Business process reconstruction method, device and equipment based on intelligent agent

Abstract

The invention provides an agent-based business process reconstruction method, device and equipment, and relates to the technical field of education, and the method comprises the steps: obtaining a definition file of a target business process and a preset business constraint rule; constructing a directed acyclic graph of the target business process according to the definition file; performing iterative optimization of configuration parameters on the directed acyclic graph according to a business constraint rule to obtain an optimal configuration parameter combination of the directed acyclic graph; and reconstructing the directed acyclic graph according to the optimal configuration parameter combination to obtain a target directed acyclic graph corresponding to the target business process. According to the invention, the optimal man-machine cooperation process can be obtained according to the business process.

Description

Technical Field

[0001] This application relates to the field of educational technology, and more specifically, to a business process refactoring method, apparatus, and device based on intelligent agents. Background Technology

[0002] Whether in market-facing business scenarios or internal operational support scenarios, decision-makers face a common problem: during the project initiation and solution design phases, there is a lack of reliable methods to predict the efficiency of business processes after the introduction of AI (Artificial Intelligence), making it difficult to quantify its commercial returns, and even more difficult to automatically generate validated, risk-controlled optimal human-machine collaboration solutions. This often leads enterprises to rely on experience-based decisions or local pilot projects, resulting in high investment costs, long trial-and-error cycles, and difficulty in achieving large-scale, scientific intelligent upgrades. Summary of the Invention

[0003] This application provides a business process refactoring method, apparatus, and device based on intelligent agents, which can obtain the optimal human-machine collaboration process based on the business process.

[0004] In a first aspect, embodiments of this application provide a business process refactoring method based on intelligent agents, the method comprising: Obtain the definition file and preset business constraint rules of the target business process; Based on the definition file, construct a directed acyclic graph corresponding to the target business process; wherein, each process node in the directed acyclic graph includes: human nodes and intelligent agent nodes; According to the business constraint rules, the configuration parameters of the directed acyclic graph are iteratively optimized to obtain the optimal configuration parameter combination of the directed acyclic graph. The optimal configuration parameter combination includes: the optimal configuration parameters of each artificial node and the optimal configuration parameters of each intelligent agent node. Based on the optimal configuration parameter combination, the directed acyclic graph is reconstructed to obtain the target directed acyclic graph corresponding to the target business process.

[0005] Optionally, constructing the directed acyclic graph corresponding to the target business process based on the definition file includes: Parse the definition file to obtain each process node and the dependencies between nodes; Based on the process nodes and the dependencies between nodes, an initial directed acyclic graph is constructed; Based on the metadata tags of each process node in the initial directed acyclic graph, the type of each process node in the initial directed acyclic graph is identified, and the operation category tag of each process node is determined. The operation category tag is used to indicate whether the process node is a human node or an intelligent agent node. Add the operation category labels of each process node to the initial directed acyclic graph to obtain the directed acyclic graph corresponding to the target business process.

[0006] Optionally, the step of iteratively optimizing the configuration parameters of the directed acyclic graph according to the business constraint rules to obtain the optimal combination of configuration parameters for the directed acyclic graph includes: The preset strategy network corresponding to the business constraint rule selects a set of configuration parameter combinations from the corresponding action space; Based on the configuration parameter combination, adjust the configuration parameters of each process node in the directed acyclic graph, and obtain the simulation quantization performance data of the directed acyclic graph after the configuration parameter adjustment; A preset reward function is used to calculate the reward value corresponding to the combination of configuration parameters based on the simulated quantitative performance data. Based on the reward value, the policy network is updated using a near-end policy optimization algorithm; The updated policy network is used to select a new combination of configuration parameters from the action space and obtain a new reward value corresponding to the new combination of configuration parameters until the new reward value meets the preset stopping iteration condition. The optimal configuration parameter combination is the combination that yields the highest reward value when the preset stopping iteration condition is met.

[0007] Optionally, the step of reconstructing the directed acyclic graph according to the optimal configuration parameter combination to obtain the target directed acyclic graph corresponding to the target business process includes: Based on the optimal configuration parameter combination, the topology of the directed acyclic graph is optimized to obtain the reconstructed topology. Based on the optimal configuration parameter combination, a collaborative attribute is defined for the node relationship between agent nodes and human nodes in the reconstructed topology, resulting in the target directed acyclic graph. The collaborative attribute of the node relationship is used to indicate the collaborative mode of human-led agent assistance, the automatic mode of agent-led human supervision, or the shadow mode of parallel operation of agent and human.

[0008] Optionally, the method further includes: Using a purely manual baseline model, a purely manual operation simulation is performed on each process node in the target directed acyclic graph to obtain the manual indicator data corresponding to the target directed acyclic graph. A collaborative optimization model is used to perform collaborative simulation of manual operations and intelligent agent operations on each process node in the target directed acyclic graph, thereby obtaining collaborative index data corresponding to the target directed acyclic graph. Based on the artificial indicator data and the collaborative indicator data, a value assessment report corresponding to the target directed acyclic graph is generated.

[0009] Optionally, generating a value assessment report corresponding to the target directed acyclic graph based on the artificial indicator data and the collaborative indicator data includes: Based on the artificial indicator data and the collaborative indicator data, differential comparison calculation is performed to obtain quantitative performance comparison data; A preset value assessment strategy is adopted to evaluate the quantitative efficiency comparison data in multiple energy efficiency assessment dimensions, and the value assessment parameters of the target directed acyclic graph in the multiple energy efficiency assessment dimensions are obtained. The value assessment report is generated based on the value assessment parameters of the multiple energy efficiency assessment dimensions.

[0010] Optionally, after obtaining the value assessment parameters of the target directed acyclic graph in the multiple energy efficiency assessment dimensions, the method further includes: The value assessment parameters of the multiple energy efficiency assessment dimensions are normalized and mapped onto a preset multi-dimensional value radar map to obtain the value radar map corresponding to the target directed acyclic graph. Based on the value assessment parameters of the cost reduction assessment dimension among the value assessment parameters of the multiple energy efficiency assessment dimensions, draw the pure labor cost curve and the collaborative cost curve to obtain the cost recovery scissor difference diagram corresponding to the target directed acyclic graph. Based on the value assessment parameters of efficiency improvement assessment dimension and quality improvement assessment dimension among the value assessment parameters of the multiple energy efficiency assessment dimensions, the congestion index of each process node in the target directed acyclic graph is marked to obtain the process bottleneck heat map corresponding to the target directed acyclic graph.

[0011] Optionally, after defining cooperative attributes for the node relationships between agent nodes and artificial nodes in the reconstructed topology based on the optimal configuration parameter combination to obtain the target directed acyclic graph, the method further includes: Based on the target directed acyclic graph, generate a human-machine collaboration business process diagram and a standardized operating procedure corresponding to the target business process.

[0012] Secondly, embodiments of this application also provide a business process refactoring apparatus based on intelligent agents, the apparatus comprising: The acquisition module is used to acquire the definition file of the target business process and the preset business constraint rules; The mapping module is used to construct a directed acyclic graph of the target business process based on the definition file; wherein, each process node in the directed acyclic graph includes: human nodes and intelligent agent nodes; The calculation module is used to iteratively optimize the configuration parameters of the directed acyclic graph according to the business constraint rules, so as to obtain the optimal configuration parameter combination of the directed acyclic graph. The optimal configuration parameter combination includes: the optimal configuration parameters of each artificial node and the optimal configuration parameters of each intelligent agent node. The mapping module is also used to reconstruct the directed acyclic graph according to the optimal configuration parameter combination to obtain the target directed acyclic graph of the target business process.

[0013] Thirdly, embodiments of this application also provide an electronic device, including: a processor, a memory, and a bus, wherein the memory stores program instructions executable by the processor, and when the electronic device is running, the processor communicates with the memory via the bus, and the processor executes the program instructions to perform the steps of the agent-based business process refactoring method as described in any of the first aspects. This application provides a business process refactoring method, apparatus, and device based on intelligent agents. The method involves first obtaining a definition file of the target business process and preset business constraint rules; then, constructing a directed acyclic graph (DAG) corresponding to the target business process, including human nodes and intelligent agent nodes, based on the definition file; next, optimizing the configuration parameters of each node in the DAG through iterative optimization according to the business constraint rules, outputting a set of optimal configuration parameter combinations covering all process nodes; finally, refactoring the original DAG based on the optimal configuration parameter combinations to generate a target DAG representing the optimized business process, thereby completing the intelligent transformation of the business process and outputting a solution. This application constructs a directed acyclic graph by distinguishing between human nodes and intelligent agent nodes, which can clearly depict the characteristics of human-machine collaborative processes. It effectively overcomes the shortcomings of ambiguous node attributes and chaotic collaboration logic in traditional modeling. At the same time, it conducts iterative optimization of configuration parameters based on business constraint rules, which can achieve refined and intelligent adjustment of process parameters, ensuring that the parameter combination is highly adapted to actual business needs. Finally, the directed acyclic graph is reconstructed based on the optimal configuration parameters, which can optimize the node collaboration logic and topology, improve process operation efficiency and risk management capabilities, and help enterprises make scientific process reengineering decisions. Attached Figure Description To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 A flowchart illustrating a business process refactoring method based on intelligent agents provided in this application; Figure 2 A schematic diagram of the process for obtaining a directed acyclic graph in the agent-based business process refactoring method provided in this application; Figure 3 A schematic diagram illustrating the process of obtaining the optimal configuration parameter combination in a business process refactoring method based on intelligent agents provided in this application; Figure 4 A schematic diagram of the process for obtaining the target directed acyclic graph in the business process refactoring method based on intelligent agents provided in this application; Figure 5 A flowchart illustrating the generation of a value assessment report in a business process refactoring method based on intelligent agents provided in this application; Figure 6 A flowchart illustrating the generation of a value assessment report in another agent-based business process refactoring method provided in this application; Figure 7 A flowchart illustrating the generation of a visual view in a business process refactoring method based on intelligent agents provided in this application; Figure 8 A schematic diagram of a business process refactoring device based on intelligent agents provided in this application; Figure 9 A schematic diagram of an electronic device provided in this application. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0016] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0017] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0018] Before providing a detailed explanation of this application, let's first introduce its application scenarios.

[0019] The Education Technology Group's business processes involve diverse scenarios, including Type 1 processes for parents / students (such as study abroad enrollment consultation and course registration Q&A), Type 2 processes with high-frequency resource allocation and strong timeliness requirements (such as campus scheduling and teaching material supply chain coordination), Type 3 processes serving internal staff and teaching operations (such as attendance consultation and teaching research material review), and Type 4 processes with low frequency but high value (such as new campus construction planning and core curriculum system construction).

[0020] Among these, Type 1 processes often face challenges such as high-concurrency inquiries during enrollment seasons, lack of response at night, and loss of long-tail leads, directly impacting conversion efficiency and revenue growth. Type 2 processes are hampered by resource conflicts and response delays that can easily occur with manual scheduling, thus affecting the continuity of teaching services. Type 3 processes suffer from high operating costs due to a large volume of repetitive inquiries and the risk sensitivity of policy interpretation scenarios (such as inquiries about attendance rules and payroll calculations, which are prone to disputes due to misunderstandings). While Type 4 processes bear the burden of core asset accumulation and long-term implementation needs, a balance must be struck between process standardization and scientific decision-making. Furthermore, the education industry has stringent requirements for process compliance, data accuracy, and service stability; therefore, AI transformation of all types of processes must achieve a precise balance between efficiency improvement and risk control.

[0021] Based on this, this application provides a business process refactoring method, apparatus, and device based on intelligent agents. First, it obtains the definition file of the target business process and preset business constraint rules. Next, based on the definition file, it constructs a directed acyclic graph (DAG) corresponding to the target business process, including human nodes and intelligent agent nodes. Then, according to the business constraint rules, iterative optimization is performed on the configuration parameters of each node in the DAG to output a set of optimal configuration parameter combinations covering all process nodes. Finally, based on the optimal configuration parameter combinations, the original DAG is refactored to generate a target DAG representing the optimized business process, thereby completing the intelligent transformation and solution output of the business process. This can improve process operation efficiency and risk management capabilities, and help enterprises make scientific decisions on process reengineering.

[0022] The following explanation, in conjunction with the accompanying drawings, uses several embodiments to illustrate the concepts. Figure 1 A flowchart illustrating a business process refactoring method based on intelligent agents provided in this application is shown below. Figure 1 As shown, this agent-based business process refactoring method includes: S101, Obtain the definition file of the target business process and the preset business constraint rules.

[0023] The target business process refers to a standardized combination of business steps formed by an enterprise in scenarios such as production operations, customer service, and internal management. Examples include supply chain procurement processes, customer complaint handling processes, financial expense approval processes, and employee onboarding processes. The definition document for the target business process can be in formats including, but not limited to, BPMN (Business Process Model and Notation) files, Visio files, JSON format process description files, or enterprise-customized process definition specifications. The definition document clearly includes all the steps constituting the process, the execution order and associated conditions between each step, the roles or responsible parties associated with each step, and the key data objects and input / output specifications involved in the process flow.

