An artificial intelligence-based customer management data processing system and method

Abstract

The present application relates to the technical field of customer management data processing, in particular to a customer management data processing system and method based on artificial intelligence, which generates candidate customer management strategies and action units through an artificial intelligence model, deduces the influence range and version of a graph, identifies instruction pollution and calculates the risk of persistent pollution, implements authorization and generates a token according to the influence radius and risk index, performs secondary verification when an execution request arrives, and outputs a rejection, freezing or resubmission of a manual review instruction for a request that does not pass, the present application generates candidate customer management strategies and action units through an artificial intelligence model, deduces the influence radius and version summary of a customer relationship graph, identifies suspected instruction pollution and calculates the risk index of persistent pollution, and then generates a customer management action authorization token and performs secondary verification of an execution request, thereby improving boundary identification, pollution prevention and control, and execution security.

Description

Technical Field

[0001] This invention relates to the field of customer management data processing technology, and more specifically, to a customer management data processing system and method based on artificial intelligence. Background Technology

[0002] As customer service, business marketing, contract management, after-sales processing, and customer relationship maintenance become increasingly digitized, a large amount of customer interaction data and basic customer management data has accumulated during the customer management process. Customer interaction data typically originates from online conversations, emails, service tickets, telephone records, and business notes. Basic customer management data typically includes customer object data, business-related data, business status data, and authorization status data. Existing customer management technologies can leverage this data to complete customer information collection, business process flow, management strategy recommendations, and the automated execution of some operations, playing a positive role in improving data processing efficiency and business response speed.

[0003] When AI models are applied to customer management scenarios, they can perform semantic analysis on fragmented customer interactions and generate customer management strategies by combining them with basic customer management data. However, the management strategies output by AI models typically include multiple management actions that may trigger data reading, writing, process activation, or API calls. The impact of these management actions on customer data is not limited to a single customer record. For example, adjusting customer service levels may be linked to multiple data points such as customer contracts, pricing rules, credit status, service processes, and approval permissions. Existing processing methods usually focus on determining whether a single action has execution authority. They lack a unified quantitative analysis method for the indirect impact of management actions along customer relationships, business relationships, and process dependencies, making it difficult to accurately determine the actual impact boundaries before strategy execution.

[0004] Furthermore, customer interaction data is a crucial input for AI model-generated management strategies. This data may contain genuine customer business needs, but it can also include content with a tendency to interfere, such as requests to ignore approval conditions, alter permission judgments, expand the scope of write operations, or specify abnormal interface call methods. Such content may originate from unverified customer descriptions, transcription records, historical work order references, or multi-layered data forwarding. If this content is incorporated into candidate customer management strategies and further written into basic customer management data, its impact may continue to spread through data replication, business references, and process flows. Existing security controls typically focus on the identity of the access subject, role permissions, sensitive fields, and individual interface requests. However, there is a lack of interconnected risk assessment methods regarding the semantic interference level of abnormal instruction content, the credibility of the data source, the stability of data retention after writing, the scope of propagation, and the difficulty of rollback.

[0005] When a customer management strategy enters the execution phase, its execution is typically determined by permission checks or fixed approval rules established during strategy generation. However, within the time interval between strategy generation and action execution, customer business relationships, authorization status, risk status, or manual review results may have changed. Furthermore, the action fields, write fields, or API call scope in the actual execution request may differ from the original authorization content. Relying solely on a single authorization during strategy generation is insufficient to cover version changes, risk shifts, and execution content deviations. Therefore, how to implement tiered authorization by considering the relationship impact scope of candidate customer management actions and the risk of persistent contamination of interaction data, and how to re-verify the authorization basis, version status, execution scope, and manual review status upon arrival of downstream execution requests, becomes a technical challenge that needs to be addressed when artificial intelligence participates in customer management data processing. Summary of the Invention

[0006] In view of the shortcomings of existing technologies, the purpose of this invention is to provide a customer management data processing system and method based on artificial intelligence.

