickly understand the overall situation of the task based on this, without having to re-ask the basic information and background of the business object, thus shortening the intervention preparation time.

[0044] In step S502, the manual processing node triggers the manual intervention button or clicks the takeover link on the manual operation terminal to generate a manual intervention operation instruction. Specifically, a manual intervention button is set in the unified task management interface, and the structured to manual notification carries a takeover link; when the manual processing node clicks the manual intervention button or clicks the takeover link, the manual operation terminal generates a manual intervention operation instruction carrying a task identifier and sends the manual intervention operation instruction to the manual intervention module.

[0045] Example 6 This embodiment provides a detailed description of the specific implementation process of the auxiliary capability module in Embodiment 1 above, which provides real-time auxiliary capabilities to the manual processing node based on the current session content, including the following steps: Step S601: Based on the current question of the business object in the current session content, retrieve standard response content from the knowledge base and push the standard response content to the human operation terminal in real time. Specifically, extract the semantic features of the current question of the business object, perform similarity matching between the semantic features and the semantic features of each knowledge entry in the knowledge base, and push the standard response content corresponding to the knowledge entry with the highest similarity to the human operation terminal in real time for reference or selection by the human processing node.

[0046] Step S602: Based on preset compliance rules, the response content entered by the manual processing node in the current session is detected in real time. When a violation is detected, a warning is issued on the manual operation terminal. Specifically, the preset compliance rules include a prohibited word library and a violation expression pattern; the response content entered by the manual processing node is matched with the prohibited word library, and the response content is pattern recognized according to the violation expression pattern; when the response content matches a prohibited word or a violation expression pattern is identified, it is determined that the response content contains a violation, and a warning is issued on the manual operation terminal through a pop-up prompt or a highlighted mark.

[0047] Step S603: Based on the current conversation content, extract key negotiation information in real time and display it in the sidebar of the manual operation terminal. The key negotiation information includes the amount, timeframe, and key points of the negotiated plan. Specifically, the large language model extracts elements from the current conversation content in real time, and dynamically updates and displays the extracted key negotiation information in the sidebar of the manual operation terminal.

[0048] Step S604: Based on the current session content, an unfinished task list is automatically generated upon the end of the session and displayed on the human operator's terminal. The unfinished task list includes tasks agreed upon during the session and their corresponding deadlines. Specifically, upon receiving a session end request, the large language model summarizes the current session content, extracts the unfinished tasks agreed upon during the session and their deadlines, generates the unfinished task list, and displays it on the human operator's terminal.

[0049] Example 7 This embodiment details the specific rules for transitioning task states using a finite state machine. The finite state machine includes four states: in progress, requiring manual intervention, requiring manual processing, and completed; and four types of triggering events: request for manual intervention, request for manual intervention, request for archiving, and request for immediate termination. The state transition rules are as follows: When a task is in progress and the triggering event is a request to transfer to human intervention, the task status changes to a human intervention required status. When a task is in progress and the triggering event is an archive request, the task status changes to the completed status. When a task is in progress and the triggering event is a direct termination request, the task status changes to the completed status. When a task is in a state requiring manual intervention and the triggering event is manual intervention, the task status changes to a manual processing state. When a task is in manual processing and the triggering event is an archive request, the task status changes to completed.

[0050] Except for the transition paths specified in the above state transition rules, all other state transition paths are prohibited. This finite state machine ensures the determinism and predictability of task state transitions, avoids state deadlock or illegal jumps, and enables real-time tracking of the task's state throughout its entire lifecycle.

[0051] Example 8 This embodiment provides a detailed explanation of the specific implementation process in Embodiment 1 above, which involves aggregating communication records from the AI ​​Agent stage and the manual processing stage into the same task file and performing unified database storage. The process includes the following steps: Step S801: Based on the task identifier, associate the communication records of the AI ​​Agent stage and the communication records of the manual processing node stage under the same task identifier, so that the communication records of the two stages belong to the same task file.