[0024] The pre-defined business constraint rules are used to standardize the execution of processes and limit the scope of optimization. They mainly revolve around three dimensions: financial constraints, human resource constraints, and SLA (Service Level Agreement) objectives. Specifically, they include: 1) Financial constraints: Clearly defining cost control standards during process execution, including the overall budget ceiling, the computing power cost threshold of intelligent agent nodes, cost structure restrictions, etc., to ensure that process optimization is always carried out within the enterprise's financial budget; 2) Human resource constraints: Based on the enterprise's organizational structure and staffing plan, setting resource allocation restrictions for human nodes, including the maximum number of human agents, the maximum concurrent processing capacity of a single human node, personnel utilization thresholds, etc., to ensure that human resource allocation is accurately matched with process operation needs; 3) SLA objectives: Defining the timeliness and quality standards for process execution, including the upper limit of response time for the entire process, task completion accuracy requirements, and exception handling timeliness, to ensure the stability of process execution and service quality.

[0025] In this embodiment, after obtaining the definition file, the category of the target business process will be identified based on the process metadata of the definition file, and adaptive weights will be configured to provide a decision basis for subsequent parameter optimization.

[0026] Specifically, feature vectors are extracted based on the process metadata in the definition file. ,in, Indicates whether the service recipient is an external customer (0 / 1); Indicates whether direct capital inflow is generated (0 / 1); Characterizes the tolerance for business interruption (high / low); Characterizes the frequency of process calls (high / low).

[0027] Meanwhile, based on the extracted feature vectors, the target business process is mapped to one of four standard types: Type 1 refers to processes that directly face customers and generate revenue (such as student recruitment and sales); Type 2 refers to high-frequency processes with resource scheduling and strong timeliness requirements (such as class scheduling and supply chain); Type 3 refers to internal services and cost-sensitive processes (such as attendance and expense reimbursement); and Type 4 refers to low-frequency but high-value processes that involve the accumulation of core assets (such as the construction of new campuses and data governance).

[0028] Simultaneously, a dynamic five-dimensional value weight matrix is ​​maintained, and the corresponding weight vector is loaded according to the type of the target business process. .

[0029] The matrix definition sets five value dimensions: revenue increase ( ), cost reduction ( ), improve efficiency ( ), quality improvement ( ),assets( For type one processes, load weights. The core strategy is aggressive growth, tolerating higher costs in pursuit of increased revenue and improved user experience; for Type II processes, weighting is applied. The core strategy is to maximize efficiency, focusing on optimizing processing speed and throughput; for type three processes, loading weights... The core strategy prioritizes cost, maximizing cost savings while ensuring compliance and basic efficiency; for Type 4 processes, weighting is applied. The core of the strategy is long-termism, with a focus on the accumulation of knowledge assets and process quality.

[0030] In practical applications, data can be obtained in various ways, such as receiving the definition file and business constraint rules of the target business process provided by the user or upstream system. These will then be stored in a designated database to provide data support for subsequent process modeling and parameter optimization.

[0031] S102, Based on the definition file, construct the directed acyclic graph corresponding to the target business process.

[0032] The process nodes in the directed acyclic graph include: human nodes and intelligent agent nodes.

[0033] In this step, the definition file of the acquired target business process is parsed and modeled, transforming it into a structured graph representation, namely a Directed Acyclic Graph (DAG). In a DAG, vertices correspond to independent business processing steps (i.e., process nodes), and directed edges correspond to the sequential execution relationships and data flow directions between nodes. Logical checks ensure no circular dependencies, guaranteeing the orderly and efficient progress of the process.

[0034] During construction, based on the complete description of the business process in the definition file, the processing links of the business process and the execution order and associated conditions of each link are extracted. Each processing link is mapped as an independent process node, and the execution relationship and associated conditions between links are transformed into directed edges between nodes, forming a directed acyclic graph with a complete structure and clear logic.

[0035] After the graph structure is constructed, a preset heterogeneous node definer is invoked to identify the type and assign attributes to each process node in the graph. Process nodes are explicitly classified as manual nodes. and agent nodes There are two categories. Human nodes correspond to business processes requiring human decision-making, subjective judgment, compliance review, or complex interactions; rule-based software (such as writing to a database); or RPA (Robotic Process Automation) based on fixed rules. The processing of such nodes relies on human experience and intervention. Given the same input and context, the output of these nodes is theoretically deterministic and repeatable; they can also be called deterministic nodes. Agent nodes... Leveraging artificial intelligence technologies (such as large language models, rule engines, and automation tools), standardized and repetitive business operations can be completed automatically without human intervention, producing results. The output of these nodes is generated based on probability distributions trained on massive amounts of data; therefore, even when faced with the same input problem, the intelligent agent nodes... It may also generate different wording of answers, or refuse to answer when the confidence level is insufficient, or even generate hallucinatory content, which can also be called probabilistic nodes.

[0036] For those identified as artificial nodes For each process node, create and populate a specific attribute vector. This vector contains key parameter slots describing manual or traditional software rule behavior, mainly including: hourly wage (representing the human cost of processing the task at this node), concurrency limit (representing the maximum number of tasks that the node can process in parallel at the same time), and processing time distribution (…). (used to simulate the fluctuation of processing time for a single task) and fatigue threshold (used to simulate the critical point of decreased efficiency or increased error rate after long-term work).

[0037] For nodes identified as intelligent agents For each process node, create and populate another attribute vector. This vector contains key parameter slots for the AI ​​agent, mainly including: model type (such as GPT-5, GLM-4, Gemini-3, etc., different models correspond to different capabilities and costs, as well as illusion rate distribution), token unit price (based on the computational cost of the number of input and output tokens), and illusion rate. (Probability that AI-generated content contains factual errors), Refusal rate (The probability that AI will refuse to answer due to insufficient confidence), response latency jitter (Variations in network or computation latency) and confidence threshold (The critical value by which an AI node determines the reliability of its output). During the initialization phase, especially in zero-shot modeling scenarios, reasonable initial values ​​are set for the attribute parameters of the agent node; for example, the hallucination rate of the agent node. It can be initialized to obey a Distribution to simulate its low-probability but potentially hallucinatory nature; Refusal rate It can then be assigned an initial value (such as 10%) as the starting point for the optimization process.

[0038] In the directed acyclic graph of the constructed target business process, each process node is explicitly divided into those carrying... Artificial nodes carrying attributes The attributed agent nodes provide basic model support for subsequent configuration parameter optimization and process refactoring.

[0039] S103. Based on the business constraint rules, iteratively optimize the configuration parameters of the directed acyclic graph to obtain the optimal combination of configuration parameters for the directed acyclic graph.

[0040] The optimal configuration parameter combination includes: the optimal configuration parameters of each artificial node and the optimal configuration parameters of each intelligent agent node.

[0041] Specifically, the optimal configuration parameter combination refers to the set of configuration parameters that, under strict adherence to business constraints, maximizes the overall operational efficiency of the target business process. This includes the optimal configuration parameters for each human node and each agent node. Specifically, the configuration parameter combination is the set of adjustable configuration parameters for all nodes in a directed acyclic graph. For human nodes, the adjustable configuration parameters primarily include the number of human resources allocated. For example, adjusting the number of staff for a specific approval position; for intelligent agent nodes, the adjustable configuration parameters mainly include confidence thresholds. Choose different AI model versions with varying costs and performance for the intelligent agent nodes.

[0042] Using business constraint rules as boundary conditions and maximizing the overall efficiency of the business process as the optimization objective, multiple rounds of iterative optimization are conducted on the configuration parameters of all process nodes in the directed acyclic graph (DAG) to obtain the optimal combination of configuration parameters for the DAG. To achieve this goal, a digital twin simulation environment is first constructed, in which accurate and efficient configuration parameter optimization can be achieved even when facing new business processes without historical operation logs.

[0043] To achieve a realistic simulation environment, a differentiated strategy is employed to construct the simulation request flow based on the data foundation of the target business process. For new business processes lacking historical operation logs (cold start scenario), a preset demand generator is first invoked. Leveraging the role-playing capabilities of a large language model, it automatically generates virtual user request queues that conform to the distribution of real business scenarios based on the business description of the target business process (e.g., "attendance inquiry," "admissions Q&A"). For existing business processes with accumulated historical operation logs, data such as real user request sequences, intent distributions, sentiment tags, and processing times are directly extracted from the historical logs. After anonymization and standardization, a virtual user request queue that highly replicates the historical load pattern is constructed. Through these two methods, each generated virtual request contains rich metadata. Examples of parameters include intent complexity (distinguishing between simple queries and complex appeals), emotional valence (simulating different emotional states such as user anger, calmness, or impatience), and the expected token length to process the request. These parameters are used to simulate the ever-changing user questioning patterns in the real world, covering different intent types and emotional states. This ensures that the simulation test data is no longer a single, repetitive set of data, but rather a comprehensive test of the robustness and adaptability of the process under different pressures and scenarios.

[0044] When a virtual user request flows through a directed acyclic graph in the simulation environment, a pre-defined supply-side actuator is responsible for driving the actual behavior simulation of each agent node in the graph. For the flow through the agent nodes... Each request will be processed based on the probability parameter (rejection rate) in the attribute. Hallucination rate Confidence threshold It executes a set of dynamic routing logic. Specifically, it generates a random number. And based on the comparison with probability parameters, the request routing path is determined, for example: if Refusal rate If the Route_Reject routing rule is triggered, the task will be processed according to the timeout and transferred to the human node queue; if Confidence threshold And hallucinations occur (based on hallucination rate) If the task fails to respond correctly (illusion), the Route_Hallucination routing rule is triggered, indicating an incorrect response (illusion) and the task flows to the downstream node. It is also marked as a potential risk work order, triggering manual error correction in subsequent stages. In other cases, the Route_Success routing rule is triggered, the task responds correctly and the loop is automatically closed, fully restoring the real business flow logic of human-machine collaboration.

[0045] In the specific iterative optimization process, the initial configuration parameter combination is first assigned to the attribute slots of each process node in the directed acyclic graph. Then, a virtual user request queue generated by the demand side is put into the simulation environment. High-concurrency simulation is performed by the supply-side executor, and the full-process quantitative performance data is collected during the simulation. Subsequently, within the parameter range limited by business constraint rules, the node configuration parameters are adjusted to form a new parameter combination. The virtual high-concurrency simulation is performed again, and performance data is collected. Through multiple rounds of parameter adjustment, virtual simulation, and performance comparison, the parameter configuration direction is continuously optimized, and all parameter adjustments do not violate the business constraint rules.

[0046] Each iteration records the configuration parameter combination and the corresponding simulated quantitative performance data. Based on the type of the target business process and the corresponding five-dimensional value weight vector, the comprehensive value score corresponding to the parameter combination in that round is calculated. This score is obtained by mapping the performance data to five dimensions and weighting and summing them according to their weights, thereby quantitatively evaluating the overall performance of the current configuration parameter combination. Based on the score, parameter combinations that do not meet the performance standards are gradually eliminated, and convergence is made to the optimal parameter range until a set of configuration parameter combinations that can balance the performance of multiple dimensions is found. This is the optimal configuration parameter combination of the directed acyclic graph, providing accurate and actual business-aligned parameter basis for subsequent process refactoring.

[0047] S104. Based on the optimal combination of configuration parameters, the directed acyclic graph is reconstructed to obtain the target directed acyclic graph corresponding to the target business process.

[0048] Among them, the target directed acyclic graph is an optimized process model that integrates the optimal combination of configuration parameters and adapts to the actual business operation requirements. It can directly guide the automated execution and management of the target business process, and can transform the results of parameter optimization into structured process logic, thereby maximizing the efficiency of human-machine collaboration.

[0049] In this step, based on the optimal configuration parameter combination obtained in S103, the original directed acyclic graph is comprehensively reconstructed and optimized to ensure that each process node runs according to the optimal configuration parameters, and to optimize the overall process efficiency and resource utilization.

[0050] Specifically, firstly, each parameter in the optimal configuration parameter combination is assigned to the corresponding attribute slot of the node, thus updating the node attributes. For manually configured nodes... Adjust its attribute vector according to the optimal configuration parameters. For example, the number of human resources allocated based on the optimal combination of configuration parameters. Adjustment Medium hourly wage, concurrency limit, processing time distribution, and fatigue threshold; for agent nodes Adjust its attribute vector according to the optimal configuration parameters. For example, a confidence threshold based on the optimal combination of configuration parameters. Adjustments to AI model versions Model type, token price, illusion rate Refusal rate Response latency jitter and confidence threshold .

[0051] Secondly, the topology of the original directed acyclic graph is optimized and adjusted based on the optimal configuration parameters, eliminating redundant links, simplifying unnecessary flow links, and optimizing the dependencies between nodes. For example, if the optimal configuration parameters show a significant increase in the concurrency limit of a certain artificial node, its upstream and downstream nodes with similar functions can be merged to shorten the process; if the illusion rate of the agent node is... Optimization to an extremely low level allows for adjustment of the workflow path between the process and downstream manual review nodes, reducing unnecessary manual intervention. At the same time, based on the optimal processing capacity and concurrency limit of each process node, the task allocation path is optimized to avoid resource waste where some nodes are overloaded while others are idle, ensuring a smoother overall workflow.

[0052] In one possible implementation, based on the updated node attributes and optimized topology, the reconstructed target directed acyclic graph is fully validated to verify its acyclicity, logical coherence, and compliance with business constraints. This ensures that the reconstructed process model not only meets the requirements of the optimal configuration parameter combination but also adapts to actual business scenarios, ultimately forming a target directed acyclic graph that can be directly implemented and executed.

