A method for generating post-loan early warning and differentiated collection schemes based on AI intelligent agents

By constructing a post-loan customer status migration map and using AI agents to generate simulated collection actions, the problems of untimely risk identification and lack of dynamic adjustment of collection strategies in existing post-loan management have been solved. This has enabled more accurate risk identification and the generation of differentiated collection plans, thereby improving the efficiency and effectiveness of post-loan management.

CN123118239APending Publication Date: 2026-11-17ZHONGKE JUXIN INFORMATION TECH BEIJING CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611178638.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-05
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing post-loan management methods struggle to promptly identify dynamic shifts in borrowers from normal repayment status to overdue, unreachable, or uncooperative states, and collection strategies lack dynamic adjustments, leading to problems such as untimely risk identification and high costs associated with disrupting customers.

Method used

By acquiring borrowers' post-loan behavior data, a post-loan customer status migration map is constructed. AI agents are used to generate simulated collection actions and simulated feedback results, thereby selecting the collection path with the lowest customer disturbance cost and generating differentiated collection solutions.

Benefits of technology

It improves the timeliness of identifying potential overdue and loss of contact risks, reduces the risks of excessive outreach and non-compliant communication, and enhances the efficiency and accuracy of post-loan management.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118239A_ABST
    Figure CN123118239A_ABST
Patent Text Reader

Abstract

This invention relates to the field of post-loan risk management technology, specifically to a method for generating post-loan early warning and differentiated collection plans based on AI intelligent agents, comprising the following steps: S1: acquiring borrower's post-loan behavior data and constructing a post-loan customer state transition map; S2: generating corresponding post-loan early warning levels; S3: outputting action-feedback trajectories; S4: determining the feasible path with the lowest customer disturbance cost as the target collection path; S5: generating collection script text and communication channel arrangement instructions to obtain differentiated collection plans. This invention achieves risk warning identification through a post-loan customer state transition map, and combines dual AI intelligent agent action-feedback deduction and a customer disturbance cost screening mechanism to generate differentiated collection plans that balance repayment effectiveness, compliance requirements, and customer outreach experience, thereby improving the timeliness of post-loan early warning and the accuracy of collection execution.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of post-loan risk management technology, and in particular to a method for generating post-loan early warning and differentiated collection schemes based on AI intelligent agents. Background Technology

[0002] With the expansion of consumer finance, internet lending, and micro-loan businesses, post-loan management requires continuous tracking of borrowers' repayment ability, willingness to repay, contact responsiveness, and delinquency trends. Traditional post-loan management typically relies on rule thresholds, experience-based manual outbound calls, or static customer segmentation to uniformly remind or collect payments from customers. This approach struggles to capture the dynamic process of borrowers transitioning from normal repayment status to delinquency, loss of contact, or negative communication. In recent years, AI agents have been able to perform task planning, strategy deduction, and feedback simulation based on customer behavior data. This provides new technological pathways for post-loan risk warning, loss of contact recovery, differentiated collection action matching, and post-loan behavior time-series analysis, enabling post-loan management to gradually shift from passive collection to proactive identification, simulation assessment, and automatic strategy generation.

[0003] However, existing methods for generating post-loan early warning and collection plans still have many shortcomings: On the one hand, existing methods mostly rely on the number of overdue days, the amount owed, or a single credit score as the basis for risk assessment, lacking modeling of the dynamic migration relationship between repayment ability, repayment willingness, and contact response status. This results in insufficient and untimely identification of potential overdue payments, loss of contact tendencies, and adverse communication risks. On the other hand, existing collection strategies are usually matched according to fixed rules or human experience, making it difficult to simulate the impact of different collection actions on borrowers' repayment feedback, contact feedback, and negative feedback before execution. This can easily lead to problems such as excessively high contact intensity with low-risk customers, insufficient response from high-risk customers, unstable compliance of communication scripts, and high customer disturbance costs. Therefore, it is necessary to provide a method for generating post-loan early warning and differentiated collection plans based on AI intelligent agents to solve the above problems. Summary of the Invention

[0004] To achieve the above objectives, this invention provides a method for generating post-loan early warning and differentiated collection schemes based on AI intelligent agents.

[0005] A method for generating post-loan early warning and differentiated collection solutions based on AI intelligent agents includes the following steps: S1: Obtain the borrower's post-loan behavior data, determine the corresponding repayment ability characteristic value, repayment willingness characteristic value and contact response characteristic value, and form a post-loan customer state migration map with the borrower's post-loan status representation as nodes and the migration probability and migration rate between post-loan states as directed edges. S2: Based on the post-loan customer status migration map, identify the risk level of a borrower's migration from a normal repayment status to an overdue status, a lost contact status, or a reverse communication status, and generate the corresponding post-loan early warning level. S3: Input the post-loan customer status transition map carrying the post-loan early warning level into the digital twin simulation engine, generate simulated collection actions through the first AI agent, and simulate the borrower's feedback to the simulated collection actions through the second AI agent to obtain the action-feedback trajectory. S4: Filter feasible paths from the action-feedback trajectory to achieve the preset repayment target and meet compliance constraints, and determine the customer disturbance cost based on the collection intensity and feedback results corresponding to the feasible path. The feasible path with the lowest customer disturbance cost is determined as the target collection path. S5: Convert the target collection path into a sequence of collection actions arranged in chronological order, and generate collection script text and communication channel arrangement instructions based on the intensity level, channel type and psychological appeal category of each collection action to obtain a differentiated collection plan.

[0006] Optionally, determining the corresponding repayment ability characteristic value, repayment willingness characteristic value, and contact response characteristic value in S1 includes: S11: Obtain the borrower's historical repayment records, overdue days, account fund fluctuation records, contact response records, and repayment commitment fulfillment records to form a post-loan behavior dataset; S12: Determine the repayment ability characteristic value to characterize changes in the borrower's repayment ability based on historical repayment records, overdue days, and account fund fluctuation records; S13: Based on the records of voluntary repayment and extension negotiation in the repayment commitment performance record, determine the repayment willingness feature value used to characterize the borrower's tendency to voluntarily repay; S14: Determine the contact response feature value used to characterize the borrower's reachability based on the connection status, response delay, number of rejections, and duration of disconnection in the contact response record.

[0007] Optionally, the node representing the borrower's post-loan status specifically includes: Based on the repayment ability characteristic value, repayment willingness characteristic value, and contact response characteristic value, the borrower's post-loan status representation corresponding to the current post-loan stage is determined; The post-loan status is categorized into normal repayment status, overdue status, out-of-contact status, and reverse communication status. Each post-loan status is represented as a node in the post-loan customer status transition graph, and each node is configured with a corresponding status time label, repayment ability feature value, repayment willingness feature value, and contact response feature value.

[0008] Optionally, forming a post-loan customer state transition graph using the transition probabilities and rates between post-loan states as directed edges includes: Obtain the borrower's status transition records between adjacent post-loan states, and count the number of transitions from the first post-loan state to the second post-loan state; The migration probability from the first post-loan state to the second post-loan state is determined based on the number of transfers and the total number of occurrences corresponding to the first post-loan state. Based on the difference in state duration between the first and second post-loan states, and the magnitude of change in the corresponding eigenvalues, the migration rate from the first post-loan state to the second post-loan state is determined. Finally, using migration probability and migration rate as directed edge attributes, the corresponding post-loan state representation nodes are connected to form the post-loan customer state migration graph.