[0007] To achieve the above objectives, the present invention provides the following technical solution: An artificial intelligence-based customer management data processing system includes: The strategy generation module is used to generate candidate customer management strategies based on customer interaction data and basic customer management data through an artificial intelligence model, and to determine the candidate customer management action units corresponding to the candidate customer management strategies. The graph inference module is used to infer the graph impact of candidate customer management action units and generate the customer relationship graph impact radius and customer relationship graph version summary. The contamination simulation module is used to identify suspected contaminated instruction fragments based on customer interaction data and candidate customer management action units; generate persistent contamination propagation chains for suspected contaminated instruction fragments and determine contamination assessment characteristics; calculate persistent contamination risk index based on persistent contamination propagation chains and contamination assessment characteristics; and generate a persistent contamination propagation chain version summary based on persistent contamination propagation chains; wherein, the contamination assessment characteristics include source credibility characteristics; The token control module is used to authorize and control the management strategies of candidate customers based on the influence radius of the customer relationship graph and the persistent pollution risk index, and to generate authorization tokens for customer management actions. The execution verification module is used to perform a second authorization verification on the customer management action authorization token before execution when it receives an execution request from the downstream customer management execution system. If the second authorization verification fails, it outputs an execution control instruction to the downstream customer management execution system. The execution control instruction includes a rejection instruction, a freeze instruction, or a resubmission instruction for manual review.

[0008] In one embodiment, the customer management basic data includes customer object data, business association data, business status data, and authorization status data. The candidate customer management strategy includes at least one candidate customer management action unit, which carries an action execution field, a write control field, an interface control field, and a source tracking field.

[0009] In one embodiment, the generation method of the customer relationship graph influence radius and the customer relationship graph version summary includes: constructing a customer relationship graph based on customer management basic data; generating an action influence subgraph in the customer relationship graph based on candidate customer management action units; calculating the customer relationship graph influence radius based on the action influence subgraph; obtaining graph nodes, relationship edges, node attributes, and edge attributes from the customer relationship graph; sorting the graph nodes, relationship edges, node attributes, and edge attributes according to a preset sorting rule; generating a graph state sequence; and performing a hash operation on the graph state sequence to obtain the customer relationship graph version summary.

[0010] In one embodiment, the calculation method of the customer relationship graph influence radius includes: determining the path level, risk weight and time decay coefficient of each action propagation path in the action influence subgraph, calculating the weighted influence value of each action propagation path, and determining the maximum weighted influence value as the customer relationship graph influence radius.

[0011] In one embodiment, the customer relationship graph includes customer management objects and graph structure data corresponding to the relationships between customer management objects. The customer management objects include customer objects, business records, status constraint items and management process items. The action impact subgraph includes the set of action reachable nodes, the set of action propagation paths, the set of risk weights, the set of change rollback costs, and the set of time-decrease coefficients.

[0012] In one embodiment, the contamination assessment features include contamination semantic features, source credibility features, and persistent write features. The contamination semantic features are determined based on the degree of directive intervention by the suspected instruction contamination fragment in customer interaction data, candidate customer management strategies, and the interface call results of the downstream customer management execution system defined by the interface control field. The source credibility features are determined based on the verifiability of the data source corresponding to the suspected instruction contamination fragment and the source propagation level. The persistent write features are determined based on the retention stability, replication range, and rollback difficulty after the suspected instruction contamination fragment is written into the customer management basic data.

[0013] In one embodiment, the persistent contamination risk index is calculated by: determining a contamination semantic risk value based on contamination semantic features, determining a source credibility risk value based on source credibility features, determining a persistent write risk value based on persistent write features, and determining a link correction coefficient based on the persistent contamination propagation chain; weighting and summing the contamination semantic risk value, source credibility risk value, and persistent write risk value according to preset weights, and correcting the weighted summation result according to the link correction coefficient to obtain the persistent contamination risk index corresponding to the suspected instruction contamination fragment.

[0014] In one embodiment, the token control module generates customer management action authorization tokens in the following manner: it generates a joint risk control value based on the customer relationship graph influence radius, persistent contamination risk index and source credibility characteristics, determines the execution control result based on the joint risk control value, and generates customer management action authorization tokens based on the execution control result; The execution control results include a set of dynamic authorization thresholds, execution levels, and write methods. Execution levels include automatic execution level, draft generation level, manual review level, frozen execution level, and prohibited execution level. Write methods include normal write, source-restricted write, and isolated write.

[0015] In one embodiment, the customer management action authorization token carries action constraint information, write constraint information, interface constraint information, risk constraint information, version constraint information, validity verification information, execution level, and manual review status. Among them, action constraint information limits the action execution fields and scope of influence; write constraint information limits the write control fields and scope of write; interface constraint information limits the interface control fields; risk constraint information limits the influence radius of the customer relationship graph and the persistent contamination risk index; version constraint information limits the version summary of the customer relationship graph and the version summary of the persistent contamination propagation chain; validity verification information is used to verify the validity of the customer management action authorization token; and manual review status includes unreviewed status, reviewed passed status, and reviewed failed status.