[0053] In this embodiment, by distinguishing between human nodes and intelligent agent nodes to construct a directed acyclic graph, the characteristics of the human-machine collaborative process can be clearly depicted, effectively overcoming the drawbacks of ambiguous node attributes and chaotic collaboration logic in traditional modeling. At the same time, iterative optimization of configuration parameters based on business constraint rules can achieve refined and intelligent adjustment of process parameters, ensuring that the parameter combination is highly adapted to actual business needs. Finally, the directed acyclic graph is reconstructed based on the optimal configuration parameters, which can optimize the node collaboration logic and topology, improve process operation efficiency and risk management capabilities, and help enterprises make scientific process reengineering decisions.

[0054] In the above Figure 1Based on the corresponding embodiments, in order to more clearly demonstrate the process of obtaining a directed acyclic graph, this application also provides a possible implementation of obtaining a directed acyclic graph in a business process refactoring method based on intelligent agents. Figure 2 This is a schematic diagram of the directed acyclic graph obtained in the agent-based business process refactoring method provided in this application. For example... Figure 2 As shown, in S102 above, according to the definition file, a directed acyclic graph corresponding to the target business process is constructed, including: S210, parse the definition file to obtain each process node and the dependencies between nodes.

[0055] Specifically, the parsing process first adapts the definition file to its format, using a dedicated parser to read the structured data and unstructured descriptive information. For the structured data, attributes such as unique identifiers, names, functional descriptions, responsible persons, and associated business forms of process nodes are directly extracted. For unstructured descriptions (such as process description documents and flow rule notes), semantic analysis is performed using Natural Language Processing (NLP) technology to extract hidden process steps and constraints.

[0056] In the node extraction stage, all business processing steps in the definition file are traversed, and each independent operation unit is mapped to a process node. For example, in the definition file of the financial reimbursement process, each independent operation such as "submitting reimbursement application", "department manager approval", "financial invoice verification", and "financial payment" will be identified as a separate process node, and the attributes of the process node will be recorded. For example, the person in charge of the "financial invoice verification" node is the finance specialist, and the associated forms are the scanned invoice and the expense list.

[0057] Furthermore, the execution order and triggering conditions of each process node are analyzed to identify the dependencies between nodes. By analyzing the pre-requirement configurations and flow rule descriptions of nodes in the definition file, the input dependencies (i.e., the processing results that need to be completed by the upstream node) and output targets (i.e., the downstream nodes to which the processing flows after completion) of each node are clarified. For example, the parsing reveals that the pre-requirement of the "Financial Invoice Verification" node is the "Approval Passed" result of the "Department Manager Approval" node, and the downstream target is either the "Financial Payment" node (when verification is passed) or the "Rejection and Modification" node (when verification is not passed). For scenarios with parallel flow or conditional branches, the branch triggering conditions are identified (e.g., "Reimbursement amount ≤ 5000 yuan" triggers automatic payment, "Reimbursement amount > 5000 yuan" triggers secondary approval by the finance manager) and recorded as conditional dependencies between nodes.

[0058] After parsing, a standardized list of nodes and a dependency table will be output. The node list contains information such as the unique identifier, name, and attribute description of each process node; the dependency table, in the form of a matrix or linked list, clearly defines the sequence, triggering conditions, and data flow direction between process nodes.

[0059] S220: Construct an initial directed acyclic graph based on the process nodes and the dependencies between them.

[0060] First, the vertex mapping of the graph structure is completed based on the node list. Each process node is treated as an independent vertex in a directed acyclic graph. The core information such as the unique identifier, name, and functional description of the associated process node are included in the vertex attributes to ensure that there is a one-to-one correspondence between the vertex and the process node.

[0061] Subsequently, directed edges between vertices are constructed based on the dependency table. For nodes with a clear order, directed edges are drawn from upstream nodes to downstream nodes, and the flow triggering conditions are recorded in the edge attributes. For parallel flow scenarios, independent directed edges are drawn for each parallel branch, and the triggering rules corresponding to the branches are marked. For conditional branch scenarios, the branch judgment conditions are clearly marked on the directed edges to ensure that the flow logic is consistent with the definition file.

[0062] During the construction of the graph structure, acyclicity checks are performed simultaneously to ensure that the constructed graph has no circular dependencies, guaranteeing that the process can proceed in sequence. After the checks pass, an initial directed acyclic graph is formed, with vertices representing process nodes and directed edges representing the dependencies and flow conditions between nodes, presenting the overall architecture of the target business process.

[0063] S230: Based on the metadata tags of each process node in the initial directed acyclic graph, perform type identification on each process node in the initial directed acyclic graph and determine the operation category tag of each process node.

[0064] The operation category label is used to indicate whether a process node is a human node or an intelligent agent node.

[0065] After constructing the initial directed acyclic graph, it is necessary to identify the executing entity of each process node in the graph to distinguish whether it requires human intervention or can be automatically executed by an AI agent.

[0066] First, the metadata tags of each process node in the initial directed acyclic graph are read, and rapid classification is achieved based on the explicit identification of the metadata tags. For example, for definition files in formats such as BPMN 2.0 XML and Visio XML, the extended attribute fields corresponding to the process nodes are parsed. If a field is marked with type="Human" or type="Rule_System", the node is directly determined to be a human node. Assigning artificial nodes Operation category label; if the label is type="AI_Agent" or type="LLM", it is determined to be an agent node. Assign nodes to intelligent agents Operation category label.

[0067] If the metadata of a process node does not contain an explicit type label, it is supplemented by swimlane role association identification. Specifically, the process definition file is read. <laneset>The information is extracted to determine the swimlane name and associated role for each process node, establishing a "role-node type" mapping relationship. If a node is located in a swimlane corresponding to a human role such as "Attendance Specialist" or "Finance Manager," then the associated role is the specific human position, and it is determined to be a human node. If a node is located in an AI-related swimlane such as "AI Assistant," "Intelligent Customer Service," or "GPT-4 Processing," then its associated role is an AI agent, and it is determined to be an agent node. This ensures that type identification can still be completed based on business scenario association even without explicit labels.

[0068] For nodes that lack both explicit metadata tags and clear swimlane role associations, semantic analysis-based intelligent identification will be initiated. By invoking the NLP semantic analysis unit, keyword extraction and semantic parsing are performed on the node name and functional description. If keywords such as "generate," "intelligent reply," and "automatic identification" are detected, it indicates that the node possesses AI intelligent processing characteristics, and it will be automatically recommended and labeled as an intelligent agent node. If keywords such as "review," "signature," or "stamp" are detected, it indicates that the node requires manual intervention and will be automatically recommended and marked as a manual node. To ensure accuracy, a configuration confirmation interface will pop up: "The 'Life Weekly Report' node may be an AI node. Confirm?" This will present the recommended node types to the user, allowing the user to ultimately confirm the node type and avoid recognition errors caused by semantic ambiguity.

[0069] S240, add the operation category labels of each process node to the initial directed acyclic graph to obtain the directed acyclic graph corresponding to the target business process.

[0070] In this step, firstly, all process nodes in the initial directed acyclic graph are traversed, and the operation category label (human node label / agent node label) corresponding to each process node is written into the node's attributes, and stored in association with the original attributes such as the process node's unique identifier, name, and functional description. Secondly, for the labeled node type, corresponding attribute vector slots are automatically reserved for it; for example, for human nodes... ,create Attribute slots; for agent nodes Reserved Attribute slots.

[0071] Once the verification is successful, a directed acyclic graph corresponding to the target business process is obtained. This not only fully preserves the topology and flow logic of the initial directed acyclic graph, but also clarifies the type attributes, characteristics, and attribute slots of each node through operation category labels. In this embodiment, the transformation from definition files to structured directed acyclic graphs is realized. By clearly distinguishing between human nodes and intelligent agent nodes, the constructed graph structure presents process logic and node characteristics, which greatly improves the intelligence and accuracy of business process refactoring.

[0072] In the above Figure 1 Based on the corresponding embodiments, in order to more clearly demonstrate the process of obtaining the optimal configuration parameter combination, this application also provides a possible implementation of obtaining the optimal configuration parameter combination in the business process refactoring method based on intelligent agents. Figure 3 This is a schematic diagram illustrating the process of obtaining the optimal combination of configuration parameters in a business process refactoring method based on intelligent agents provided in this application. Figure 3 As shown, in S103 above, based on business constraint rules, the configuration parameters of the directed acyclic graph are iteratively optimized to obtain the optimal combination of configuration parameters for the directed acyclic graph, which includes: S310 uses the preset policy network corresponding to the business constraint rules to select a set of configuration parameter combinations from the corresponding action space.

[0073] In this step, the selection of configuration parameter combinations is completed based on a preset policy network that adapts to business constraint rules. The policy network defines a state space. This indicates that the intelligent agent is in A snapshot of the environment at any given moment, used as a feature vector to perceive the health of the process operation, specifically... These represent the current queue length, end-to-end average waiting time, cumulative token consumption, AI illusion / abnormal trigger frequency, and human agent occupancy rate for each node, respectively; and also define an action space. This represents the set of all configuration parameters that can be dynamically adjusted (i.e., optimizable variables), mainly including: firstly, AI cognitive parameters, primarily the confidence threshold for each agent node. It is a The first is the continuous variable within the interval, which directly defines the boundary between automatic closed-loop processing and manual review of the AI ​​task; the second is the AI ​​model selection parameter, which is a discrete variable representing the selection of each agent node from a predefined pool of candidate models, for example... The first step involves selecting a specific version of a large language model, with different models corresponding to varying cost rates, processing capabilities, and illusion rates. The second step involves the allocation of human resources, which is also a discrete variable, involving adjusting the number of human nodes. For example, it allows you to decide whether a customer service or approval position should be staffed with 3 or 5 people, and supports adjustments as needed.

[0074] In completing the state space With action space After the policy network is clearly defined, it is activated and connected to the simulation environment. At the beginning of each iteration, the policy network receives the current state vector from the environment feedback. Based on its internally parameterized strategy function Choose the action space given the state. Each possible action (i.e.) Model version and The combination of these parameters can be used as the initial configuration parameter combination. For example, the policy network might decide to set the confidence threshold of agent node A. The value is set to 0.85, and the GLM-4 model is selected for it. The number of artificial nodes B is also adjusted. Adjusted to 3.

[0075] S320 adjusts the configuration parameters of each process node in the directed acyclic graph according to the configuration parameter combination, and obtains the simulation quantization performance data of the directed acyclic graph after the configuration parameter adjustment.

[0076] First, update the attribute parameters of the corresponding nodes in the directed acyclic graph one by one according to the configuration parameter combinations. For agent nodes... , set confidence threshold Assigning to it Define the criteria for determining whether AI-automated closed-loop processing should be implemented or manually reviewed; based on the selected model version, synchronously update the node's token price, illusion rate distribution, and other related attributes; for manual nodes... According to the configured number of staff Adjustment The parameters such as hourly wage and concurrency limit in the attribute slots ensure that the node attributes match the selected parameter combinations.

[0077] After parameter adjustments, a complete business process simulation is initiated using this newly configured directed acyclic graph in the constructed digital twin simulation environment. Virtual user request queues generated on the demand side are used to execute high-concurrency process simulations. During the simulation, the supply-side executor processes requests according to the dynamic routing logic of the agent nodes, recreating the human-machine collaborative business flow scenario. Simultaneously, real-time simulation-quantitative performance data, including data dimensions and state space, is collected throughout the entire process. Correspondingly, this includes the current queue length of each node. End-to-end average waiting time Cumulative Token Consumption AI hallucination / abnormal trigger frequency Human agent occupancy rate All data is recorded in a structured manner, forming a simulation quantization performance dataset corresponding to the configuration parameter combination.

[0078] The S330 uses a preset reward function to calculate the reward value corresponding to the combination of configuration parameters based on the simulated quantification performance data.

[0079] Specifically, based on the preset reward function, the simulation quantification performance data obtained by S320 is analyzed and normalized in multiple dimensions to obtain the reward value corresponding to the current configuration parameter combination, providing a basis for subsequent policy network updates. The reward function adopts a normalized scalar scoring mechanism, covering five dimensions of value: increased revenue, reduced costs, improved efficiency, improved quality, and assets. The calculation formula is shown in formula (1):

[0080] Formula (1) Specifically, it can be elaborated as follows:

[0081] in, The total reward value is the final target score during parameter optimization. The higher the score, the better the overall performance of the current configuration parameter combination. For adaptive weights, these are the weight coefficients for each dimension, including... , , , , Configure according to the type differences of the target business process; Normalization, or standardization function, is used to normalize data of different dimensions. Mapped to Intervals, eliminating the influence of dimensions; By taking the reciprocal of the cost term, the cost minimization problem is transformed into a reward maximization problem.