[0009] Optionally, S2 specifically includes: S21: Extract the migration probability and migration rate of borrowers from normal repayment status to overdue status, out-of-contact status, or reverse communication status from the post-loan customer status migration map. S22: Determine the borrower's risk level based on migration probability, migration rate, and current repayment willingness characteristic values; S23: Compare the risk level with the preset warning threshold, and generate a low-risk warning level, a medium-risk warning level, or a high-risk warning level according to the comparison result; S24: Write the post-loan warning level into the post-loan customer status transition map of the corresponding borrower to obtain the post-loan customer status transition map with the post-loan warning level.

[0010] Optionally, the simulated collection actions generated by the first AI agent in S3 include: S31: The first AI agent reads the post-loan warning level, the current post-loan status representation, the migration probability, and the migration rate; S32: Determine the candidate intensity range of collection actions based on the post-loan early warning level, and determine the corresponding candidate channel type based on the current post-loan status. S33: Generate simulated collection actions within the candidate intensity range and candidate channel type, wherein the simulated collection actions include action intensity level, contact channel, contact time, and psychological appeal category.

[0011] Optionally, in step S3, the second AI agent simulates the borrower's feedback to the simulated debt collection actions, resulting in an action-feedback trajectory including: S34: The second AI agent reads the repayment ability feature value, repayment willingness feature value, contact response feature value, and simulates collection actions; S35: Based on the impact of the simulated collection actions on the borrower's willingness to repay and the contact response status, output the corresponding feedback results, including repayment feedback, contact feedback and negative feedback; S36: Record each simulated collection action and its corresponding feedback result in sequence according to the simulation time to form the action-feedback trajectory.

[0012] Optionally, the feasible paths in S4 that achieve the preset repayment target and meet compliance constraints from the action-feedback trajectory include: S41: Extract candidate paths from the action-feedback trajectory, which consist of multiple simulated collection actions and corresponding feedback results; S42: Determine whether the candidate path has reached the preset repayment target within the preset repayment period, and take the candidate path that has reached the preset repayment target as the initial selection path; S43: Perform compliance verification on the initial selected path, the compliance constraints including contact frequency constraints, contact time constraints, channel authorization constraints, and script content constraints; determine the initial selected path that passes the compliance verification as the feasible path.

[0013] Optionally, determining the feasible path with the lowest customer disturbance cost as the target collection path in S4 includes: S45: Extract the action intensity level, contact frequency, contact channel, and feedback result corresponding to each simulated collection action in each feasible path; S46: Determine the customer disturbance cost for each feasible path based on the action intensity level, reach frequency, channel sensitivity corresponding to the reach channel, and degree of negative feedback in the feedback results; S47: Compare the customer disturbance costs corresponding to each feasible path, and determine the feasible path with the lowest customer disturbance cost as the target collection path.

[0014] Optionally, S5 specifically includes: S51: Divide the target collection path into multiple collection action nodes according to the time sequence, and determine the execution order, intensity level, channel type and psychological appeal category corresponding to each collection action node; S52: Match the corresponding script template from the preset script template library according to the category of psychological needs, and generate collection script text by combining the borrower's post-loan warning level and current post-loan status. S53: Generate communication channel orchestration instructions based on execution order, channel type, and intensity level; S54: Associate collection action nodes, collection script texts, and communication channel arrangement instructions to obtain differentiated collection solutions that include execution time, contact channels, collection script texts, and triggering conditions.

[0015] The beneficial effects of this invention are: This invention, by acquiring borrowers' post-loan behavior data, determines repayment ability characteristic values, repayment willingness characteristic values, and contact response characteristic values, and constructs a post-loan customer status migration map. This allows the risk of borrowers changing from a normal repayment status to an overdue, unreachable, or reversed communication status to be dynamically expressed. Compared to methods that rely solely on the number of overdue days or fixed scores, this invention combines status migration probability, migration rate, and changes in repayment willingness to generate post-loan early warning levels, thereby improving the timeliness of identifying potential overdue and unreachable risks and providing a more accurate risk basis for generating subsequent collection strategies.

[0016] This invention provides a method to generate simulated collection actions using a first AI agent and simulate borrower feedback under different collection actions using a second AI agent, forming an action-feedback trajectory. Feasible paths that meet repayment goals and compliance constraints are then selected, and the target collection path is determined based on minimizing customer disruption costs. Therefore, the collection plan no longer relies on fixed templates or human experience, but can generate differentiated collection scripts and communication channel arrangement instructions based on different borrowers' risk levels, status changes, and feedback results. This improves collection efficiency while reducing the risks of over-reaching, customer resistance, and non-compliant scripts. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the method for generating post-loan early warning and differentiated collection schemes according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the process for generating post-loan early warning levels according to an embodiment of the present invention. Detailed Implementation

[0019] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0020] It should be noted that the use of terms such as "an embodiment," "an embodiment," "an exemplary embodiment," and "some embodiments" in the specification indicates that the described embodiment may include a specific feature, structure, or characteristic, but not every embodiment necessarily includes that specific feature, structure, or characteristic. Furthermore, when a specific feature, structure, or characteristic is described in connection with an embodiment, implementing such a feature, structure, or characteristic in conjunction with other embodiments (whether explicitly described or not) should be within the knowledge of those skilled in the art.

[0021] Generally, terms can be understood at least partly from their use in context. For example, depending at least partly on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in a singular sense, or a combination of features, structures, or characteristics in a plural sense. Additionally, the term "based on" can be understood not necessarily to convey an exclusive set of factors, but rather, alternatively, depending at least partly on the context, to allow for the presence of other factors that are not necessarily explicitly described.

[0022] like Figures 1-2 As shown, a method for generating post-loan early warning and differentiated collection solutions based on AI intelligent agents includes the following steps: S1: Obtain the borrower's post-loan behavior data, determine the corresponding repayment ability characteristic value, repayment willingness characteristic value and contact response characteristic value, and form a post-loan customer state migration map with the borrower's post-loan status representation as nodes and the migration probability and migration rate between post-loan states as directed edges. In a specific implementation, the process of acquiring the borrower's post-loan behavior data and determining the corresponding repayment ability characteristic value, repayment willingness characteristic value, and contact response characteristic value includes the following steps: First, the system acquires the borrower's historical repayment records, overdue days, account balance fluctuation records, contact response records, and promised repayment fulfillment records generated during the post-loan management period. These records are then aggregated according to the borrower's unique identifier and the time of data occurrence to form a post-loan behavior dataset. The historical repayment records include the repayment date, amount due, amount actually repaid, and number of historical overdue payments. Overdue days represent the borrower's current or historical overdue status. Account balance fluctuation records include account inflows, outflows, and balance changes. Contact response records include call connection status, SMS or app message responses, response delay, number of rejected calls, and duration of continuous non-response. Promised repayment fulfillment records include the time the borrower made the repayment commitment, the committed repayment amount, the committed repayment date, and the actual fulfillment result. After obtaining the post-loan behavior dataset, the monetary, frequency, and time-related data are standardized to ensure that data of different dimensions can be included in the same feature calculation process. For any original indicator... The following methods can be used for normalization: ,in, This represents the normalized index value. This represents the borrower's original indicator value. and These represent the minimum and maximum values ​​of this indicator in a sample of borrowers of the same type, respectively. When an indicator is negative, the normalization result is reversed so that a larger value indicates a better state.