[0052] Step S802: Add a session type marker as Agent session to the communication records in the AI ​​Agent stage, and add a session type marker as Manual session to the communication records in the manual processing node stage, in order to distinguish the processing stage to which the communication records belong.

[0053] Step S803: Record the timestamps for each stage. The timestamps for each stage include the Agent start timestamp, the transition to human intervention timestamp, the human intervention timestamp, and the archiving timestamp. Among them, the Agent start timestamp is the time when the AI ​​Agent starts the follow-up session, the transition to human intervention timestamp is the time when the task status changes to the human intervention status, the human intervention timestamp is the time when the task status changes to the human processing status, and the archiving timestamp is the time when the unified database is written.

[0054] Step S804: Record the processing entity identifier, which includes the identifier of the AI ​​Agent and the identifier of the manual processing node, and is used to identify the processing entity of the communication record at each stage.

[0055] Step S805 involves atomically writing the session type marker, timestamps for each stage, processing entity identifier, and complete session content to the archive database using a database transaction mechanism. This database transaction mechanism binds multiple write operations into a single execution: when all write operations succeed, the results are committed uniformly; when any write operation fails, all write operations are rolled back. This mechanism ensures atomic writing of data between the AI ​​Agent stage and the manual processing stage, preventing data inconsistencies caused by partial archiving. After unified storage, a complete session archive is formed under the same task. Subsequent quality inspection, auditing, dispute tracing, and model training can directly access the complete session archive without needing to stitch data across systems.

[0056] Example 9 This embodiment provides a human-machine collaborative AI agent task management and seamless transfer system, which implements the methods described in any of the above embodiments. For example... Figure 2 As shown, Figure 2 This is a schematic diagram of a human-machine collaborative AI Agent task management and seamless transfer system provided in this embodiment. The system includes: a task management module, a business object management module, an Agent engine module, a manual transfer tool module, a status synchronization module, a manual intervention module, an auxiliary capability module, and an archiving module.

[0057] The task management module is used to receive task creation requests in a unified task management interface, create task records with tasks as the core management entity, bind preset AI agents to tasks, and maintain and display task status. The business object management module is used to maintain existing business object information and support the entry of new business objects; The Agent engine module is used to drive the AI ​​Agent to automatically start follow-up conversations with business objects on communication channels, and make a judgment on transferring to human based on the conversation context. When it is determined that a transfer to human is required, the human transfer tool module is called. The "Transfer to Human Agent" module receives the business instance identifier, channel identifier, and transfer to human agent summary from the Agent engine module, and generates a structured transfer to human agent notification. The status synchronization module is used to synchronize the task status in real time between the Agent engine module and the human operation terminal. When the task status changes to the human-required status, the human-required status and the human-required summary are broadcast to the human operation terminal where the preset human processing node is located. The manual intervention module is used to receive manual intervention operation instructions issued by the manual processing node at the manual operation terminal, send a termination instruction to the Agent engine module, cause the AI ​​Agent to exit the current follow-up session, and switch the task status to manual processing status. The auxiliary capability module is used to provide real-time auxiliary capabilities to the manual processing node based on the current session content during the session. The real-time auxiliary capabilities include at least one of the following: knowledge assistance, compliance verification, dynamic summary, and to-do list organization. The archiving module is used to switch the task status to completed after receiving a session end request, and to collect the communication records of the AIAgent stage and the communication records of the manual processing stage into the same task archive for unified storage.

[0058] The interaction relationships between the above modules are as follows: the task management module interacts with the business object management module to complete task creation; the agent engine module triggers the transfer through the manual intervention tool module; the status synchronization module broadcasts the task status change to the task management module and the manual intervention module; the manual intervention module sends a termination command to the agent engine module; the auxiliary capability module runs in parallel with the session stream; and the archiving module aggregates all data and performs unified database storage when the task is completed.