[0082] During the calculation, based on the simulated quantification performance data, a lightweight procedural proxy algorithm is first used to estimate the original index values ​​of the five dimensions of value. This includes: 1) Revenue-increasing guidance items Accumulate expected output for high-value tasks successfully completed within the simulation cycle, according to Calculation, where The function is a closed-loop indicator (0 / 1) for successful task completion. This is the preset output benchmark value for the corresponding position of this task. This is the capacity conversion coefficient. This represents the total number of tasks processed in this simulation, traversed... From 1 to All subsequent parameters are synonymous; 2) Cost reduction-oriented items This is the weighted sum of computing power and labor costs within the simulation period, calculated as follows: ,in For the cumulative number of tokens, The average price per token. For manual busy time, 3) Efficiency-oriented items (average hourly wage for workers) , which is the reciprocal of the average dwell time of all tasks, calculated by calculate, Waiting time for tasks to be queued. A small constant introduced to prevent the denominator from being zero; the larger the value, the higher the process efficiency; 4) Quality improvement-oriented items , which is the net success rate of no serious errors occurring during the simulation period, calculated as follows: calculate, Instructions to trigger AI hallucinations For compliance violation indications, the closer the value is to 1, the higher the quality of process processing; 5) Asset-oriented items , representing the number of structured data logs successfully generated during the simulation period, categorized by... , This is a count of unstructured dialogues being transformed into structured assets (i.e., AI not only answers questions but also condenses data according to a schema).

[0083] In calculating the original indicators of five-dimensional value Throughout the process, the collected simulation-quantified performance data serves as the input. For example, the end-to-end average waiting time is directly used for efficiency-oriented items. Average dwell time calculation; cumulative token consumption as a cost reduction guideline. The cost of computing power is calculated by combining the human agent occupancy rate with the corresponding human busy time to determine the overall cost; the frequency of AI hallucinations / abnormal triggers is used to calculate quality improvement-oriented items. .

[0084] Complete the original indicator values ​​of the five dimensions After estimation, each indicator is normalized, and then combined with an adaptive weight matrix matching the current business type, the weighted calculation is completed by substituting it into the reward function, resulting in the final value. This is the reward value corresponding to the current combination of configuration parameters. Also, when calculating the reward value... During the process, if the confidence threshold of the agent node Setting the confidence level too high will cause a large number of tasks that could be automatically closed by intelligent agents to be transferred to human nodes because they cannot meet the confidence requirements, directly resulting in a high occupancy rate of human agents. High-end average waiting time The increase ultimately leads to a higher total reward value. Decrease; conversely, if the confidence threshold of the agent node increases. If the settings are too low, the agent will automatically process tasks too frequently due to insufficient knowledge, directly resulting in a higher frequency of AI hallucinations / anomalies. A significant increase, ultimately leading to a decrease in total reward value Reduced. Through this design, the reward function... It can automatically identify efficiency bottlenecks caused by parameter configuration imbalances and transform this identification into reward signals, thereby guiding the policy network to find a balance between avoiding human congestion and controlling the AI ​​error rate in subsequent iterations.

[0085] S340 updates the policy network using a near-end policy optimization algorithm based on the reward value.

[0086] In this step, based on the reward value calculated by S330, the parameters of the policy network are updated through the Proximal Policy Optimization (PPO) algorithm, enabling the policy network to learn and output better configuration parameter combinations, gradually approaching the global optimum. At the same time, Monte Carlo Tree Search (MCTS) is supported as an alternative algorithm to adapt to business process optimization scenarios with different complexities.

[0087] First, the reward value corresponding to the current configuration parameter combination obtained from S330 is compared with the reward value data recorded during historical iterations to clarify the performance of the current strategy. Then, the iterative logic of the proximal policy optimization algorithm is adopted. This algorithm optimizes the parameters of the policy network through gradient ascent, with maximizing cumulative reward as its core objective. Specifically, the algorithm first calculates the gradient between the probability distribution of action selection under the current policy and the reward value, and then fine-tunes the mapping relationship parameters between the state space and action space in the policy network along the direction of gradient ascent, avoiding training instability caused by excessively large parameter update amplitudes.

[0088] During parameter updates, the policy network incorporates actual feedback from simulations to adjust the output logic of various parameters within the action space. For example, when the reward value displays the current confidence threshold... When an excessively high confidence threshold leads to an increase in the manual review rate and costs, the algorithm will automatically adjust the output parameter of the confidence threshold to a better value (such as 0.78). When the illusion rate corresponding to the model selection is too high and affects the quality improvement index, the algorithm will tend to output the parameters of the model version with a lower illusion rate.

[0089] For business processes with multiple branches and high complexity, the Monte Carlo tree search algorithm can be switched to. By efficiently searching and evaluating the possible space of parameter combinations, parameter directions with potential high reward values ​​can be selected, and then the policy network can be updated by combining the gradient ascent method.

[0090] After the parameters are updated, the new policy network will generate a set of optimized configuration parameters for the next round of simulation and reward calculation. Through multiple iterations, the policy network will continue to evolve, and the output parameter combinations will become increasingly aligned with the five-dimensional value orientation of the target business process.

[0091] S350 uses the updated policy network to select a new combination of configuration parameters from the action space and obtains the new reward value corresponding to the new combination of configuration parameters until the new reward value meets the preset stopping iteration condition.

[0092] First, the updated policy network of S340 is invoked, and combined with real-time feedback from the current state space (including core indicators such as queue length and average waiting time of each node), a new set of configuration parameter combinations is re-selected from the action space and output (i.e., Model version and The new configuration parameter combination is then substituted into the S320 process to simultaneously complete the adjustment of the directed acyclic graph node parameters, the high-concurrency simulation of the digital twin environment, and the full-process quantization performance data acquisition. Then, through the reward function of S330, a new reward value is calculated based on the acquired simulation quantization performance data, forming an iterative closed loop of "parameter selection → simulation simulation → reward calculation".

[0093] During the iteration process, the changing trend of the new reward value in each round will be monitored in real time, and convergence will be determined according to the preset stopping iteration conditions: on the one hand, if the fluctuation range of the reward value is less than the preset threshold for several consecutive rounds (such as 500 rounds), the convergence will be determined. If the reward function stabilizes, it is considered that the iteration has converged. On the other hand, if the number of iterations reaches a preset limit (such as 10,000 times), the iteration process is terminated regardless of whether the reward value is completely stable, in order to avoid resource consumption caused by overtraining.

[0094] The optimization process is considered to have converged when any of the above stopping iteration conditions are met.

[0095] S360: Obtain the configuration parameter combination that yields the highest reward value when the preset stopping iteration condition is met as the optimal configuration parameter combination.

[0096] After the S350 iteration is terminated, key data recorded throughout the entire iteration process will be extracted, including the configuration parameter combinations, simulation quantization performance data and reward values ​​corresponding to each iteration, to build a complete iteration data ledger.

[0097] Subsequently, all configuration parameter combinations in the ledger will be screened: among all configuration parameter combinations, the one with the highest reward value will be selected as the optimal configuration parameter combination. If multiple parameter combinations have the same reward value, the performance of each configuration parameter combination in the core dimensions will be compared by combining the five-dimensional value weighting guidance and business constraint rules of S101 (e.g., Type 1 processes focus on revenue increase and quality improvement indicators, while Type 3 processes focus on cost reduction and efficiency improvement indicators) to finally determine the unique optimal configuration parameter combination.

[0098] After the selection is complete, detailed information on the optimal configuration parameter combination will be output, specifying the number of manual nodes to be created. Confidence threshold for each agent node The AI ​​model version is synchronized with its corresponding simulation quantification performance data and reward value details to ensure the traceability of output results and provide accurate parameter basis for subsequent process reconstruction of directed acyclic graphs.

[0099] In this embodiment, iterative optimization using a policy network enables automated and intelligent optimization of business process configuration parameters. It can simulate high-concurrency business scenarios in a virtual environment and adaptively adjust the combination of configuration parameters. While satisfying multiple constraints, it maximizes the overall efficiency of the process, effectively solving the problems of low efficiency and difficulty in balancing multiple objectives in traditional manual parameter tuning. It can quickly output the optimal parameter combination verified by massive simulations, significantly improving the adaptability of human-machine collaboration.

[0100] In the above Figure 1 Based on the corresponding embodiments, in order to more clearly demonstrate the process of obtaining the target directed acyclic graph, this application also provides a possible implementation of obtaining the target directed acyclic graph in the agent-based business process refactoring method. Figure 4 This is a schematic diagram illustrating the process of obtaining the target directed acyclic graph in the agent-based business process refactoring method provided in this application. For example... Figure 4 As shown, in S104 above, the directed acyclic graph is reconstructed based on the optimal configuration parameter combination to obtain the target directed acyclic graph corresponding to the target business process, which includes: S410: Based on the optimal combination of configuration parameters, optimize the topology of the directed acyclic graph to obtain the reconstructed topology.

[0101] This step, based on the optimal configuration parameter combination, combines simulated performance data to optimize the original directed acyclic graph topology through branch solidification and route insertion, constructing a topology adapted to the optimal configuration parameters. Here, simulated performance data refers to the real-time, structured, and recorded end-to-end operational data collected during the iterative optimization of the configuration parameter combination, when a high-concurrency simulation is performed in a digital twin simulation environment configured with the optimal parameter combination. This data includes key indicators such as queue length at each node, end-to-end average waiting time, cumulative token consumption, AI illusion / anomaly trigger frequency, and human agent occupancy rate. It also includes detailed confidence levels of agent nodes handling various tasks and records of risk events in different task scenarios.

[0102] Specifically, optimization includes branch solidification and route insertion. Branch solidification is a pruning process based on comparing simulated quantitative performance data with a preset high confidence threshold. The decision parameters are based on the AI ​​confidence statistics generated during the simulation. These results extract the confidence details of the agent node's task processing under the optimal configuration parameter combination from the simulation statistics. The preset decision threshold can be set to 0.99. For a specific type of standardized task (such as querying annual leave or inquiring about attendance rules), multiple high-concurrency simulations are performed in the digital twin simulation environment defined by the optimal configuration parameter combination. The confidence data of the agent node in processing this type of task is recorded. If the statistical results show that the confidence output by the agent node when processing this type of task is consistently higher than the preset threshold (0.99), it is determined that the error rate of this step is extremely low and no manual intervention is required. Then, the original manually reviewed nodes connected to the downstream of the agent node in the acyclic graph are deleted, and the process path of this type of task is modified to a fully automatic path, thereby improving processing efficiency.

[0103] Route insertion is a hardening process based on comparing risk penalty values ​​with safety red lines. The decision parameters are the risk penalty values ​​from the simulation results. Based on this, the variables used to calculate the risk penalty value are all derived from simulation statistics driven by the optimal configuration parameter combination. The calculation is shown in formula (2): Formula (2) in, This represents the number of work orders that experienced AI hallucinations during the simulation and were not intercepted. The average time required to manually fix this error; Expert hourly rate refers to the unit-time salary standard for human roles that possess professional qualifications in the business field and are responsible for handling high-risk / high-complexity error work orders, adapting to the expert human resource cost accounting needs in different business scenarios; Business impact penalty coefficient ( ,like This indicates that additional communication costs are required. The weighting coefficients, set based on user emotion data from simulation statistics, are 1.0 when users are calm and 3.0 when they are negative / angry. This simulates escalating complaints due to user anger, requiring three times the manpower for public relations or appeasement, thus making cost calculations more realistic. When performing simulations of certain highly complex tasks (such as sick leave appeals and expense overrun approvals), the simulation quantifies the number of work orders that exhibit AI hallucinations and are not intercepted, monitoring the number of such work orders from the quantified performance data. Or the corresponding business impact penalty coefficient The value is relatively large, and the risk penalty value is calculated based on data. If the preset safety threshold is exceeded, it is determined that relying solely on AI in this step poses a high risk. Subsequently, a manual review node is forcibly inserted after the intelligent agent node and before the end node, and a connection line is automatically generated. The flow rule for this type of task is set to require manual confirmation to complete the entire closed loop.

[0104] S420 defines cooperative attributes for the node relationships between agent nodes and artificial nodes in the reconstructed topology based on the optimal combination of configuration parameters, and obtains the target directed acyclic graph.

[0105] Among them, the collaborative attribute of node relationships is used to indicate the collaborative mode of human-led intelligent agent assistance, the automatic mode of intelligent agent-led human supervision, or the shadow mode of intelligent agent and human running in parallel.

[0106] In this step, based on the optimal combination of configuration parameters, the collaborative attributes are defined by the interaction relationship between the intelligent agent node and the human node in the reconstructed topology, the human-machine collaborative operation mode is clarified, and finally a target directed acyclic graph that can be implemented is formed.

[0107] Specifically, the collaborative mode of human-led intelligent agent assistance is also known as the Copilot driver mode. This mode typically corresponds to complex tasks or high-risk scenarios. In this mode, the agent's main behavior is to generate a task processing draft, then transfer the task to a human node for final modification and confirmation. The agent acts as an assistant, while the human is the decision-maker. The trigger condition for this mode is when the confidence threshold in the optimal configuration parameter combination is reached. In the high range (e.g.) This indicates that the agent sets an extremely high automatic pass threshold to avoid risks, causing most tasks to be diverted to the manual queue. According to the routing logic of the supply-side executor, random numbers... The probability of success will decrease, and most tasks will not meet the conditions for automatic approval, thus triggering Route_Reject or being transferred to the manual queue. In terms of reconstruction actions, the collaborative relationship between human and machine nodes will be defined as a mode of AI generating drafts and then human confirmation.