[0016] In one embodiment, an artificial intelligence-based customer management data processing method includes: Step 1: Generate candidate customer management strategies based on customer interaction data and basic customer management data using an artificial intelligence model, and determine the candidate customer management action units corresponding to the candidate customer management strategies; Step 2: Perform graph impact analysis on candidate customer management action units to generate customer relationship graph impact radius and customer relationship graph version summary; Step 3: Identify suspected instruction contamination fragments based on customer interaction data and candidate customer management action units; Step 4: Generate a persistent contamination propagation chain for suspected instruction contamination fragments and determine contamination assessment features, including reliable source characteristics; Step 5: Calculate the persistent pollution risk index based on the persistent pollution propagation chain and pollution assessment characteristics, and generate a persistent pollution propagation chain version summary based on the persistent pollution propagation chain; Step 6: Based on the influence radius of the customer relationship graph and the persistent contamination risk index, authorize and control the management strategies for candidate customers, and generate authorization tokens for customer management actions; Step 7: Receive the execution request sent by the downstream customer management execution system, perform a second authorization verification on the customer management action authorization token before execution, and output the execution control command if the verification fails. Among them, execution control instructions include instructions to refuse execution, freeze execution instructions, or resubmit for manual review.

[0017] Compared with the prior art, the present invention has the following beneficial effects: This invention uses an artificial intelligence model to generate candidate customer management strategies based on customer interaction data and basic customer management data, and further identifies candidate customer management action units. This ensures that customer management strategies are no longer just general suggestions, but are broken down into action units with clear execution directions. On this basis, the invention performs graph impact deduction on the candidate customer management action units, generating the customer relationship graph influence radius and customer relationship graph version summary. This enables the identification of the impact range of the action on customer objects, business-related data, business status data, and related authorization status before authorization control, improving the accuracy of customer management action boundary identification and version traceability. By identifying suspected contaminated instruction fragments based on customer interaction data and candidate customer management action units, and generating persistent contamination propagation chains for suspected contaminated instruction fragments, and further combining contamination assessment features including source credibility characteristics to calculate a persistent contamination risk index, it is possible to link the risk of abnormal instruction intervention in customer interaction content with the actual execution impact of candidate customer management actions. This facilitates the identification of the continuous propagation risk of contaminated content in strategy generation, action formation, data writing, and process flow, and reduces the possibility that suspected contaminated instruction fragments are mistakenly included in customer management strategies and form a persistent impact. Based on the influence radius of the customer relationship graph and the persistent contamination risk index, authorization control is applied to candidate customer management strategies, and customer management action authorization tokens are generated. This ensures that the execution of candidate customer management strategies is jointly constrained by risk level, scope of influence, and authorization status. Simultaneously, upon receiving an execution request from the downstream customer management execution system, a second authorization verification is performed on the customer management action authorization token before execution. If the verification fails, an execution rejection instruction, a freeze instruction, or a resubmission manual review instruction is output. This avoids situations where execution requests deviate from the original authorization scope or continue to be executed despite version status changes, thereby improving the security, controllability, and execution consistency of the customer management data processing process. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the system architecture of an artificial intelligence-based customer management data processing system; Figure 2 This is a schematic diagram of the pollution simulation module's processing flow. Figure 3 A flowchart illustrating the process of generating authorization tokens for customer management actions for the token control module; Figure 4 A schematic diagram of the structure of authorization tokens for customer management actions. Detailed Implementation

[0019] Example 1: Refer to Figure 1 An artificial intelligence-based customer management data processing system includes: The strategy generation module is used to generate candidate customer management strategies based on customer interaction data and basic customer management data through artificial intelligence models, and to determine the candidate customer management action units corresponding to the candidate customer management strategies.