[0059] In summary, this invention integrates the AI ​​Agent phase and the manual processing phase into a framework centered on tasks as the core management entity. Based on a conversation context-driven model for determining whether to switch to manual processing, it performs multi-factor scoring on the conversation context, automatically triggering the manual switching process. Furthermore, a status synchronization module synchronizes the required manual status and structured manual switching notifications to the manual operation terminal in real time. This allows manual processing nodes to directly continue the conversation from the existing communication channels when taking over the session, without needing to rebuild the connection. They can also obtain the communication context and manual switching summary generated in the AI ​​Agent phase, avoiding redundant descriptions of business objects. Simultaneously, an archiving module collects the communication records from both the AI ​​Agent and manual processing phases into a unified task file, achieving complete traceability throughout the entire task lifecycle.

[0060] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for human-machine collaborative AI agent task management and seamless handover, characterized in that, include: Based on the task creation request received in the unified task management interface, obtain the business object information carried in the task creation request, create a task record with the task as the core management entity in the unified task management module, and bind a preset AI Agent to the task. Based on the task binding result, the AI ​​Agent is triggered to automatically start a follow-up session on the business object and set the task status to in progress. Based on the conversation context generated by the AI ​​Agent during the follow-up conversation, the human intervention decision model built into the Agent engine module performs multi-factor scoring on the conversation context to obtain the human intervention score result. Based on the comparison between the score of the transfer to human intervention and the preset threshold, when it is determined that a transfer to human intervention is required, the transfer to human intervention tool module is invoked to generate a structured transfer to human intervention notification, and the task status is changed from in progress to human intervention required. Based on the structured notification to be converted to manual, the status requiring manual processing is synchronized in real time to the manual operation terminal where the preset manual processing node is located through the status synchronization module. Based on the manual intervention operation command issued by the manual processing node at the manual operation terminal, the AI ​​Agent exits the current follow-up session and changes the task status from the manual intervention required state to the manual processing state. Based on the result of the task status transitioning to manual processing status, the auxiliary capability module is triggered to provide real-time auxiliary capabilities to the manual processing node based on the current session content; Based on the received session end request, the task status is changed from manual processing to completed, and the archiving module is triggered to collect the communication records of the AI ​​Agent stage and the communication records of the manual processing stage into the same task file and execute unified database storage.

2. The method according to claim 1, characterized in that, The business object information includes at least one of the following: communication channel, channel number, business object classification, associated product type, pending processing time, and pending processing amount; The communication channel serves as the communication carrier for the AI ​​Agent to execute follow-up sessions and for the human processing nodes to conduct sessions.

3. The method according to claim 1, characterized in that, The model for determining whether to use a human agent performs multi-factor scoring on the conversation context to obtain a human agent scoring result, including: Based on the conversation context, extract the score for recognizing the human request intent, the score for the intensity of negative emotions, the value for exceeding the ability boundary, and the count of repeated rejections; Based on the score for recognizing the intention of the human request, the score for the intensity of negative emotions, the value of exceeding the ability boundary, and the count of repeated rejections, the scores are weighted and summed with the corresponding preset weight coefficients to obtain the result of the human scoring. Among them, the score for recognizing the intention of the human request is 0 or 1, the score for the intensity of negative emotion is between 0 and 1, the score for exceeding the ability boundary is 0 or 1, the repeated rejection count is normalized to the range of 0 to 1, and the sum of each preset weight coefficient is 1.

4. The method according to claim 1, characterized in that, The generation of the structured notification to human notification includes: Based on the session context, the Agent engine module automatically identifies the core requirements of the business object and obtains a description of the core requirements. Based on the session context, the Agent engine module determines the reason why the AI ​​Agent cannot handle the task and obtains a description of the reason why it cannot be handled. Based on the session context, the Agent engine module extracts key business information to obtain a key business information summary. Generate a summary for manual processing based on the core request description, the reason for the inability to process the request, and a summary of key business information. A structured notification for transferring to a human operator is generated based on the summary of the transfer to human operator, the business instance identifier of the current task, and the channel identifier of the communication channel.