[0108] The automated mode where an intelligent agent is in charge and human supervision is also known as the Autopilot autonomous driving mode, corresponding to the confidence threshold determined in the optimal combination of configuration parameters. In the low to medium range (e.g.) In this scenario, the intelligent agent demonstrates that it can reliably handle most tasks with a high degree of confidence and the risks are controllable. In this mode, the intelligent agent is capable of handling most tasks. The agent can directly execute and send task processing results, with the workflow logic set to only respond to exceptions or when the task processing confidence level falls below a confidence threshold. Only when necessary is the task transferred to a human node. The core function of the human node is to handle abnormal tasks as a fallback. The intelligent agent is positioned as the main decision-maker, while the human node serves as a safety net. The trigger condition for this mode is, on the one hand, a confidence threshold. Within the low to medium range, according to the routing logic, a random number is generated. The probability is relatively high, and most tasks will trigger the Route_Success automatic loop closure; on the other hand, the risk penalty value calculated under this threshold... When it is in a low position, the corresponding reconstruction action will retain the automatic closed-loop path, and only the manual node will be used as a supplementary node for exception handling.

[0109] The shadow mode, where intelligent agents and human nodes operate in parallel, is also known as Shadow Mode. It's typically used during the cold start phase or the validation period for extremely high-risk tasks. In this mode, the intelligent agent runs the task processing flow in the background but doesn't send results outwards, while the human node actually executes the tasks and compares the processing results of the intelligent agent and the human node in the background, using this information for model training or policy validation. The two operate independently, allowing for data accumulation and validation of the optimal policy. This mode is triggered in two ways: the first is based on the risk penalty value. Extremely high, regardless of how the confidence threshold is adjusted. Simulation calculation The persistently high level is usually due to the business impact penalty coefficient of the process. Set it to be very large, for example In one scenario, the model itself has a high illusion rate. In the corresponding reconstruction action, it will trigger the routing insertion logic, forcibly insert a human review node, and even cut off the agent's write permissions, allowing it to run in the background and record logs, and perform parallel verification in shadow mode. In another scenario, the business is in the cold start stage. At this time, the source of the virtual request queue is generated virtual data rather than real historical logs. The simulation results only represent theoretical feasibility and have not yet been verified by real business. In this case, the collaboration attribute of the agent node and the human node will be forcibly defined as shadow mode, and data accumulation and strategy iteration will be completed through parallel operation.

[0110] After all nodes have defined their collaboration attributes, the triggering conditions and exception handling mechanisms for task flow will be improved in parallel to ensure that each collaboration mode can match the optimal configuration parameter combination. The final result is a directed acyclic graph that integrates the optimal configuration parameter combination, reconstructs the topology and collaboration modes.

[0111] In this embodiment, business process reengineering and human-machine collaboration are realized. On the one hand, redundant manual nodes are eliminated by branch solidification, shortening the workflow of high-confidence tasks; on the other hand, targeted supplementation of high-risk link control nodes is achieved by routing insertion, reducing business losses caused by AI illusions. Based on the optimal configuration parameter combination, three human-machine collaboration modes are matched. The autonomous driving mode can release AI efficiency, the driver mode can ensure the decision-making quality of high-risk tasks, and the shadow mode can complete cold start verification and data accumulation, balancing efficiency and safety, maximizing the efficiency of human-machine collaboration, and providing reliable support for the implementation of the target directed acyclic graph.

[0112] In the above Figure 1 Based on the corresponding embodiments, in order to more clearly demonstrate the process of generating a value assessment report, this application also provides a possible implementation of generating a value assessment report in a business process refactoring method based on intelligent agents. Figure 5 This is a flowchart illustrating the generation of a value assessment report in a business process refactoring method based on intelligent agents provided in this application. Figure 5 As shown, based on the above S101-S104, the method further includes: S510 uses a purely manual baseline model to perform purely manual operation simulations on each process node in the target directed acyclic graph, obtaining the manual index data corresponding to the target directed acyclic graph.

[0113] After generating the target directed acyclic graph, a baseline simulation is first initiated to quantitatively evaluate the true value of the optimization compared to the traditional model. Specifically, a purely artificial baseline model is constructed and run based on the same target directed acyclic graph. Benchmark comparison data was obtained through fully manual simulation and deduction, providing a reference for subsequent value assessment. In the purely manual baseline model... In this model, all process nodes (including the original agent nodes) are treated as being operated manually. Their attribute parameters (such as processing time, concurrency limit, and hourly wage) are configured according to a preset pure manual processing mode. Then, in the digital twin simulation environment, the pure manual operation simulation is performed on each process node in the target directed acyclic graph, and a virtual user request queue is added to the pure manual baseline model. The simulation was conducted at high speeds (e.g., 1000x speed) to simulate business operations over a future period (e.g., one year). During the simulation, human performance data was collected and recorded in real time, including but not limited to: total man-hours consumed by each human node and corresponding human costs, end-to-end processing time distribution, task completion rate, number of human errors, and queue backlog at each node. These data represent the operational costs, efficiency ceiling, and quality performance of the business process without the introduction of intelligent agents and relying entirely on human processing, providing a benchmark for subsequent value comparisons.

[0114] S520 employs a collaborative optimization model to perform collaborative simulations of manual and intelligent agent operations on each process node in the target directed acyclic graph, thereby obtaining collaborative index data corresponding to the target directed acyclic graph.

[0115] In another thread that executes in parallel, a collaborative optimization model is employed. Simulations are performed on a directed acyclic graph (DAG) with the same objective. The model strictly employs the optimal configuration parameter combination obtained through optimization. Each node in the graph performs a full-process human-machine collaborative operation simulation based on its type (human node or intelligent agent node). In a digital twin environment, a virtual user request queue is added to the collaborative optimization model to simulate the collaborative operation of human and intelligent agent operations at each process node in the target DAG. The simulation process also employs 1000x time-space acceleration technology, compared with a purely artificial baseline model. Within the same business cycle in the simulation, the agent nodes are based on their confidence threshold. The selected AI model and its probability parameters (hallucination rate, refusal rate) are used for dynamic routing; human nodes are configured based on the number of human resources. Work. During the simulation, collaborative indicator data is collected in real time, including but not limited to: token consumption, human-machine collaboration time consumption, overall operating cost, end-to-end processing time distribution, AI illusion / abnormal trigger frequency, human agent occupancy rate, and other dimensions, which represent the optimal expected operating state that the optimized intelligent agent and human collaboration process can achieve.

[0116] S530 generates a value assessment report corresponding to the directed acyclic graph of the target based on manual indicator data and collaborative indicator data.

[0117] In this step, a value assessment report of the target directed acyclic graph is generated through multi-dimensional comparative analysis of manual and collaborative indicator data, completing the final value verification of the optimal configuration parameter combination. First, the two sets of indicator datasets are dimensionally aligned and normalized to eliminate differences in units of measurement and ensure data comparability. At the same time, relevant data on asset accumulation generated throughout the simulation cycle are integrated to form a standardized and structured comparative analysis dataset.

[0118] Based on this, according to the business process categories identified by S101, the matching weight vectors are automatically loaded, and parallel calculations are performed on five dimensions: revenue increase, cost reduction, efficiency improvement, quality improvement, and asset accumulation. Quantitative analysis is carried out from core dimensions such as business revenue growth, full-cycle cost control, process operation efficiency, task processing quality, and data asset accumulation, to comprehensively measure the value increment of the human-machine collaborative process compared to the pure manual baseline process under the optimal configuration parameter combination.

[0119] After completing multi-dimensional quantitative analysis, the built-in rule engine generates targeted intelligent decision-making suggestions based on the analysis results of each dimension. For example, at the parameter tuning level, based on the correlation data between confidence threshold and manual review cost, specific parameter adjustment suggestions are output, such as suggesting that the confidence threshold be lowered from 0.8 to 0.75, which is expected to reduce manual review cost by 15% within a controllable risk range. At the resource allocation level, resource replenishment suggestions are output for node backlog issues detected during simulation. For example, if a serious task backlog is detected at the "financial payment" node, it is recommended to add one RPA robot or human staff member at this node to improve workflow efficiency. At the project approval level, a clear conclusion is given based on the analysis results of each dimension and investment return cycle data. For example, if the current project's comprehensive value score reaches 85 points and the ROI (return on investment) payback period is less than 12 months, it meets the group's Class A project approval standards and is recommended to be promoted for implementation.

[0120] Meanwhile, the report deeply integrates the results of financial dimension analysis, incorporating key financial indicators such as ROI (reflecting capital efficiency), NPV (Net Present Value) (judging the absolute profitability of a project), and the break-even point (clearly defining the critical point between risk and return), forming a financially-grade assessment conclusion with quantitative support. Specifically, ROI is calculated primarily based on labor costs and abnormal loss costs under a pure manual baseline model simulation, as well as labor costs, token consumption, and risk penalties under a collaborative optimization model, combined with incremental business revenue. NPV is calculated primarily based on the net revenue difference between the collaborative optimization model and the pure manual baseline model in each period, derived using a preset discount rate. The break-even point is calculated primarily based on total implementation costs and net revenue per task (the difference between unit task revenue and cost under the collaborative optimization model, compared to the baseline model), estimating the total number of tasks or the corresponding time period required to achieve cost coverage. These indicators collectively transform the technical efficiency of process optimization into intuitive economic benefits, providing clear and reliable quantitative basis for enterprise decision-making.

[0121] Ultimately, all quantitative analysis results, comparative charts, core conclusions, and specific recommendations are automatically integrated and packaged to generate a comprehensive value assessment report tailored to different decision-making levels. This report not only clearly demonstrates the full value of intelligent transformation of business processes but also provides a complete action guide, offering decision-making basis and execution details for enterprises to make investment decisions and promote implementation.

[0122] In this embodiment, by constructing and running a parallel comparative simulation of a purely manual baseline model and a collaborative optimization model, the objective quantification of the value of business process transformation is achieved. This enables the simulation of long-term, high-concurrency business operations in the future, comprehensively collects multi-dimensional comparative data, and automatically generates an in-depth value assessment report covering key financial indicators and intelligent implementation suggestions. This changes the traditional decision-making model that relies on manual estimation and lacks data support, allowing enterprises to accurately predict the overall benefits, investment return cycle, and potential risks of transformation before project investment. It provides a scientific and reliable data basis for investment decisions in digital transformation, significantly improving decision quality and resource allocation efficiency.

[0123] In the above Figure 5 Based on the corresponding embodiments, in order to more clearly demonstrate the process of generating a value assessment report, this application also provides a possible implementation of generating a value assessment report in a business process refactoring method based on intelligent agents. Figure 6 This is a flowchart illustrating the generation of a value assessment report in another agent-based business process refactoring method provided in this application. Figure 6 As shown, in S530 above, based on the manual indicator data and the collaborative indicator data, a value assessment report corresponding to the target directed acyclic graph is generated, including: S610 performs differential comparison calculations based on manual indicator data and collaborative indicator data to obtain quantitative performance comparison data.

[0124] Specifically, based on a purely artificial baseline model Collected manual indicator data and data based on collaborative optimization models The collected collaborative indicator data includes paired indicator data, including: 1) Resource consumption pairs: including Total man-hours and Token consumption (exist The following item is usually 0 or simulates the computational cost of a traditional system, and Total man-hours for human-machine collaboration and token consumption ;2) Time-delay distribution pairs: including End-to-end processing time of P90 quantile values ,and End-to-end processing time of P90 quantile values 3) Abnormal event pairs: including Number of manual error processing ,and AI hallucination / abnormal trigger frequency 4) Asset accumulation: including The scale of structured data assets accumulated during the simulation cycle The number of automatically generated Standard Operating Procedure (SOP) documents The structured data assets are derived from reusable business logs, decision link records, user requirement tags, etc., generated when intelligent agents process tasks, while the SOP document is an operational specification document that can be directly implemented and extracted from collaborative simulation based on the optimal combination of configuration parameters.

[0125] After completing the pairwise data extraction, perform difference comparison calculations on the two sets of indicators, for example, calculate... Comparison Work hour savings Calculate the savings in labor costs as follows: Calculate the efficiency improvement rate of P90 Calculate the total number of statistical anomalies that have decreased. And so on. Meanwhile, due to the simulation process... and The entire operational data chain was collected, and differential calculation can cover far more dimensions than just this. It can further calculate a series of multi-dimensional indicators such as the comprehensive difference between token costs and labor costs, the proportion of AI automation processing, the increase in structured data assets, and the number of new standardized operating procedures. These quantitative efficiency comparison data obtained through differential comparison cover areas such as resource consumption, timeliness, anomaly control, and asset accumulation, mapping the calculation needs of multiple energy efficiency assessment dimensions. This not only intuitively reflects the specific impact of intelligent transformation in various dimensions, but also provides a comprehensive, accurate, and traceable input data foundation for subsequent multi-dimensional value assessment.

[0126] S620 employs a pre-defined value assessment strategy to evaluate quantitative performance comparison data across multiple energy efficiency assessment dimensions, thereby obtaining the value assessment parameters of the target directed acyclic graph across multiple energy efficiency assessment dimensions.

[0127] After obtaining quantitative performance comparison data, based on the type of the target business process identified in stage S101, a pre-defined value assessment strategy matching it is loaded. This strategy is manifested as a set of dynamic weight vectors. It stipulates the increase in income ( ), cost reduction ( ), improve efficiency ( ), quality improvement ( ),assets( The weight that each of these five value dimensions should hold. (Based on the weights) China and Africa use zero-dimensional precision to call the corresponding preset calculation models to ensure that the evaluation process is standardized and consistent.