[0023] The repayment ability characteristic value is determined based on historical repayment records, overdue days, and account balance fluctuation records. Specifically, the borrower's on-time repayment ratio, actual repayment amount as a percentage of the outstanding amount, account balance stability, recent inflow stability, and the inverse indicator of overdue days are extracted and weighted according to preset weights to obtain the repayment ability characteristic value. : ,in, This represents a characteristic value of repayment ability. This indicates the proportion of payments due on schedule. This indicates the percentage of the amount actually repaid relative to the amount due. Indicates the stability of the account balance. This indicates the stability of recent revenue. This indicates a reverse indicator of the number of overdue days. Let each represent the weight of the corresponding indicator, and satisfy the following condition: Among them, the reverse indicator of overdue days Based on the current number of overdue days Compared with the preset maximum overdue reference days Sure: The longer the overdue period, The smaller the value, the lower the corresponding repayment ability characteristic value; when the borrower has a high on-time repayment ratio, a high actual repayment ratio, and a high stability of account funds, the repayment ability characteristic value will increase accordingly.

[0024] The repayment willingness characteristic value is determined based on the borrower's commitment repayment performance record, proactive repayment record, and extension negotiation record. Specifically, the borrower's proactive repayment frequency, commitment repayment performance rate, negotiation response frequency, and the reverse indicator of commitment default frequency are extracted and weighted to obtain the repayment willingness characteristic value, expressed as: ,in, This represents a characteristic value indicating the willingness to repay. Indicators representing proactive repayment behavior. This indicates the promised repayment rate. Indicates the response indicators for the postponement of negotiations. This is a reverse indicator of the number of times a commitment has been defaulted. Let each represent the weight of the corresponding indicator, and satisfy the following condition: Among them, the promised repayment fulfillment rate It can be determined by the following formula: ,in, This indicates the number of times the borrower has made the promised repayments. This indicates the total number of times a borrower has made repayment commitments. The repayment willingness characteristic value increases when a borrower actively repays, negotiates, or fulfills their commitments multiple times; conversely, the repayment willingness characteristic value decreases when a borrower makes multiple commitments but fails to fulfill them.

[0025] The contact response characteristic value is determined based on the connection status, response delay, number of rejections, and duration of lost contact recorded in the contact response log. Specifically, the borrower's call connection rate, message response rate, average response delay inverse indicator, number of rejections inverse indicator, and duration of lost contact inverse indicator are extracted and weighted to obtain the contact response characteristic value. : ,in, Indicates the characteristic value of the contact response. Indicates the call connection rate. Indicates message response rate. This represents the inverse indicator of average response time. This is a reverse indicator of the number of rejected calls. This indicates the inverse indicator of the duration of the loss of contact. Let each represent the weight of the corresponding indicator, and satisfy the following condition: Among them, the average response delay reverse index It can be determined by the following formula: ,in, This indicates the borrower's average response time. This represents the preset maximum response delay reference value. When the borrower's connection rate and message response rate are high, and the response delay is short, the contact response characteristic value is high; when the number of borrower rejections increases or the continuous period of unreachable contact lengthens, the contact response characteristic value decreases. Through the above processing, scattered post-loan behavior data can be converted into repayment ability characteristic values, repayment willingness characteristic values, and contact response characteristic values.

[0026] The specific nodes represented by the borrower's post-loan status include: Based on the repayment ability characteristic value, repayment willingness characteristic value, and contact response characteristic value, the borrower's post-loan status representation corresponding to the current post-loan stage is determined; The post-loan status is categorized into normal repayment status, overdue status, out-of-contact status, and reversed communication status. Each post-loan status is represented as a node in the post-loan customer status transition graph, and each node is configured with a corresponding status time label, repayment ability feature value, repayment willingness feature value, and contact response feature value.

[0027] Specifically, after obtaining the borrower's repayment ability characteristic value, repayment willingness characteristic value, and contact response characteristic value, these three types of characteristic values ​​are used as the basis for describing the borrower's current post-loan status, thus determining the borrower's post-loan status representation. The post-loan status representation is used to indicate the borrower's comprehensive status at a specific post-loan point in time or within a post-loan time interval. It includes at least the repayment ability characteristic value, repayment willingness characteristic value, and contact response characteristic value, and can also be further associated with the time of occurrence or duration of this status. Specifically, for the first... Each observation period can be used to characterize the borrower's post-loan status as follows: ,in, Indicates the first Post-loan status characterization within an observation period This represents the characteristic value of repayment ability within the observation period. This represents the characteristic value of repayment willingness within the observation period. This represents the characteristic value of the contact response within the observation period. This indicates the current number of overdue days within the observation period. This indicates the time label corresponding to the post-loan status.

[0028] After obtaining the post-loan status characteristics, the borrower's current post-loan status is classified based on repayment ability characteristics, repayment willingness characteristics, contact response characteristics, and the current number of overdue days. The preset overdue judgment threshold is set at one calendar day. That is, if a borrower fails to make full payment of the due amount more than one calendar day after the agreed repayment date, they are considered to have met the initial criteria for overdue status. This threshold is set based on the following: In post-loan management, the agreed repayment date is the direct time boundary for judging whether the borrower has fulfilled their obligations on time. Failure to make full payment after the agreed repayment date already reflects a change in the borrower's performance status. Furthermore, setting the threshold to one calendar day allows for timely initiation of risk identification and collection strategy simulation processes in the early stages of overdue payments, avoiding delays in issuing warnings after a long period of overdue payment, which would affect the timeliness of post-loan warnings.

[0029] Specifically, if the current number of overdue days If the repayment ability characteristic value, repayment willingness characteristic value, and contact response characteristic value are all within the normal range, the post-loan status is classified as a normal repayment status; if the current overdue days are... If the loan status is overdue for a certain number of natural days, but the contact response characteristic value is not lower than the preset loss of contact threshold, and the borrower does not exhibit obvious refusal to communicate, the loan status will be classified as overdue. If the contact response characteristic value is lower than the preset loss of contact threshold, or the continuous non-response duration exceeds the preset loss of contact duration threshold, the loan status will be classified as lost. If the repayment willingness characteristic value is lower than the preset willingness threshold, and the contact response record shows refusal to communicate, strong rejection, complaint tendency, or multiple promises of default, the loan status will be classified as reverse communication. The current overdue days are listed below. It can be determined in the following way: ,in, Indicates the first The current number of overdue days corresponding to each observation period. Indicates the current statistical date. This indicates the agreed repayment date. When... hour, This indicates that the borrower has not yet exceeded the agreed repayment date; when The current overdue days are determined based on the number of calendar days exceeding the agreed repayment date. To avoid misjudgments caused by delayed repayment arrivals or end-of-day batch processing by the system, some implementations may use the end-of-day settlement results on the repayment date as the basis for judgment. That is, only if no full repayment record is detected on the first calendar day after the agreed repayment date will the overdue period be considered overdue. The status is used for overdue status classification. This ensures timely overdue identification and reduces misclassification of status due to short-term payment delays.