5. The method according to claim 1, characterized in that, The manual intervention instructions issued by the manual processing node at the manual operation terminal include: Based on the status requiring human intervention synchronized to the human operation terminal by the status synchronization module, the summary of the transfer to human intervention and the past communication content of the AI ​​Agent stage are displayed in the unified task management interface on the human operation terminal. The manual processing node triggers the manual intervention button or clicks the takeover link on the manual operation terminal to generate a manual intervention operation command.

6. The method according to claim 1, characterized in that, The triggering assistance module provides real-time assistance capabilities to the manual processing node based on the current session content, including: Based on the current question of the business object in the current conversation content, retrieve the standard response content from the knowledge base and push the standard response content to the human operation terminal in real time; Based on preset compliance rules, the system monitors the responses entered by manual processing nodes in the current session in real time. When a violation is detected, a warning is issued on the manual operation terminal.

7. The method according to claim 6, characterized in that, The triggering assistance module provides real-time assistance capabilities to the manual processing node based on the current session content, and also includes: Based on the current conversation content, extract key negotiation information in real time and display it in the sidebar of the manual operation terminal; Based on the current conversation content, an to-do list is automatically generated when the conversation ends and displayed on the human operator's terminal.

8. The method according to any one of claims 1 to 7, characterized in that, The task states are transitioned according to a finite state machine, which includes: Status: In Progress, Requires Manual Handling, Requires Manual Processing, and Completed; When a task is in progress and the triggering event is a request to transfer to human intervention, the task status changes to a human intervention required status. When a task is in progress and the triggering event is an archive request or a direct termination request, the task status changes to the completed status. When a task is in a state requiring manual intervention and the triggering event is manual intervention, the task status changes to a manual processing state. When a task is in manual processing and the triggering event is an archive request, the task status changes to completed.

9. The method according to claim 1, characterized in that, The process of collecting communication records from the AI ​​Agent phase and the manual processing phase into the same task file and performing unified database storage includes: Based on the task identifier, the communication records of the AI ​​Agent stage and the communication records of the manual processing node stage are associated with the same task ID; Add a session type label to the communication records in the AI ​​Agent stage as Agent session, and add a session type label to the communication records in the manual processing node stage as manual session; Record timestamps for each stage, including Agent start timestamp, transition to human intervention timestamp, human intervention timestamp, and archiving timestamp. Record the identifier of the processing entity, which includes the identifier of the AI ​​Agent and the identifier of the human processing node; The session type marker, timestamps for each stage, processing entity identifier, and complete session content are atomically written to the archived database through a database transaction mechanism.

10. A human-machine collaborative AI agent task management and seamless transfer system, the system implementing the method as described in any one of claims 1 to 9, characterized in that, include: The task management module is used to receive task creation requests in a unified task management interface, create task records with tasks as the core management entity, bind preset AI agents to tasks, and maintain and display task status. The business object management module is used to maintain existing business object information and support the entry of new business objects; The Agent engine module is used to drive the AI ​​Agent to automatically start follow-up conversations with business objects on communication channels, and make a judgment on transferring to human based on the conversation context. When it is determined that a transfer to human is required, the human transfer tool module is called. The "Transfer to Human Agent" module receives the business instance identifier, channel identifier, and transfer to human agent summary from the Agent engine module, and generates a structured transfer to human agent notification. The status synchronization module is used to synchronize the task status in real time between the Agent engine module and the human operation terminal. When the task status changes to the human-required status, the human-required status and the human-required summary are broadcast to the human operation terminal where the preset human processing node is located. The manual intervention module is used to receive manual intervention operation instructions issued by the manual processing node at the manual operation terminal, send a termination instruction to the Agent engine module, cause the AI ​​Agent to exit the current follow-up session, and switch the task status to manual processing status. The auxiliary capability module is used to provide real-time auxiliary capabilities to the manual processing node based on the current session content during the session. The real-time auxiliary capabilities include at least one of the following: knowledge assistance, compliance verification, dynamic summary, and to-do list organization. The archiving module is used to switch the task status to completed after receiving a session end request, and to collect the communication records of the AIAgent stage and the communication records of the manual processing stage into the same task archive for unified storage.

Images