[0128] First, calculate the increase in revenue. (If the increase in revenue has a weighting...) The calculation calculates revenue increases through direct conversion or indirect capacity release. Specifically, a dual-track model is adopted. First, through a pre-set job attribute mapping library, the user IDs (Identity Documents) of those who initiate or benefit from the business process are read, and associated with role information in the virtual user queue to determine whether the job attribute is revenue-generating (sales, enrollment consultants, teachers) or support (HR, finance, IT). Then, based on traffic data, conversion rate difference data, and number of beneficiaries data in the quantitative efficiency comparison data, the corresponding calculation path is matched: for revenue-generating positions, a pre-set direct business revenue increase model is used to coordinate the traffic of the optimization model and the pure manual baseline model. AI conversion rate With artificial conversion rate The difference, the preset average order value Using the main parameter, the direct revenue increase value is calculated using formula (3). : Formula (3) For support roles, a pre-set model for indirect empowerment and revenue increase is adopted to shorten the processing time. (For example, the onboarding process is shortened from 7 days to 2 days,) ), number of beneficiary business personnel (e.g., 50 newly hired top-performing instructors) and the projected daily revenue target for each position. (e.g., 2000 yuan / day), calculate the indirect empowerment and income-increasing value using formula (4). : Formula (4) Finally, the value assessment parameters for the revenue-generating dimension are calculated based on formula (5). : Formula (5) By using a binary model to calculate revenue growth, we can analyze the time reduction achieved through process optimization and correlate it with the benefiting business roles and their revenue-generating capabilities to calculate the indirect revenue growth resulting from efficiency improvements. For example, in an evaluation of HR attendance and onboarding processes, we identified that the collaborative participation of intelligent agents shortened the new employee onboarding process from 7 days to 2 days. Based on this, combined with the relevant information on enrollment consultant positions (with an expected daily revenue of 2,000 yuan / day), we calculated that due to the accelerated process, 50 newly recruited enrollment consultants could start working 5 days earlier. Therefore, the indirect revenue growth brought to the group by this optimization is: 50 people × 5 days × 2,000 yuan / day = 500,000 yuan. This calculation result shows that even for back-end support processes, efficiency improvements can translate into considerable indirect revenue growth.

[0129] Then, cost reduction calculations are performed. A pre-defined TCO (Total Cost of Ownership) scissors difference assessment model is used, based on resource consumption data from quantitative performance comparison data, to achieve full-cycle cost accounting. First, the operation log data of the pure manual baseline model is read to extract the total number of work orders. Manual processing time per session and the pre-set comprehensive hourly wage for the corresponding position The pure labor cost is calculated using formula (6). : Formula (6) Then, calculate the three types of costs of the collaborative optimization model one by one, and calculate the total cost of the optimization model according to formula (7). : Formula (7) Among them, the remaining labor costs Based on the collaborative optimization model, the human-machine collaboration time, the amount of tasks returned to manual processing, and the hourly wage of the corresponding positions are calculated to represent the cost of the portion of the work returned to manual processing due to AI rejection or low confidence; AI computing power cost. Using a pre-defined token-based billing standard, the total number of agent calls in the collaborative optimization model is extracted. , and the The amount of input Tokens for each call Output Token Quantity Combined with the preset prompt word unit price Preset unit price of generated content AI computing power cost is calculated using formula (8). :

[0130] Formula (8) Hidden risks / correction costs, also known as risk penalty values. The cost reduction dimension value assessment parameters are calculated based on formula (9) obtained through formula (2) above. : Formula (9) It also simultaneously outputs the estimated annual savings, detailed differences in various costs, and the break-even point time coordinates, providing quantitative support for cost control.

[0131] Next, efficiency improvement calculations are performed. A pre-defined timeliness analyzer is used, with timeliness distribution pairs and queue backlog data from the quantitative efficiency comparison data as the main inputs for evaluation. First, based on simulation records, for each task... Calculate its end-to-end processing time See formula (10): Formula (10) in, This represents the virtual queue waiting time (reflecting the impact of concurrency and queuing). For node processing time, This represents the transmission time between nodes. Based on the above calculations, two complete sets of models are obtained: the purely artificial baseline model and the collaborative optimization model. The datasets were then analyzed using distributional statistics to calculate the P90 (90th percentile) and P99 (99th percentile) values ​​for each dataset, thus obtaining long-tail timeliness performance. Based on the obtained time distribution, a timeliness threshold corresponding to a pre-defined SLA target was introduced. As a time-sensitive threshold (e.g., completion within 1 hour), statistics were performed separately on the two models to determine which met the requirements. The number of work orders under the given conditions is used to calculate the SLA achievement rate for each, thereby obtaining the SLA achievement rate improvement value. See formula (11): Formula (11) here, and This refers to the collaborative optimization model and the purely manual baseline model, where each task... value, This represents the total number of work orders.

[0132] Based on the completed timeliness analysis, the concurrent throughput capacity of the process is further evaluated through stress test simulation, and the concurrent throughput improvement factor is calculated. .

[0133] Specifically, in a virtual simulation environment, with an initial, low request sending frequency. Initially, the load pressure on the virtual demand generator is continuously and steadily increased. The entire testing process involves close monitoring of key health indicators, including queue backlog on each node, end-to-end response time distribution, and resource utilization. The test will continue until a performance indicator exceeds a preset circuit breaker threshold, at which point the corresponding request input frequency... This is recorded as the maximum concurrent throughput of the model under the current configuration. .

[0134] To ensure the objectivity and business relevance of the test results, circuit breaker boundary conditions for determining when the load capacity limit has been reached were defined before the test. The stress test terminates if any one of these conditions is met: the first condition is SLA circuit breaker failure, meaning the end-to-end P90 response time exceeds the maximum value tolerated by the business. For example, set to 24 hours); the second is queue overflow, which means that the length of the queue of tasks to be processed at any process node shows a continuous and irreversible increase, indicating that the backlog speed has permanently exceeded the processing speed; the third is resource saturation, which specifically means that the utilization rate of manual nodes reaches 100% and cannot be released through internal scheduling, forming a rigid bottleneck.

[0135] Specifically, stress tests are performed on purely artificial baseline models. For example... The load begins, and after each simulation round, if the queue is empty and the operator is idle, the load is increased by a fixed step size (e.g., increasing by a certain amount each time). The load is increased in stages. This process continues until the aforementioned circuit breaker condition is triggered, for example, when the load increases to... At that time, the simulation detected an infinite backlog of node queues and a breach of the SLA target due to complete reliance on manual processing. The load value at this moment was recorded as the maximum concurrent throughput of the pure manual baseline model, i.e. .

[0136] Next, stress test the collaborative optimization model, which is loaded with the optimal configuration parameter combination obtained through optimization. The test can start with a low load, or more efficiently, directly from the bottleneck values ​​of the baseline model. Getting Started. Because the processing power of the agent nodes in the model is extremely high (typically on the millisecond level), it remains more than capable even when the load reaches the traditional human bottleneck. Therefore, exponential or large-scale jump strategies can be used to quickly increase the load (e.g., testing sequentially). Under high load, the bottleneck shifts from the agent node to the human review node that handles complex anomalies (such as AI rejections or low-confidence tasks). As the load continues to increase until the human review node reaches resource saturation or queue overflow, triggering the circuit breaker (e.g., in...), the bottleneck persists. When the test terminates, this load value is recorded as the maximum concurrent throughput of the collaborative optimization model. .

[0137] Finally, the concurrent throughput improvement factor is obtained using formula (12). : Formula (12) Substituting the example values ​​above, we get... times.

[0138] Then, quality improvement calculations are performed. A pre-defined consistency scoring model is used, and the evaluation is conducted based on anomaly pairs in the quantitative performance comparison data and the main inputs of the process execution data. The value assessment parameter for the quality improvement dimension is the comprehensive quality improvement index. See formula (13): Formula (13) Among them, the compliance rate has improved. The calculation is shown in formula (14):

[0139] Formula (14) During the calculation, the AI ​​illusion rate is calculated based on the number of AI illusions, the number of human errors, and the total business volume for each anomalous event pair. Compared with human error rate Thus obtain .

[0140] Among them, the degree of standardization has been improved. The calculation is shown in formula (15): Formula (15) The calculation is based on the number of standard steps executed fully automatically by AI. Total number of steps Combined with the standard deviation of the preset process execution path (The variance of the AI ​​execution path is extremely small), thus obtaining .

[0141] Finally, asset value is calculated. A pre-defined data asset entropy reduction model is used, taking asset accumulation data as input, quantifying unstructured data and transforming it into the implicit value of the asset, outputting asset-dimensional value assessment parameters. See formula (16): Formula (16) Among them, data asset value-added The calculation is shown in formula (17): Formula (17) During calculation, For the total number of work orders, extract the amount of unstructured interaction logs corresponding to the scale of structured data assets in the quantitative performance comparison data. (No. The effective data volume of each unstructured interaction log (such as dialogue text), combined with the preset structured cleaning rate. (AI's ability to convert a single unstructured log entry into a structured tag, such as 0.8; failure to convert is recorded as 0), and the market valuation of a single structured data entry. (A fixed threshold is set, such as valuing a single valid piece of structured data containing employee turnover intentions at 50 yuan, to ensure that the value of a single piece of data does not exceed the limit), thus obtaining... .

[0142] Among them, the value of knowledge accumulation The calculation is shown in formula (18): Formula (18) During the calculation, the number of newly added standardized operating procedures was extracted from the quantitative performance comparison data. Combined with the cost of traditionally manually writing SOP documents of the same quality Thus obtain .

[0143] final As a parameter for value assessment in the asset dimension.

[0144] After calculating the value assessment parameters for multiple energy efficiency assessment dimensions, the value assessment parameters for each dimension are standardized according to the normalization rules of the preset value assessment strategy (mapping parameters of different dimensions to a unified standard). (Interval) combined with the matching weight vector (The weights of each dimension are configured according to the differences in process types identified by S101), and the comprehensive value assessment parameters are calculated by formula (19). This provides a comprehensive and accurate quantitative basis for the generation of subsequent value assessment reports.

[0145] Formula (19) in, Representing value dimensions, corresponding to increased revenue in sequence ( ), cost reduction ( ), improve efficiency ( ), quality improvement ( ),assets( These five value dimensions; Original value assessment parameters corresponding to each dimension (such as) , wait), Normalization, or standardization function, is used to normalize raw proxy indicator data of different dimensions and ranges. Mapped to interval, The adaptive weights are the weight coefficients for each dimension, based on the weight vector. Sure.

[0146] S630 generates a value assessment report based on value assessment parameters from multiple energy efficiency assessment dimensions.

[0147] First, the output value assessment parameters from each dimension are formatted and compiled. Based on these parameters, the final verified value indicators are calculated and highlighted, such as: the exact annual cost reduction, quantifiable efficiency improvement multiples, a clear investment payback period, and the calculated comprehensive value score. These indicators together constitute an authoritative quantitative summary of the project's value.

[0148] Meanwhile, intelligent decision-making suggestions are refined based on parameters of various dimensions. For example, when outputting specific parameter adjustment suggestions at the parameter tuning level, the corresponding cost reduction and risk penalty value are marked; when outputting resource supplementation suggestions at the resource allocation level, the input cost of new resources and expected cost reduction are clearly defined; when giving clear conclusions at the project initiation level, the achievement requirements of parameters such as comprehensive value score and ROI are clearly defined.

[0149] Ultimately, the value indicators, multi-dimensional quantitative parameters, and decision-making suggestions with specific data support are automatically packaged into a well-structured value assessment report that can directly support decision-making. This report not only builds upon the quantitative results of the S620 but also provides practical suggestions to support the decision-making and advancement of enterprise intelligent transformation.

[0150] In this embodiment, by constructing and executing parallel comparative simulations of a purely manual baseline model and a collaborative optimization model, the objective and accurate quantification of the value of business process transformation is achieved. Multi-dimensional indicator data is automatically collected and differential calculations are performed to generate a quantitative performance comparison dataset. Based on this, value assessment is completed across multiple dimensions. Finally, by integrating multi-dimensional assessment parameters and quantitative decision-making suggestions, a value assessment report containing core value indicators, financial analysis, and clear action guidelines is automatically generated. This provides enterprises with a scientific and comprehensive basis for decision-making before project investment, significantly improving the accuracy of investment decisions and the efficiency of resource allocation.

[0151] In the above Figure 6 Based on the corresponding embodiments, in order to more clearly demonstrate the process of generating a visual view, this application also provides a possible implementation of generating a visual view in a business process refactoring method based on intelligent agents. Figure 7 This is a flowchart illustrating the generation of a visual view in a business process refactoring method based on intelligent agents provided in this application. For example... Figure 7 As shown, in S630 above, after obtaining the value assessment parameters of the target directed acyclic graph in multiple energy efficiency assessment dimensions, the method further includes: S710 normalizes the value assessment parameters of multiple energy efficiency assessment dimensions and maps them onto a preset multi-dimensional value radar chart to obtain the value radar chart corresponding to the target directed acyclic graph.