[0020] Collect customer interaction data and basic customer management data. Customer interaction data includes requests, confirmations, objections, and supplementary explanations generated by customers in telephone communications, online conversations, work order records, and business processing pages. Basic customer management data includes customer object data, business association data, business status data, and authorization status data. Taking enterprise customer renewal scenarios as an example, customer interaction data records customer requests to change contact persons, upgrade service levels, and extend billing periods; customer object data records customer entities, contact persons, and account identifiers; business association data records the relationships between contracts, orders, and service items; business status data records the current renewal stage, overdue payment status, and complaint handling status; and authorization status data records the permissions of the personnel in charge, the scope of customer authorization, and approval records. Under the joint constraints of the above data, the artificial intelligence model generates candidate customer management strategies. These candidate customer management strategies do not directly fall into business execution but form a set of decomposable and verifiable management intentions. At least one candidate customer management action unit is determined according to the candidate customer management strategy. Each candidate customer management action unit carries an action execution field, a write control field, an interface control field, and a source tracking field. The action execution field records the action content such as contact person change, service level adjustment, and billing period change. The write control field records the range of customer object data, business-related data, business status data, or authorized status data that can be written. The interface control field records the customer management execution interface, approval interface, or record query interface involved in the action. The source tracking field records the customer interaction data fragments and the source of customer management basic data corresponding to the candidate customer management action unit. When a candidate customer management strategy includes multiple candidate customer management action units, a single candidate customer management action unit is used as the minimum authorized object, and a customer management action authorization token is generated for each. Each customer management action authorization token corresponds to only one candidate customer management action unit, and the action execution field, write control field, interface control field, and source tracking field of that candidate customer management action unit are recorded. When multiple candidate customer management action units belong to the same candidate customer management strategy, the customer relationship graph influence radius and persistent contamination risk index of each candidate customer management action unit are calculated separately. The overall execution level of the candidate customer management strategy is determined according to the execution level with the highest risk constraint strength among all candidate customer management action units. The risk constraint strength of the execution level increases sequentially in the following order: automatic execution level, draft generation level, manual review level, frozen execution level, and prohibited execution level.

[0021] For example, for the candidate customer management action unit for improving service level, the action execution field records the service level adjustment matters, the write control field restricts the writing of the service level field in the business status data, the interface control field restricts the calling of the service change interface, and the source tracking field is associated with the customer's online session content, existing contract records, and authorization approval records.

[0022] The graph inference module is used to infer the graph impact of candidate customer management action units, and generate the customer relationship graph impact radius and customer relationship graph version summary.

[0023] A customer relationship graph is constructed based on basic customer management data. The customer relationship graph includes graph structure data corresponding to customer management objects and the relationships between customer management objects. Customer management objects include customer objects, business records, status constraints, and management process items. During the construction process, customer object data is mapped to customer object nodes, contracts, orders, service orders, and follow-up records are mapped to business record nodes, credit limits, service levels, prohibited items, and billing rules are mapped to status constraint item nodes, and approval flows, review flows, renewal flows, and complaint handling flows are mapped to management process item nodes. Relationship edges are established according to customer affiliation, business dependency, status constraint, and process control relationships. At the same time, node attributes are configured for graph nodes, and edge attributes are configured for relationship edges. For example, when a corporate customer applies for a contact person change and requests a service level upgrade, a multi-level association is formed between the customer object node and the contact person business record node, the service contract business record node, the service level status constraint item node, and the change approval management process item node. The graph structure data records the business dependencies and permission restrictions between the customer management objects. Based on the candidate customer management action unit, an action impact subgraph is generated in the customer relationship graph. The action impact subgraph includes a set of action reachable nodes, a set of action propagation paths, a set of risk weights, a set of change rollback costs, and a set of time-effect decay coefficients. In this process, the action starting node is determined based on the action execution field, the node attributes or edge attributes that can be changed are determined based on the write control field, the management process items that may be reached are determined based on the interface control field, and the source of the action is verified by combining the source tracking field. Actions such as contact person change, service level adjustment, and billing period change are extended along the relationship edges to form a set of action propagation paths. For the set of action propagation paths in the action influence subgraph Any path The path hierarchy is denoted as The first in the path The risk weight of each relation edge is denoted as , No. The risk weight of each graph node is denoted as . The risk coefficient for the action type is denoted as The aging attenuation coefficient is denoted as The normalized value of the change rollback cost is denoted as The rollback cost amplification factor is denoted as . and The values ​​range from 0 to 1. The value range is 0 to 2. Path Weighted impact value Determine using the following formula:

[0024] Customer Relationship Map Influence Radius Determine using the following formula:

[0025] When the action propagation path set is empty, the influence radius of the customer relationship graph is set to 0. The risk weights of relationship edges and graph nodes are obtained by mapping relationship edge type, node type, field sensitivity level, and process constraint level; the action type risk coefficient is obtained by mapping the action category corresponding to the action execution field; the timeliness decay coefficient is obtained by mapping the time difference between the business record update time and the graph projection time; the normalized value of change rollback cost is obtained by normalizing the historical rollback time, number of manual approvals, and number of affected data records according to the normalization benchmark recorded in the system security policy library, and then weighting them according to their corresponding weights, with the sum of all weights being 1. A normalization result exceeding 1 is taken as 1, and a normalization benchmark of 0 is taken as 0 for the corresponding normalization item. The process constraint level is obtained from the customer management process configuration.