[0030] After completing the state segmentation, each post-loan state representation is used as a node in the post-loan customer state transition graph. Each node is configured with at least a node identifier, borrower identifier, state type, state time label, repayment ability feature value, repayment willingness feature value, and contact response feature value. Specifically, the node identifier distinguishes different post-loan state nodes formed by the same borrower in different observation periods; the borrower identifier associates the node with the corresponding borrower; the state type indicates whether the node belongs to a normal repayment state, an overdue state, an unreachable state, or a reverse communication state; and the state time label indicates the post-loan time position corresponding to the node. Through this method, the originally scattered post-loan behavior data can be transformed into continuous post-loan state nodes, enabling a structured expression of the borrower's risk changes at different post-loan stages. The resulting nodes not only contain the current state category but also carry three key attributes: repayment ability, repayment willingness, and contact response.

[0031] In a specific implementation, a post-loan customer state transition graph is formed using the transition probabilities and rates between post-loan states as directed edges, specifically including the following process: First, based on the multiple post-loan status representation nodes formed in claim 3, the post-loan status representations of the same borrower are sorted by time according to the status time label to obtain the borrower's post-loan status sequence. The post-loan status sequence represents the process by which a borrower changes from one post-loan status to another within the post-loan management period. When statistically analyzing status transition records, records where the status type changes within adjacent observation periods are considered valid status transition records. For example, when the first... The status type for the first observation period is normal repayment status. If the status type of the first observation period is overdue, then a status transition record is created, pointing from the normal repayment status to the overdue status; when the first... The status type for the first observation period is overdue. If the state type for an observation period is "out of contact," then a state transition record is formed, pointing from the overdue state to the out of contact state. For cases where the state type remains unchanged but the state attributes continuously deteriorate, this can also be considered a continuous migration record within the same state, used to characterize the deepening of risk in that state.

[0032] Then, count the number of transitions from the first post-loan state to the second post-loan state. Let the first post-loan state be... The second post-loan status is From the first post-loan status Transition to the second post-loan status The number of transitions is denoted as Post-loan status The total number of occurrences in the post-loan status sequence is denoted as . Then the first post-loan status Towards the second post-loan status The migration probability can be determined by the following formula: ,in, This indicates that the borrower has transitioned from the first post-loan status. Entering the second post-loan status The probability, Indicates the number of state transitions. Indicates the first post-loan status The total number of occurrences. A higher transition probability indicates a higher likelihood that the borrower will transition from the current state to the target risk state. To avoid zero or excessively fluctuating transition probabilities for certain low-frequency states due to a small sample size, the implementation method can smooth the transition probabilities: ,in, This represents the smoothing coefficient, typically ranging from 0.5 to... This indicates the number of transferable post-loan status types. This approach reduces the impact of occasional status records on the migration probability calculation results, making the post-loan customer status migration map more stable.

[0033] After obtaining the transition probabilities, the transition rate from the first post-loan state to the second post-loan state is further determined. The transition rate represents how quickly the borrower's post-loan state deteriorates or changes; it is determined based on the difference in state duration between the first and second post-loan states, and the magnitude of change in the corresponding eigenvalues. Specifically, let the first post-loan state... The corresponding state representation is Second post-loan status The corresponding state representation is The corresponding eigenvalue change range Represented as: ,in, Indicates the first post-loan status With the second post-loan status The range of changes in the comprehensive characteristics between them and These represent the characteristic values ​​of repayment ability before and after the transfer, respectively. and These represent the characteristic values ​​of repayment willingness before and after the transfer, respectively. and These represent the characteristic values ​​of the connection response before and after the transfer, respectively. and These represent the current overdue days before and after the transfer, respectively. Let each represent the weight of the corresponding feature, and satisfy the following: ;Difference in state duration Represented as: ,in, Indicates the state after the first loan. Transition to the second post-loan status The time interval that has elapsed, and These represent the state time labels of two adjacent post-loan states.

[0034] Therefore, the first post-loan status Towards the second post-loan status migration rate It can be determined by the following formula: ,in, Indicates the state transition rate. When Larger and When the value is small, it indicates that the borrower's post-loan status has changed significantly in a short period of time, corresponding to a higher migration rate; when When the transition is small or the state change lasts for a long time, the corresponding migration rate is low. After determining the migration probability and migration rate, the post-loan state representation node corresponding to the first post-loan state is used as the starting point of the directed edge, and the post-loan state representation node corresponding to the second post-loan state is used as the ending point of the directed edge. The migration probability is then set... and migration rate The state transition edges are obtained by writing the edge attributes of the directed edge. Each state transition edge includes at least the starting node identifier, the target node identifier, the transition probability, the transition rate, and the transition time. In this way, the post-loan customer state transition graph can not only record the borrower's current post-loan state, but also represent the probability and speed of its transition from the current state to the overdue state, the out-of-contact state, or the reverse communication state.

[0035] S2: Based on the post-loan customer status migration map, identify the risk level of a borrower's migration from a normal repayment status to an overdue status, a lost contact status, or a reverse communication status, and generate the corresponding post-loan early warning level. In a specific implementation, the borrower's risk level is identified based on the post-loan customer status migration map, and a corresponding post-loan early warning level is generated. The specific process includes the following: First, the directed edge attributes corresponding to the borrower's current post-loan status node are extracted from the post-loan customer status migration graph. These attributes include the migration probability and migration rate from a normal repayment status to an overdue status, an unreachable status, or a reversal of communication status. The migration probability represents the likelihood of the borrower entering a target risk status, while the migration rate represents the speed at which the borrower's status deteriorates. To standardize the calculation, the migration probabilities of the borrower moving to overdue, unreachable, and reversal of communication status are combined to obtain the comprehensive migration probability. The corresponding migration rates are combined to obtain the overall migration rate. The overall migration probability and overall migration rate can be determined by weighting according to the importance of the target risk status, or the migration probability and migration rate corresponding to the migration direction with the highest risk level can be selected as the basis for calculation.

[0036] Subsequently, based on the comprehensive migration probability Overall migration rate and current repayment willingness characteristic value Calculate the borrower's risk level : ,in, This indicates the borrower's risk level. Indicates the overall migration probability. Indicates the overall migration rate. This represents the characteristic value of the current willingness to repay. Let each represent a corresponding weight, and satisfy the following condition: In this embodiment, it can be The values ​​are set to 0.4, 0.3, and 0.3 respectively. The rationale for these settings is as follows: the migration probability directly reflects the likelihood of a borrower entering a risky state, and therefore is given a high weight; the migration rate reflects the urgency of risk deterioration, and the repayment willingness characteristic value reflects the borrower's subjective cooperation level. Both are used together to adjust the risk level.

[0037] After obtaining the risk level Then, it is compared with a preset early warning threshold to generate a corresponding post-loan early warning level. Specifically, when At 0.35, a low-risk warning level is generated; when When a medium-risk warning level is generated; when At that time, a high-risk warning level is generated. A threshold of 0.35 is used to distinguish between normal fluctuations and risk changes requiring attention, while 0.65 is used to distinguish between general risk and significantly worsening risk. This threshold division categorizes post-loan customer status into three warning ranges: low, medium, and high.

[0038] Finally, the generated post-loan warning level is written into the post-loan customer status transition graph of the corresponding borrower, serving as the warning attribute of the current post-loan status node or the corresponding borrower graph, thus obtaining the post-loan customer status transition graph carrying the post-loan warning level, providing an input basis for the generation of subsequent simulated collection actions.