[0152] After obtaining value assessment parameters from multiple energy efficiency evaluation dimensions, a value radar chart can be generated. First, the original value assessment parameters for the five dimensions—revenue increase, cost reduction, efficiency improvement, quality improvement, and assets—are normalized and uniformly mapped to… The data is divided into intervals to eliminate differences in dimensions and achieve cross-dimensional comparability. These normalized scores are then used as coordinate points for the five dimensions, projected onto a pre-defined five-axis radar chart coordinate system, and the points are connected to form a closed polygon. The value radar chart, in an intuitive and visual way, clearly reveals the balance and asset focus of the optimized business process across various value dimensions. For example, a radar chart showing a strong emphasis on revenue growth and efficiency improvement in a Type 1 process, while the cost reduction dimension is relatively flat, indicates that the project has aggressive growth and efficiency-first characteristics; conversely, a Type 3 process may show a stable dominance in cost reduction and quality improvement. This visualization helps decision-makers quickly determine the alignment between the project and the current asset direction of the group.

[0153] S720, based on the value assessment parameters of the cost reduction assessment dimension among the value assessment parameters of multiple energy efficiency assessment dimensions, draws the pure labor cost curve and the collaborative cost curve, and obtains the cost recovery scissor difference diagram corresponding to the target directed acyclic graph.

[0154] Based on detailed value assessment parameters from the cost reduction evaluation dimension, a cost recovery scissors chart is generated. The cost recovery scissors chart plots time (usually months or years) on the horizontal axis (X-axis) and cumulative investment cost on the vertical axis (Y-axis), showing two key curves: one is the pure labor cost curve, which simulates the linear growth trend of labor costs with business volume without intelligent transformation, appearing as a relatively steep straight line or quasi-straight line; the other is the collaborative optimization cost curve, which realistically reflects the cost after adopting the intelligent agent transformation solution. This curve starts higher in the initial stage due to the inclusion of one-time sunk costs such as technology R&D deployment, but its growth slope slows significantly in the subsequent operation stage due to automation substitution and efficiency improvement. The intersection point of the two curves, i.e., the break-even point, is automatically calculated and marked, and the corresponding recovery period is clearly indicated (e.g., expected cost recovery in 8.5 months). This chart intuitively reveals the long-term economics of technology investment and serves as a basis for assessing the financial feasibility of the project.

[0155] S730, based on the value assessment parameters of efficiency improvement assessment dimension and quality improvement assessment dimension among the value assessment parameters of multiple energy efficiency assessment dimensions, marks the congestion index of each process node in the target directed acyclic graph, and obtains the process bottleneck heat map corresponding to the target directed acyclic graph.

[0156] To further diagnose the internal operational status of the optimized business process, a bottleneck heatmap is generated by combining evaluation parameters from efficiency-improvement dimensions (such as node processing time and queue waiting time) and quality-improvement dimensions (such as error rate and anomaly frequency). The bottleneck heatmap uses a directed acyclic graph (DAG) passing through the target as its base map. Based on congestion indices calculated from simulation data (comprehensively reflecting queue length, processing delay, and error frequency), each process node in the graph is colored using a color gradient (e.g., green for smooth flow, yellow for mild congestion, and red for severe congestion). The bottleneck heatmap not only displays the pressure distribution within the process but also highlights new bottlenecks identified under the new collaborative model. For example, it may clearly reveal a new contradiction not seen in traditional processes: the extremely fast processing speed of upstream intelligent agent nodes leads to a severe backlog of tasks at the downstream sole manual review node. The bottleneck heatmap can guide targeted resource allocation or process fine-tuning, ensuring that the optimization solution continues to achieve its expected effectiveness in actual operation.

[0157] In this embodiment, by generating a value radar chart, a cost recovery gap chart, and a process bottleneck heatmap, the multidimensional evaluation results are transformed into an intuitive and visual decision-making view, which greatly reduces the understanding threshold for decision-makers and improves the efficiency and accuracy of data-driven decision-making and optimization.

[0158] In the above Figure 1 Based on the corresponding embodiments, to more clearly demonstrate the process of generating feasible business documents, this application also provides a possible implementation of generating feasible business documents in a business process refactoring method based on intelligent agents. Optionally, in S420 above, after defining collaborative attributes for the node relationships between intelligent agent nodes and artificial nodes in the refactored topology structure according to the optimal configuration parameter combination, and obtaining the target directed acyclic graph, the method further includes: S810 generates a human-machine collaborative business process diagram and standardized operating instructions corresponding to the target business process based on the target directed acyclic graph.

[0159] In this embodiment, based on the target directed acyclic graph (DAG), a flowchart is first rendered. The topology and collaboration attributes of the target DAG are read, and an intuitive "To-Be" human-machine collaboration business flowchart is automatically generated. To intuitively distinguish process differences and collaboration logic, differentiated visualization marking rules are adopted: nodes where the intelligent agent replaces the original human are marked in green; newly added exception handling or human review branches are marked in red; and the critical point of human-machine switching is marked in yellow. At the same time, annotations are added next to the nodes. For example, next to the "Attendance Consultation" intelligent agent node, it is noted: "The AI ​​confidence threshold here is set to 0.78. If it is lower than this value, it will automatically be transferred to a second-line human customer service representative." This ensures that the flowchart not only restores the logical structure but also clearly defines the collaboration attributes.

[0160] Simultaneously with flowchart generation, configuration files that can be directly imported into business systems can be generated, supporting formats such as YAML and JSON. These configuration files contain complete optimal configuration parameters, including AI model selection (e.g., AI_Model:GPT-4), model tuning parameters (e.g., Temperature: 0.2), confidence thresholds (e.g., Handover_Threshold: 0.78), and maximum retries (e.g., Max_Retries: 3). These can be directly imported into downstream business systems or automation platforms for rapid strategy deployment.

[0161] Finally, leveraging the text generation capabilities of large language models, standardized operating procedures (SOPs) for human-machine collaboration are automatically generated. The SOPs detail the division of responsibilities, triggering conditions, operational procedures, and exception handling processes for each step in natural language. For example, regarding the "attendance exception appeal" scenario, the SOP clearly states: AI's responsibilities: automatically collect attendance screenshots, complete optical character recognition and scheduling comparison; Human's responsibilities: only review the system-pushed work order within 2 hours if the AI-marked image is blurry or logically conflicting; Collaboration benefits: expected to reduce your image search time by 85%; Collaboration process description: employee initiates appeal → AI automatically verifies → result confidence assessment → direct feedback upon meeting standards → transfer to human for non-compliance → feedback after human review, thus providing clear and actionable work guidelines for business operators, ensuring accurate and consistent implementation of business processes.

[0162] In this embodiment, the target directed acyclic graph after simulation optimization is transformed into a visual human-machine collaboration flowchart, a directly deployable configuration file, and an executable SOP, providing a comprehensive and standardized execution basis for the implementation of the human-machine collaboration mode of the target business process, and significantly reducing the threshold for process implementation and the cost of manual adaptation.

[0163] Based on the above embodiments, optionally, an AI business process reengineering simulation decision-making cockpit interface can also be provided to help users intuitively and efficiently grasp the results of process optimization.

[0164] The interface features a functional navigation bar on the left (including process configuration, AI model configuration, value assessment, etc.); the central core area presents information in multiple modules: the five-dimensional value assessment uses a radar chart to compare the multi-dimensional performance of the collaborative optimization model and the purely manual extreme model, simultaneously displaying indicators such as ROI, net return on investment, and end-to-end efficiency; the optimized process topology and heat map module visualizes each process node (such as AI intelligent analysis and second-line expert review), annotating parameters such as AI pass rate, average response time, and bottleneck warnings; the bottom intelligent decision suggestion module provides implementation solutions such as parameter optimization and capacity warnings, and also provides a comprehensive rating (based on multi-dimensional value assessment and risk analysis) to guide users on whether to initiate a project.

[0165] This interface allows users to quickly obtain multi-dimensional value insights, key operational parameters, and precise implementation suggestions for process optimization, significantly improving decision-making efficiency.

[0166] Based on the above embodiments, the present application will be explained and described below through specific scenarios.

[0167] Example 1: AI Restructuring of International Study Abroad Enrollment Consultation Process. This example focuses on the AI ​​restructuring and optimization of the international study abroad enrollment consultation process, specifically targeting Type 1, which emphasizes direct revenue growth and high concurrency. During the annual summer study abroad enrollment season, consultation volumes surge. Traditional processes rely on human sales consultants, leading to issues such as missing nighttime responses (parents' inquiries go unanswered) and lost long-tail leads (consultants focus on high-intent clients, neglecting hesitant clients), resulting in low conversion rates. To address this, zero-sample modeling was performed. A Directed Acyclic Graph (DAG) was constructed by reading the study abroad enrollment brochure. An AI sales agent with 24 / 7 responsiveness was introduced, and 10,000 virtual parent requests were generated to simulate different purchase intentions (high / low) and active time periods throughout the day, completing scenario modeling.

[0168] Subsequently, it was identified that this process directly faces external customers and generates transaction volume, thus classifying it as Type 1, and automatically loading a weight vector. The core strategy is to prioritize increasing revenue, allowing for a significant increase in token costs.

[0169] In the configuration parameter combination optimization phase, within the virtual simulation environment, the optimization action space includes: the agent's confidence threshold. AI model selection and the number of nighttime customer service staff. In the initial stage, the agent attempted an aggressive strategy (setting a low threshold). (AI followed the order throughout the process). Although the response was fast, the inability to handle complex emotions led to a lower success rate and lower reward value. A conservative strategy was also tried (setting a high...). (This almost immediately led to manual intervention), resulting in unattended nighttime traffic and unsatisfactory reward values.

[0170] After several rounds of simulation, the policy network continuously adjusted the parameter combinations and eventually converged to an optimal set of configuration parameters. This combination explicitly indicates that the optimal confidence threshold should be set for the agent nodes responsible for intention cleaning. GPT-4 was selected; at the same time, the number of staff allocated for nighttime care was increased. Optimized to the minimum necessary level. Under this configuration, a highly efficient collaborative mode is achieved, in which the intelligent agent is responsible for front wheel cleaning and fully managed at night, and accurately switches to human intervention only when the confidence level is insufficient.

[0171] Based on this optimal configuration parameter combination, differential value quantification was performed. Simulation results show that, for revenue increase calculations, the AI's instant response capability improved the lead conversion rate from 2% to 2.5%, resulting in a direct revenue increase of 1 million yuan. Simultaneously, the agent converted 10 additional orders from low-intent leads abandoned manually and nighttime leads, generating an additional 200,000 yuan in revenue. For cost reduction calculations, although some manual initial screening time was saved, the consumption of GPT-4 tokens increased operating costs by 50,000 yuan, resulting in a net increase in revenue of 1.15 million yuan.

[0172] During the decision-making output stage, the value radar chart showed a perfect score in the revenue increase dimension and a low score in the cost reduction dimension, giving a strong recommendation to initiate the project: although the operating cost increases by 50,000 yuan, it can bring a direct revenue increase of 1.2 million yuan, with an ROI as high as 24:1. At the same time, it automatically generates a human-machine collaboration flowchart, providing a feasible reference solution for the AI ​​restructuring of Type 1 processes.

[0173] Example 2: AI Transformation of HR Attendance Consultation and Exception Handling Processes. This example focuses on Type 3, HR attendance consultation and exception handling processes that emphasize cost reduction and efficiency improvement and are risk-sensitive, and implements AI transformation. The group's annual consultation volume for this process reaches 2200 per person, requiring 40 attendance specialists. During peak periods at the end of the month, specialists are often busy responding to repeated inquiries such as unsuccessful clock-ins, leading to delays in payroll calculation. Simultaneously, AI processing of this type of business has a 5% illusion rate; if an incorrect interpretation of attendance policies is given, it may trigger employee complaints or even labor arbitration, constituting a high-risk scenario. Based on this, probabilistic modeling is performed to clarify the illusion rate of AI nodes and the corresponding risk consequences, completing scenario modeling.

[0174] Subsequently, it was identified that the service recipients of this process were internal employees and that it did not generate direct revenue, thus classifying it as Type 3, and automatically loading the weight vector. The core strategy is to reduce costs and improve efficiency, strictly control the calculation of revenue increase to avoid data fraud, and set the weight of revenue increase to 0.

[0175] In the configuration parameter combination optimization phase, within the virtual simulation environment, the optimization action space includes: the agent's confidence threshold. AI model selection and the number of attendance specialists needed. After several rounds of simulation, the policy network found that for tasks such as "vacation balance inquiry," even with a low confidence threshold, the agent's accuracy was extremely high, suggesting that branching should be fixed for such task paths. However, for high-risk, hallucination-prone tasks such as "apology for abnormal work schedules," to avoid errors, the confidence threshold for these agent nodes was ultimately lowered. The version was adjusted to 0.92, meaning that the vast majority of such issues are now forcibly transferred to manual review, constituting route insertion and forcing high-risk tasks to be reviewed manually. At the same time, it is recommended to use a model version with a lower illusion rate but slightly higher cost, and to manually compile the data. reduce.

[0176] Based on this optimal configuration parameter combination, differential value quantification was performed. Simulation results show that, for cost reduction calculations, the pure manual baseline model costs 600,000 yuan per year for 40 specialists working 10% of the time; the collaborative optimization model requires only 10% of the manpower, reducing manpower costs to 60,000 yuan, AI computing power token consumption to 20,000 yuan, and reserving 10,000 yuan for error correction costs to address hidden risks. The total annual operating cost is reduced from 600,000 yuan to 90,000 yuan, achieving a net cost reduction of 510,000 yuan. For efficiency improvement calculations, the average response time is shortened from a peak queue of 4 hours to an instant response of 1 minute, and the SLA achievement rate is significantly improved by 45%. For revenue increase calculations, since it is a back-end process, the revenue increase is determined to be 0, but the implicit quality improvement value of significantly improved employee satisfaction is also pointed out.

[0177] During the decision-making output stage, the value radar chart showed outstanding scores in cost reduction and efficiency improvement, generating the "Attendance Robot Collaboration Specification" SOP, which clearly defines risk control clauses such as "for explanations involving salary deductions, AI must display a disclaimer and provide a button to transfer to human assistant." At the same time, a human-machine collaboration flowchart was generated, providing a feasible reference solution for the AI ​​transformation of Type 3 processes.

[0178] The following describes the agent-based business process refactoring apparatus and electronic device provided in this application, which are used to perform the application. The specific implementation process and technical effects are described above and will not be repeated below.

[0179] Figure 8 A schematic diagram of a business process refactoring device based on intelligent agents provided in this application is shown below. Figure 8 As shown, the agent-based business process refactoring device includes: The acquisition module 1000 is used to acquire the definition file of the target business process and the preset business constraint rules.

[0180] The mapping module 2000 is used to construct a directed acyclic graph corresponding to the target business process based on the definition file; the process nodes in the directed acyclic graph include: human nodes and intelligent agent nodes.

[0181] The calculation module 3000 is used to iteratively optimize the configuration parameters of the directed acyclic graph according to the business constraint rules, so as to obtain the optimal configuration parameter combination of the directed acyclic graph. The optimal configuration parameter combination includes the optimal configuration parameters of each artificial node and the optimal configuration parameters of each intelligent agent node.

[0182] The mapping module 2000 is also used to reconstruct the directed acyclic graph based on the optimal combination of configuration parameters to obtain the target directed acyclic graph corresponding to the target business process.

[0183] Optionally, the mapping module 2000 is specifically used to parse the definition file to obtain each process node and the dependencies between nodes; construct an initial directed acyclic graph based on the process nodes and the dependencies between nodes; identify the type of each process node in the initial directed acyclic graph based on the metadata tags of each process node in the initial directed acyclic graph, and determine the operation category tag of each process node; wherein, the operation category tag is used to indicate whether the process node is a human node or an intelligent agent node; add the operation category tags of each process node to the initial directed acyclic graph to obtain the directed acyclic graph corresponding to the target business process.

[0184] Optionally, the calculation module 3000 is specifically used to select a set of configuration parameter combinations from the corresponding action space using a preset strategy network corresponding to the business constraint rules; adjust the configuration parameters of each process node in the directed acyclic graph according to the configuration parameter combinations, and obtain the simulation quantization performance data of the directed acyclic graph after the configuration parameter adjustment; calculate the reward value corresponding to the configuration parameter combination according to the simulation quantization performance data using a preset reward function; update the strategy network using a near-end strategy optimization algorithm according to the reward value; select a new configuration parameter combination from the action space again using the updated strategy network, and obtain the new reward value corresponding to the new configuration parameter combination, until the new reward value meets the preset stopping iteration condition; and obtain the configuration parameter combination with the highest reward value when the preset stopping iteration condition is met as the optimal configuration parameter combination.

[0185] Optionally, the mapping module 2000 is specifically used to optimize the topology of the directed acyclic graph according to the optimal combination of configuration parameters to obtain the reconstructed topology; and to define the cooperation attributes of the node relationships between agent nodes and human nodes in the reconstructed topology according to the optimal combination of configuration parameters to obtain the target directed acyclic graph; wherein, the cooperation attributes of the node relationships are used to indicate the cooperation mode of human-led agent assistance, the automatic mode of agent-led human supervision, or the shadow mode of parallel operation of agent and human.

[0186] Optionally, the agent-based business process reconfiguration device further includes: a simulation module 4000, used to perform pure manual operation simulation on each process node in the target directed acyclic graph using a pure manual baseline model to obtain the manual indicator data corresponding to the target directed acyclic graph; and to perform collaborative simulation of manual operation and agent operation on each process node in the target directed acyclic graph using a collaborative optimization model to obtain the collaborative indicator data corresponding to the target directed acyclic graph.

[0187] Optionally, the calculation module 3000 is also used to generate a value assessment report corresponding to the target directed acyclic graph based on the manual indicator data and the collaborative indicator data.

[0188] Optionally, the calculation module 3000 is specifically used to perform differential comparison calculations based on manual indicator data and collaborative indicator data to obtain quantitative efficiency comparison data; adopt a preset value assessment strategy to evaluate the quantitative efficiency comparison data in multiple energy efficiency assessment dimensions to obtain the value assessment parameters of the target directed acyclic graph in multiple energy efficiency assessment dimensions; and generate a value assessment report based on the value assessment parameters of multiple energy efficiency assessment dimensions.

[0189] Optionally, the mapping module 2000 is also used to normalize the value assessment parameters of multiple energy efficiency assessment dimensions and map them onto a preset multi-dimensional value radar chart to obtain the value radar chart corresponding to the target directed acyclic graph; based on the value assessment parameters of the cost reduction assessment dimension among the value assessment parameters of multiple energy efficiency assessment dimensions, it draws the pure labor cost curve and the collaborative cost curve to obtain the cost recovery scissor difference chart corresponding to the target directed acyclic graph; based on the value assessment parameters of the efficiency improvement assessment dimension and the quality improvement assessment dimension among the value assessment parameters of multiple energy efficiency assessment dimensions, it marks the congestion index of each process node in the target directed acyclic graph to obtain the process bottleneck heat map corresponding to the target directed acyclic graph.

[0190] Optionally, the mapping module 2000 is also used to generate a human-machine collaborative business process diagram and a standardized operating instruction manual corresponding to the target business process based on the target directed acyclic graph.

[0191] These modules can be one or more integrated circuits configured to implement the above methods, such as one or more Application Specific Integrated Circuits (ASICs), one or more digital signal processors (DSPs), or one or more Field Programmable Gate Arrays (FPGAs). Alternatively, when a module is implemented using processing element scheduler code, the processing element can be a general-purpose processor, such as a Central Processing Unit (CPU) or other processor capable of calling program code. Furthermore, these modules can be integrated together as a system-on-a-chip (SOC).

[0192] Figure 9 This is a schematic diagram of an electronic device provided in this application. The device may be a computing device or a server with computing processing capabilities.

[0193] The electronic device 10 includes a processor 11, a storage medium 12, and a bus 13. The storage medium 12 stores program instructions executable by the processor 11. When the electronic device 10 is executed, the processor 11 communicates with the storage medium 12 via the bus 13, and the processor 11 executes the program instructions to perform the above-described method embodiment. The specific implementation and technical effects are similar and will not be described in detail here.

[0194] Optionally, this application also provides a program product, such as a computer-readable storage medium, including a program that, when executed by a processor, performs the above-described method embodiments.

[0195] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0196] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0197] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in a combination of hardware and software functional units.

[0198] The integrated units implemented as software functional units described above can be stored in a computer-readable storage medium. These software functional units, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0199] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.< / laneset>

Claims

1. A business process refactoring method based on intelligent agents, characterized in that, The method includes: Obtain the definition file and preset business constraint rules of the target business process; Based on the definition file, construct a directed acyclic graph corresponding to the target business process; wherein, each process node in the directed acyclic graph includes: human nodes and intelligent agent nodes; According to the business constraint rules, the configuration parameters of the directed acyclic graph are iteratively optimized to obtain the optimal configuration parameter combination of the directed acyclic graph. The optimal configuration parameter combination includes: the optimal configuration parameters of each artificial node and the optimal configuration parameters of each intelligent agent node. Based on the optimal configuration parameter combination, the directed acyclic graph is reconstructed to obtain the target directed acyclic graph corresponding to the target business process.

2. The method according to claim 1, characterized in that, The step of constructing the directed acyclic graph corresponding to the target business process based on the definition file includes: Parse the definition file to obtain each process node and the dependencies between nodes; Based on the process nodes and the dependencies between nodes, an initial directed acyclic graph is constructed; Based on the metadata tags of each process node in the initial directed acyclic graph, the type of each process node in the initial directed acyclic graph is identified, and the operation category tag of each process node is determined. The operation category tag is used to indicate whether the process node is a human node or an intelligent agent node. Add the operation category labels of each process node to the initial directed acyclic graph to obtain the directed acyclic graph corresponding to the target business process.

3. The method according to claim 1, characterized in that, The step of iteratively optimizing the configuration parameters of the directed acyclic graph (DAG) according to the business constraint rules to obtain the optimal combination of configuration parameters for the DAG includes: The preset strategy network corresponding to the business constraint rule selects a set of configuration parameter combinations from the corresponding action space; Based on the configuration parameter combination, adjust the configuration parameters of each process node in the directed acyclic graph, and obtain the simulation quantization performance data of the directed acyclic graph after the configuration parameter adjustment; A preset reward function is used to calculate the reward value corresponding to the combination of configuration parameters based on the simulated quantitative performance data. Based on the reward value, the policy network is updated using a near-end policy optimization algorithm; The updated policy network is used to select a new combination of configuration parameters from the action space and obtain a new reward value corresponding to the new combination of configuration parameters until the new reward value meets the preset stopping iteration condition. The optimal configuration parameter combination is the combination that yields the highest reward value when the preset stopping iteration condition is met.

4. The method according to claim 1, characterized in that, The step of reconstructing the directed acyclic graph based on the optimal configuration parameter combination to obtain the target directed acyclic graph corresponding to the target business process includes: Based on the optimal configuration parameter combination, the topology of the directed acyclic graph is optimized to obtain the reconstructed topology. Based on the optimal configuration parameter combination, a collaborative attribute is defined for the node relationship between agent nodes and human nodes in the reconstructed topology, resulting in the target directed acyclic graph. The collaborative attribute of the node relationship is used to indicate the collaborative mode of human-led agent assistance, the automatic mode of agent-led human supervision, or the shadow mode of parallel operation of agent and human.

5. The method according to claim 1, characterized in that, The method further includes: Using a purely manual baseline model, a purely manual operation simulation is performed on each process node in the target directed acyclic graph to obtain the manual indicator data corresponding to the target directed acyclic graph. A collaborative optimization model is used to perform collaborative simulation of manual operations and intelligent agent operations on each process node in the target directed acyclic graph, thereby obtaining collaborative index data corresponding to the target directed acyclic graph. Based on the artificial indicator data and the collaborative indicator data, a value assessment report corresponding to the target directed acyclic graph is generated.

6. The method according to claim 5, characterized in that, The step of generating a value assessment report corresponding to the target directed acyclic graph based on the artificial indicator data and the collaborative indicator data includes: Based on the artificial indicator data and the collaborative indicator data, differential comparison calculation is performed to obtain quantitative performance comparison data; A preset value assessment strategy is adopted to evaluate the quantitative efficiency comparison data in multiple energy efficiency assessment dimensions, and the value assessment parameters of the target directed acyclic graph in the multiple energy efficiency assessment dimensions are obtained. The value assessment report is generated based on the value assessment parameters of the multiple energy efficiency assessment dimensions.

7. The method according to claim 6, characterized in that, After obtaining the value assessment parameters of the target directed acyclic graph in the multiple energy efficiency assessment dimensions, the method further includes: The value assessment parameters of the multiple energy efficiency assessment dimensions are normalized and mapped onto a preset multi-dimensional value radar map to obtain the value radar map corresponding to the target directed acyclic graph. Based on the value assessment parameters of the cost reduction assessment dimension among the value assessment parameters of the multiple energy efficiency assessment dimensions, draw the pure labor cost curve and the collaborative cost curve to obtain the cost recovery scissor difference diagram corresponding to the target directed acyclic graph. Based on the value assessment parameters of efficiency improvement assessment dimension and quality improvement assessment dimension among the value assessment parameters of the multiple energy efficiency assessment dimensions, the congestion index of each process node in the target directed acyclic graph is marked to obtain the process bottleneck heat map corresponding to the target directed acyclic graph.

8. The method according to claim 4, characterized in that, After defining cooperative attributes for the node relationships between agent nodes and artificial nodes in the reconstructed topology based on the optimal configuration parameter combination to obtain the target directed acyclic graph, the method further includes: Based on the target directed acyclic graph, generate a human-machine collaboration business process diagram and a standardized operating procedure corresponding to the target business process.

9. A business process refactoring device based on intelligent agents, characterized in that, The device includes: The acquisition module is used to acquire the definition file of the target business process and the preset business constraint rules; The mapping module is used to construct a directed acyclic graph of the target business process based on the definition file; wherein, each process node in the directed acyclic graph includes: human nodes and intelligent agent nodes; The calculation module is used to iteratively optimize the configuration parameters of the directed acyclic graph according to the business constraint rules, so as to obtain the optimal configuration parameter combination of the directed acyclic graph. The optimal configuration parameter combination includes: the optimal configuration parameters of each artificial node and the optimal configuration parameters of each intelligent agent node. The mapping module is also used to reconstruct the directed acyclic graph according to the optimal configuration parameter combination to obtain the target directed acyclic graph of the target business process.

10. An electronic device, characterized in that, include: The device includes a processor, a memory, and a bus. The memory stores program instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus, and the processor executes the program instructions to perform the steps of the agent-based business process refactoring method as described in any one of claims 1 to 8.

Images