[0039] S3: Input the post-loan customer status transition map with post-loan early warning level into the digital twin simulation engine, generate simulated collection actions through the first AI agent, and simulate the borrower's feedback to the simulated collection actions through the second AI agent to obtain the action-feedback trajectory. In the implementation method, the simulated debt collection actions are generated by a first AI agent, specifically including the following process: First, the first AI agent reads the post-loan customer status migration map carrying the post-loan warning level, and obtains the borrower's post-loan warning level, current post-loan status representation, migration probability, and migration rate. Among them, the post-loan warning level is used to determine the basic intensity range of collection actions, the current post-loan status representation is used to determine the type of status the borrower is currently in, the migration probability is used to represent the probability that the borrower will enter a state of delinquency, loss of contact, or reverse communication, and the migration rate is used to represent the speed at which the borrower's status deteriorates.

[0040] Then, the first AI agent determines the candidate intensity range of collection actions based on the post-loan warning level. In this implementation, the intensity of collection actions is divided into levels 1 to 5, where level 1 represents low-intensity reminders, level 2 represents regular repayment reminders, level 3 represents negotiation-based outreach, level 4 represents focused follow-up outreach, and level 5 represents high-risk intensive outreach. Correspondingly, low-risk warning levels match the candidate intensity range of levels 1 to 2, medium-risk warning levels match the candidate intensity range of levels 2 to 4, and high-risk warning levels match the candidate intensity range of levels 3 to 5. This avoids configuring excessively high-intensity collection actions for low-risk customers and also avoids configuring only low-intensity reminder actions for high-risk customers.

[0041] Meanwhile, the first AI agent determines candidate channel types based on the current post-loan status. Specifically, if the current post-loan status indicates normal repayment or mild delinquency, candidate channel types include SMS, app messages, and automated voice reminders; if the current post-loan status indicates delinquency and the contact response characteristic value is within reach, candidate channel types include SMS, phone calls, and app messages; if the current post-loan status indicates being out of contact, candidate channel types prioritize historically effective channels and authorized backup contact channels; if the current post-loan status indicates a reverse communication state, candidate channel types prioritize low-irritation text channels or negotiation-based communication channels.

[0042] After determining the candidate intensity range and candidate channel type, the first AI agent generates simulated collection actions within these ranges. Each simulated collection action includes at least the action intensity level, the contact channel, the contact time, and the psychological appeal category. The psychological appeal category indicates the communication direction when subsequent collection scripts are generated, and can include performance reminders, repayment convenience prompts, negotiation support, and credit impact warnings. To facilitate subsequent feedback simulations by the second AI agent, the second AI agent can... A simulated collection action is represented as follows: ,in, Indicates the first A simulated debt collection action, Indicates the level of intensity of the action. Indicates the channels reached. Indicates the time of arrival. This indicates the category of psychological needs. Through the above method, the simulated collection actions output by the first AI agent are not generated by a single fixed rule, but are simultaneously constrained by the post-loan warning level, the current post-loan status representation, the migration probability, and the migration rate. The generated simulated collection actions serve as input for the subsequent simulated borrower feedback results of the second AI agent, forming an action-feedback trajectory and further supporting the selection of the target collection path.

[0043] The first AI agent is a strategy-based agent used to generate simulated collection actions. It includes a graph feature encoding layer, a state fusion layer, a strategy generation layer, and a compliance constraint output layer. Specifically, the graph feature encoding layer reads the current post-loan state representation, migration probability, and migration rate from the post-loan customer state transition graph and encodes them into a graph state vector. The state fusion layer fuses the graph state vector with the post-loan early warning level to obtain a strategy decision vector. The strategy generation layer generates candidate simulated collection actions based on the strategy decision vector. The compliance constraint output layer verifies the intensity range, contact channels, and contact time of the candidate simulated collection actions and outputs simulated collection actions that meet the constraints.

[0044] In this implementation, the action-feedback trajectory is obtained by simulating the borrower's feedback to simulated debt collection actions through a second AI agent. The specific process includes the following: First, the second AI agent reads the borrower's current repayment ability characteristic value, repayment willingness characteristic value, contact response characteristic value, and the simulated collection actions output by the first AI agent. The repayment ability characteristic value indicates whether the borrower has the actual conditions to repay; the repayment willingness characteristic value indicates whether the borrower has the intention to actively fulfill the obligation or negotiate; and the contact response characteristic value indicates whether the borrower can be effectively reached. The simulated collection actions include the action intensity level, contact channel, contact time, and psychological appeal category. To facilitate feedback simulation by the second AI agent, the k-th simulated collection action can be represented as: ,in, Indicates the first A simulated debt collection action, Indicates the level of intensity of the action. Indicates the channels reached. Indicates the time of arrival. This indicates the category of psychological needs. The second AI agent compares the simulated debt collection actions with the borrower's repayment ability characteristics. Intention to repay characteristic value and contact response characteristic values Using these as inputs, the study simulates the impact of this action on the borrower's willingness to repay and their contact response status.

[0045] Then, the second AI agent outputs corresponding feedback results. These feedback results include repayment feedback, contact feedback, and negative feedback. Repayment feedback indicates whether the borrower has made a repayment, partial repayment, promised repayment, or negotiated repayment after the simulated collection actions; contact feedback indicates whether the borrower has answered, responded, delayed response, or not responded after the simulated collection actions; negative feedback indicates whether the borrower has shown a tendency to refuse communication, complain, become emotionally resistant, or further become unreachable. In the implementation, the second AI agent can be... The feedback results corresponding to each simulated collection action are represented as follows: ,in, Indicates the first Feedback results corresponding to each simulated debt collection action. This indicates feedback on repayment. Indicates contacting and providing feedback. Negative feedback is represented by classification labels or probability values. For example, when a borrower has high repayment ability, high willingness to repay, and high contact response values, the second AI agent, upon receiving low- or medium-intensity simulated collection actions, tends to output feedback results indicating connection or response, willingness to negotiate, and low negative feedback. Conversely, when a borrower has low repayment ability, low willingness to repay, and declining contact response values, the second AI agent, upon receiving high-intensity simulated collection actions, tends to output feedback results indicating no response, refusal to communicate, or increased negative feedback. Therefore, it is possible to simulate and evaluate the potential customer reactions to different collection actions before actual collection actions are executed.

[0046] To facilitate a continuous simulation process, the second AI agent, after outputting each feedback result, binds that result to the corresponding simulated collection action and records it in chronological order according to the simulation time, forming an action-feedback record. An action-feedback log can be represented as: ,in, Indicates the first Action-Feedback Log This indicates a simulated debt collection action. This indicates the corresponding feedback result. This indicates the simulation time corresponding to the action-feedback record. After multiple simulated collection actions and their corresponding feedback results are generated consecutively, the action-feedback records are arranged in chronological order to form an action-feedback trajectory. ,in, Indicates the action-feedback trajectory. This represents the total number of action-feedback records generated during the simulation. Indicates the first Action-feedback records are generated. Action-feedback trajectories represent the changes in borrower repayment feedback, contact feedback, and negative feedback over time under the continuous application of different collection actions. Through this method, the second AI agent can correlate the borrower's post-loan status characteristics with simulated collection actions, outputting feedback results that can be used for strategy deduction, and forming action-feedback trajectories.

[0047] The inputs to the aforementioned first AI agent include post-loan early warning level, current post-loan status representation, migration probability, migration rate, and historical feedback results; the output is structured simulated collection actions, including action intensity level, contact channels, contact time, and psychological appeal category. In other words, the first AI agent in this solution does not directly output natural language collection scripts, but first outputs action parameters for simulation and deduction. Subsequently, S5 generates corresponding collection script text and communication channel arrangement instructions based on the target collection path. The first AI agent can be trained using historical post-loan collection samples. Specifically, the post-loan status representation, warning level, state transition information, and executed collection actions from historical post-loan collection samples are used as training inputs. Repayment feedback, contact feedback, and negative feedback following the corresponding collection actions are used as training results to construct an action effect sample set. First, the action effect sample set is used to conduct supervised training on the first AI agent, enabling it to learn the basic collection action selection rules under different post-loan statuses and warning levels. Then, the supervised-trained first AI agent is connected to a digital twin simulation engine for strategy iteration, updating its parameters in directions that simulate collection actions can improve repayment target achievement rates, reduce customer disturbance costs, and meet compliance constraints. During strategy iteration, a reward function can be set to evaluate the output actions of the first AI agent. ,in, Indicates the strategy reward value. This indicates that the repayment target has been achieved. This represents customer disturbance costs. This indicates compliance and non-compliance penalties. These represent the corresponding weights. This training method ensures that the simulated debt collection actions generated by the first AI agent can both match the borrower's post-loan risk status and avoid directly employing overly intensive or non-compliant debt collection actions.

[0048] The aforementioned second AI agent is a customer response simulation agent, comprising a customer feature encoding layer, an action impact analysis layer, a feedback prediction layer, and a state update layer. The customer feature encoding layer encodes repayment ability, repayment willingness, and contact response features into a customer state vector. The action impact analysis layer analyzes the impact of simulated collection actions on the customer state vector. The feedback prediction layer outputs repayment feedback, contact feedback, and negative feedback. The state update layer updates the borrower's repayment willingness and contact response states based on the feedback results, enabling subsequent simulated collection actions to continue based on the updated customer state. The second AI agent is trained using historical post-loan collection feedback samples. Specifically, the repayment ability, repayment willingness, and contact response features from historical post-loan collection samples, along with executed collection actions, are used as training inputs. The actual repayment results, contact results, and negative feedback results after the collection action are used as training labels to train the second AI agent, allowing it to learn the feedback patterns of different customer states under different collection actions. After training, the second AI agent can simulate the borrower's feedback to simulated debt collection actions in the digital twin simulation engine.

[0049] S4: Filter feasible paths from the action-feedback trajectory to achieve the preset repayment target and meet compliance constraints, and determine the customer disturbance cost based on the collection intensity and feedback results corresponding to the feasible path. The feasible path with the lowest customer disturbance cost is determined as the target collection path. In the implementation method, feasible paths that achieve the preset repayment target and meet compliance constraints are selected from the action-feedback trajectory, specifically including the following process: First, candidate paths are extracted from the action-feedback trajectory generated by the second AI agent. The action-feedback trajectory is formed by arranging multiple action-feedback records in simulation time sequence. Each action-feedback record includes a simulated collection action, the corresponding feedback result, and the simulation time. Simultaneously, candidate paths consisting of multiple consecutive action-feedback records are extracted from the action-feedback trajectory according to a preset path length or a preset repayment period. Specifically, if the action-feedback trajectory is represented as: Then any candidate path can be represented as: ,in, Indicates the first Candidate paths, This indicates the initial action – feedback record – of the candidate path. This indicates the termination action - feedback record for the candidate path, and .

[0050] Then, it is determined whether the candidate path has reached the preset repayment target within the preset repayment period. In this embodiment, the preset repayment period can be set to 7 calendar days from the time the post-loan warning level is generated, or it can be adjusted according to the post-loan warning level; for example, a low-risk warning level corresponds to 7 calendar days, a medium-risk warning level corresponds to 5 calendar days, and a high-risk warning level corresponds to 3 calendar days. This setting is based on the fact that the higher the warning level, the faster the customer's condition deteriorates, and the collection strategy needs to complete the effectiveness verification within a shorter period. The preset repayment target includes full repayment, partial repayment reaching a preset proportion, or the formation of a valid repayment commitment. In some embodiments, when the feedback results corresponding to the candidate path show full repayment feedback, or the cumulative repayment amount reaches more than 30% of the amount due, or the borrower generates a valid repayment commitment with a clear amount and a clear date, the candidate path is deemed to have reached the preset repayment target. The judgment condition can be expressed as: ,in, Indicates the first The cumulative repayment amount corresponding to each candidate path This indicates the amount the borrower currently owes. This represents the preset repayment ratio threshold, which can be 0.3. When a candidate path meets the above conditions, or when the feedback results show sufficient repayment or a valid repayment commitment, the candidate path is selected as the initial path.

[0051] Next, compliance verification is performed on the initially selected path. Compliance constraints include at least the following: contact frequency constraints, contact time constraints, channel authorization constraints, and script content constraints. Contact frequency constraints limit the number of times a single borrower can be contacted within a given time period, preventing excessive repeated contact within the same cycle. Contact time constraints limit the execution time of simulated collection actions, preventing contact with borrowers outside permitted hours. Channel authorization constraints determine whether the contact channel is authorized by the borrower or permitted by the post-loan system. Script content constraints determine whether the psychological appeal category or script template identifier corresponding to the simulated collection actions conforms to preset compliance rules. In some implementations, the frequency of contact can be constrained by no more than 3 times per day and no more than 10 times within 7 consecutive days for the same borrower; the time of contact can be constrained by the contact time being between 9:00 and 20:00; the channel authorization constraint can be constrained by allowing only channels authorized in the loan agreement, post-loan management agreement, or system records, such as SMS, telephone, and application messages; and the content of the communication script can be constrained by prohibiting the inclusion of threatening, misleading, false legal consequences, or irrelevant third-party disclosures. For any initial selection path, if each simulated collection action it contains meets the above compliance constraints, then the initial selection path is determined as a feasible path; if any simulated collection action does not meet any of the constraints on contact frequency, contact time, channel authorization, or communication script content, then the initial selection path is eliminated and will not be included in the subsequent customer disturbance cost calculation process. As can be seen from the above method, feasible paths are not simply selected based on repayment results, but are formed after further compliance verification based on achieving the preset repayment target. Therefore, when determining the target collection path, a constraint balance can be formed between effective collection and compliant reach, providing a reliable basis for selecting the target collection path with the lowest disturbance cost for the customer.

[0052] In a specific implementation, the customer disturbance cost is determined based on the collection intensity and feedback results corresponding to the feasible paths, and the feasible path with the lowest customer disturbance cost is determined as the target collection path. This process includes the following steps: First, for each feasible path selected, extract the action intensity level, contact frequency, contact channel, and feedback result corresponding to each simulated collection action in that feasible path.

[0053] Then, based on the action intensity level of each simulated collection action, the channel sensitivity corresponding to the contact channel, and the degree of negative feedback in the feedback results, the single disturbance value corresponding to that simulated collection action is calculated. Specifically, it can be determined using the following formula: ,in, Indicates the first The single perturbation value corresponding to each simulated collection action. Indicates the level of intensity of the action. This indicates the channel sensitivity corresponding to the reach channel. This indicates the degree of negative feedback in the feedback results. Let each represent a corresponding weight, and satisfy the following condition: In this embodiment, These can be set to 0.4, 0.2, and 0.4 respectively. The rationale is that the action intensity level and the degree of negative feedback directly affect customer experience and the stability of subsequent communication, therefore they have higher weights; channel sensitivity is mainly used to correct the degree of disturbance to customers from different outreach channels, therefore it is given a relatively lower but necessary weight. Among these, the action intensity level... Normalization can be performed according to levels 1 to 5, for example: ,in, This represents the normalized motion intensity value. This represents the original intensity level of the action. To standardize calculation methods, the value in the above formula... Alternatively, the normalized version can be used directly. Involved in disturbance value calculation. Channel sensitivity corresponding to the reached channel. The sensitivity of different channels to borrowers can be preset. For example, the sensitivity of application messages can be set to 0.2, SMS messages to 0.3, automated voice messages to 0.5, and human telephone calls to 0.7. This is based on the fact that asynchronous text-based outreach has a lower immediate impact on borrowers, while real-time telephone outreach has a relatively higher level of interference. (Negative feedback level) The negative feedback level can be determined based on the negative feedback output by the second AI agent. For example, a value of 0 is used when there is no negative feedback, 0.3 is used when there is a delayed response or slight rejection, 0.6 is used when there is a refusal to communicate, and 0.9 is used when there is a tendency to complain or further loss of contact. The degree of negative feedback is used to characterize the extent to which the simulated collection action has an adverse impact on the borrower's subsequent cooperation.

[0054] After determining the single-period disturbance value for each simulated collection action, and combining this with the reach frequency within the feasible path, the number of... Customer disturbance costs of a feasible path : ,in, Indicates the first Customer disturbance costs for a feasible path This indicates the number of simulated collection actions in the feasible path. Indicates the first The single perturbation value of a simulated collection action. This indicates the frequency of contact within the preset repayment period for the feasible path. This indicates the weighting of the reach frequency adjustment. By introducing a reach frequency adjustment term, we can prevent a situation where a single action on a particular path may have low intensity, but the overall level of disturbance can increase due to excessive reach. After obtaining the customer disturbance cost for each feasible path, we compare the customer disturbance costs corresponding to all feasible paths and select the feasible path with the lowest customer disturbance cost as the target collection path. ,in, Indicates the target collection path. Represents the set of all feasible paths. Indicates the first The customer disturbance cost corresponding to each feasible path is considered. If two or more feasible paths have the same customer disturbance cost, their repayment target achievement time can be further compared, and the feasible path with the shorter repayment target achievement time is determined as the target collection path. If the repayment target achievement time is still the same, the feasible path with the lower average action intensity level is selected as the target collection path. This ensures that the target collection path prioritizes collection strategies with low disruption and low negative feedback, while meeting repayment targets and compliance constraints. Through the above method, customer disturbance cost can unify action intensity, contact frequency, channel sensitivity, and negative feedback level into a single evaluation indicator, so that the determination of the target collection path no longer depends on a single repayment result, but takes into account repayment effectiveness, customer experience, and compliance risks.

[0055] S5: Convert the target collection path into a sequence of collection actions arranged in chronological order, and generate collection script text and communication channel arrangement instructions based on the intensity level, channel type and psychological appeal category of each collection action to obtain a differentiated collection plan.

[0056] In a specific implementation, the collection script text and communication channel arrangement instructions are generated to obtain a differentiated collection plan, which specifically includes the following process: First, the determined target collection path is broken down chronologically into multiple collection action nodes. Each collection action node corresponds to a simulated collection action within the target collection path, and carries the execution order, intensity level, channel type, and psychological appeal category for that simulated collection action. The execution order determines the sequential relationship between different collection actions; the intensity level determines the level of alertness; the channel type determines whether to use SMS, telephone, app messages, or other authorized channels; and the psychological appeal category determines the direction of the communication. For example, if the target collection path includes three actions in sequence: low-intensity app message alert, medium-intensity SMS negotiation prompt, and telephone negotiation contact, it can be broken down into three collection action nodes, each configured with its own execution order, intensity level, channel type, and psychological appeal category. Through this breakdown, the target collection path is transformed from a complete path into a set of collection action nodes that can be executed sequentially.

[0057] Then, based on the psychological appeal category corresponding to each collection action node, a corresponding script template is matched from a pre-set script template library. This library pre-stores script templates corresponding to different psychological appeal categories, which may include repayment reminders, repayment convenience tips, negotiation support, and credit impact warnings. Each script template has replaceable semantic slots, which at least include the borrower's title, amount due, repayment date, repayment channel, negotiation entry point, warning level information, and current post-loan status information. After matching the script template, the semantic slots in the script template are filled in based on the borrower's post-loan warning level and current post-loan status, generating the collection script text. Specifically, when the borrower is at a low-risk warning level, the collection script mainly focuses on repayment reminders and convenient repayment suggestions; when the borrower is at a medium-risk warning level, the collection script adds reminders for negotiated repayment, installment communication, or proactive contact; when the borrower is at a high-risk warning level, the collection script, while maintaining compliant expression, emphasizes timely communication, preventing further deterioration of the situation, and providing an entry point for negotiation. This ensures that the collection script is matched to the borrower's risk status and the intensity of collection actions.

[0058] Next, communication channel orchestration instructions are generated based on the execution order, channel type, and intensity level of each collection action node. These instructions instruct the post-loan management system to execute the corresponding collection actions according to a specified time, channel, and order. Each instruction includes at least the borrower identifier, action node identifier, execution time, contact channel, content identifier, trigger condition, and stop condition. The execution time is determined based on the temporal relationship in the target collection path; the contact channel is determined based on the channel type of the collection action node; the content identifier is used to associate with the corresponding collection script text; the trigger condition determines whether the current collection action node needs to be executed; and the stop condition stops subsequent actions when the borrower completes repayment, forms a valid repayment commitment, or the status changes. In this implementation, after the previous collection action node is executed, if the feedback result shows that the borrower has completed repayment or formed a valid repayment commitment, then subsequent higher-intensity collection action nodes will not be triggered; if the feedback result shows that the borrower has not responded and is still within the preset repayment period, then subsequent collection action nodes will be triggered according to the next execution order in the target collection path. This approach allows for the sequential execution of differentiated collection plans based on borrowers' feedback, avoiding further contact with borrowers even after repayment targets have been met.

[0059] Finally, each collection action node, its corresponding collection script, and communication channel arrangement instructions are linked to form a differentiated collection plan. A differentiated collection plan includes at least execution time, contact channels, collection script, action intensity level, psychological appeal category, triggering conditions, and stopping conditions. This differentiated collection plan serves as the execution input for the post-loan management system, guiding subsequent collection actions to be implemented according to the target collection path. Through this method, the target collection path can be transformed into a differentiated collection plan with a clear execution sequence, contact channels, script content, and triggering rules, ultimately translating the aforementioned post-loan early warning, agent simulation, and customer disturbance cost screening results into an executable collection strategy.

[0060] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0061] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for generating post-loan early warning and differentiated collection plans based on AI intelligent agents, characterized in that, Includes the following steps: S1: Obtain the borrower's post-loan behavior data, determine the corresponding repayment ability characteristic value, repayment willingness characteristic value and contact response characteristic value, and form a post-loan customer state migration map with the borrower's post-loan status representation as nodes and the migration probability and migration rate between post-loan states as directed edges. S2: Based on the post-loan customer status migration map, identify the risk level of a borrower's migration from a normal repayment status to an overdue status, a lost contact status, or a reverse communication status, and generate the corresponding post-loan early warning level. S3: Input the post-loan customer status transition map carrying the post-loan early warning level into the digital twin simulation engine, generate simulated collection actions through the first AI agent, and simulate the borrower's feedback to the simulated collection actions through the second AI agent to obtain the action-feedback trajectory. S4: Filter feasible paths from the action-feedback trajectory to achieve the preset repayment target and meet compliance constraints, and determine the customer disturbance cost based on the collection intensity and feedback results corresponding to the feasible path. The feasible path with the lowest customer disturbance cost is determined as the target collection path. S5: Convert the target collection path into a sequence of collection actions arranged in chronological order, and generate collection script text and communication channel arrangement instructions based on the intensity level, channel type and psychological appeal category of each collection action to obtain a differentiated collection plan.

2. The method for generating post-loan early warning and differentiated collection schemes based on AI intelligent agents according to claim 1, characterized in that, The corresponding repayment ability characteristic value, repayment willingness characteristic value, and contact response characteristic value determined in S1 include: S11: Obtain the borrower's historical repayment records, overdue days, account fund fluctuation records, contact response records, and repayment commitment fulfillment records to form a post-loan behavior dataset; S12: Determine the repayment ability characteristic value to characterize changes in the borrower's repayment ability based on historical repayment records, overdue days, and account fund fluctuation records; S13: Based on the records of voluntary repayment and extension negotiation in the repayment commitment performance record, determine the repayment willingness feature value used to characterize the borrower's tendency to voluntarily repay; S14: Determine the contact response feature value used to characterize the borrower's reachability based on the connection status, response delay, number of rejections, and duration of disconnection in the contact response record.

3. The method for generating post-loan early warning and differentiated collection schemes based on AI intelligent agents according to claim 2, characterized in that, The specific components of using the borrower's post-loan status as a node include: Based on the repayment ability characteristic value, repayment willingness characteristic value, and contact response characteristic value, the borrower's post-loan status representation corresponding to the current post-loan stage is determined; The post-loan status is categorized into normal repayment status, overdue status, out-of-contact status, and reverse communication status. Each post-loan status is represented as a node in the post-loan customer status transition graph, and each node is configured with a corresponding status time label, repayment ability feature value, repayment willingness feature value, and contact response feature value.

4. The method for generating a post-loan early warning and differentiated collection plan based on an AI intelligent agent according to claim 3, characterized in that, The process of forming a post-loan customer state transition graph using the transition probabilities and rates between post-loan states as directed edges includes: Obtain the borrower's status transition records between adjacent post-loan states, and count the number of transitions from the first post-loan state to the second post-loan state; The migration probability from the first post-loan state to the second post-loan state is determined based on the number of transfers and the total number of occurrences corresponding to the first post-loan state. Based on the difference in state duration between the first and second post-loan states, and the magnitude of change in the corresponding eigenvalues, the migration rate from the first post-loan state to the second post-loan state is determined. Finally, using migration probability and migration rate as directed edge attributes, the corresponding post-loan state representation nodes are connected to form the post-loan customer state migration graph.

5. The method for generating post-loan early warning and differentiated collection schemes based on AI intelligent agents according to claim 1, characterized in that, S2 specifically includes: S21: Extract the migration probability and migration rate of borrowers from normal repayment status to overdue status, out-of-contact status, or reverse communication status from the post-loan customer status migration map. S22: Determine the borrower's risk level based on migration probability, migration rate, and current repayment willingness characteristic values; S23: Compare the risk level with the preset warning threshold, and generate a low-risk warning level, a medium-risk warning level, or a high-risk warning level according to the comparison result; S24: Write the post-loan warning level into the post-loan customer status transition map of the corresponding borrower to obtain the post-loan customer status transition map with the post-loan warning level.

6. The method for generating post-loan early warning and differentiated collection schemes based on AI intelligent agents according to claim 1, characterized in that, The simulated debt collection actions generated by the first AI agent in S3 include: S31: The first AI agent reads the post-loan warning level, the current post-loan status representation, the migration probability, and the migration rate; S32: Determine the candidate intensity range of collection actions based on the post-loan early warning level, and determine the corresponding candidate channel type based on the current post-loan status. S33: Generate simulated collection actions within the candidate intensity range and candidate channel type, wherein the simulated collection actions include action intensity level, contact channel, contact time, and psychological appeal category.

7. The method for generating post-loan early warning and differentiated collection schemes based on AI intelligent agents according to claim 6, characterized in that, In step S3, the second AI agent simulates the borrower's feedback to the simulated debt collection actions, resulting in an action-feedback trajectory including: S34: The second AI agent reads the repayment ability feature value, repayment willingness feature value, contact response feature value, and simulates collection actions; S35: Based on the impact of the simulated collection actions on the borrower's willingness to repay and the contact response status, output the corresponding feedback results, including repayment feedback, contact feedback and negative feedback; S36: Record each simulated collection action and its corresponding feedback result in sequence according to the simulation time to form the action-feedback trajectory.

8. The method for generating post-loan early warning and differentiated collection schemes based on AI intelligent agents according to claim 1, characterized in that, The feasible paths in S4 that achieve the preset repayment target and meet compliance constraints from the action-feedback trajectory include: S41: Extract candidate paths from the action-feedback trajectory, which consist of multiple simulated collection actions and corresponding feedback results; S42: Determine whether the candidate path has reached the preset repayment target within the preset repayment period, and take the candidate path that has reached the preset repayment target as the initial selection path; S43: Perform compliance verification on the initial selected path, the compliance constraints including contact frequency constraints, contact time constraints, channel authorization constraints, and script content constraints; determine the initial selected path that passes the compliance verification as the feasible path.

9. The method for generating post-loan early warning and differentiated collection schemes based on AI intelligent agents according to claim 8, characterized in that, In S4, the feasible path with the lowest customer disturbance cost is identified as the target collection path, including: S45: Extract the action intensity level, contact frequency, contact channel, and feedback result corresponding to each simulated collection action in each feasible path; S46: Determine the customer disturbance cost for each feasible path based on the action intensity level, reach frequency, channel sensitivity corresponding to the reach channel, and degree of negative feedback in the feedback results; S47: Compare the customer disturbance costs corresponding to each feasible path, and determine the feasible path with the lowest customer disturbance cost as the target collection path.

10. The method for generating a post-loan early warning and differentiated collection plan based on an AI intelligent agent according to claim 1, characterized in that, S5 specifically includes: S51: Divide the target collection path into multiple collection action nodes according to the time sequence, and determine the execution order, intensity level, channel type and psychological appeal category corresponding to each collection action node; S52: Match the corresponding script template from the preset script template library according to the category of psychological needs, and generate collection script text by combining the borrower's post-loan warning level and current post-loan status. S53: Generate communication channel orchestration instructions based on execution order, channel type, and intensity level; S54: Associate collection action nodes, collection script texts, and communication channel arrangement instructions to obtain differentiated collection solutions that include execution time, contact channels, collection script texts, and triggering conditions.