[0026] The customer relationship graph version summary is generated based on graph nodes, relationship edges, node attributes, edge attributes, and graph status timestamps. The graph status timestamp is the time of the most recent status change in the customer relationship graph, excluding the summary calculation time. Graph nodes are arranged in ascending order by node type, node identifier, and customer object identifier; relationship edges are arranged in ascending order by starting node identifier, edge type, and ending node identifier; and node attributes and edge attributes are arranged in ascending order by attribute name and attribute value summary. The arrangement results are concatenated with the graph status timestamps and hashed to generate the customer relationship graph version summary.

[0027] Reference Figure 2 The contamination simulation module is used to identify suspected contaminated instruction fragments based on customer interaction data and candidate customer management action units; generate persistent contamination propagation chains for suspected contaminated instruction fragments and determine contamination assessment characteristics; calculate persistent contamination risk index based on persistent contamination propagation chains and contamination assessment characteristics, and generate a persistent contamination propagation chain version summary based on persistent contamination propagation chains; wherein, the contamination assessment characteristics include source credibility characteristics.

[0028] Based on customer interaction data and candidate customer management action units, suspected instruction pollution fragments are identified. Customer interaction data includes customer messages, online conversations, email content, work order descriptions, and business processing remarks. Candidate customer management action units include action execution fields, write control fields, interface control fields, and source tracking fields. During the identification process, the natural expression content in the customer interaction data is compared item by item with the action execution fields, write control fields, and interface control fields to determine whether there is content that changes the judgment boundary of the candidate customer management strategy, bypasses authorization status data, expands the scope of write control fields, or induces interface control fields to deviate from business rules. For example, if a customer writes in the work order remarks that they want to ignore the approval record and directly upgrade the service level, but the authorization status data in the customer management basic data does not record the corresponding authorization, this content will be entered into the processing flow as a suspected instruction pollution fragment. Customer interaction data is segmented into several candidate text fragments. , Number the fragment. For any candidate text fragment. Extracting approval bypass score Write to expand score Interface deviation score Action conflict score Score for permission conflicts Each score ranges from 0 to 1. (Directive intervention score) Determine using the following formula:

[0029] in, and The intervention score weights all range from 0 to 1, and The intervention score weights are recorded in the system security policy library and saved in correspondence with the threshold policy version number.

[0030] When the directive intervention score reaches the semantic intervention threshold Furthermore, the approval bypass score reached the approval bypass threshold. The write expansion score reaches the write expansion threshold. The interface deviation score reaches the interface deviation threshold. The action conflict score reaches the action conflict threshold. The permission conflict score reaches the permission conflict threshold. When at least one of the following conditions is met, the candidate text fragment is added to the suspected instruction pollution fragment set. Otherwise, the candidate text fragment is not added to the suspected instruction pollution fragment set. The values ​​of each of the above thresholds range from 0 to 1. The semantic intervention threshold, approval bypass threshold, write expansion threshold, interface deviation threshold, action conflict threshold, and permission conflict threshold are recorded by the system security policy library and saved in correspondence with the threshold policy version number.

[0031] Approval bypass score, write expansion score, and interface deviation score are obtained by normalizing the ratio of the number of corresponding rules matched by the candidate text fragment to the total number of corresponding rules, respectively. Action conflict score is obtained by normalizing the number of conflicts between the actions involved in the candidate text fragment and the action execution field in the candidate customer management action unit. Permission conflict score is obtained by normalizing the number of conflicts between the permissions involved in the candidate text fragment and the permission scope of the requesting subject. A normalization result greater than 1 is taken as 1. If the corresponding rule set is empty or the total number of corresponding rules is 0, the corresponding score is 0. The normalization benchmark for action conflict score and permission conflict score is recorded by the system security policy library; when the normalization benchmark is 0, the corresponding score is 0. The permission scope of the requesting subject is obtained from the identity and permission configuration.

[0032] When the set of suspected instruction contamination fragments is empty, the persistent contamination propagation chain is denoted as an empty chain, and the persistent contamination risk index is set to 0. The empty chain version digest is generated by sequentially concatenating the empty chain identifier, the candidate customer management action unit identifier, and the source tracking field digest, and then performing a hash operation.

[0033] A persistent contamination propagation chain is generated for suspected command contamination fragments, and the contamination semantic features, source credibility features, and persistent write features are determined. The persistent contamination propagation chain records the link relationship from customer interaction data to candidate customer management strategy, candidate customer management action unit, and interface call result of the suspected command contamination fragment. The contamination semantic features are determined based on the degree of command intervention of the suspected command contamination fragment on customer interaction data, candidate customer management strategy, and interface call result limited by interface control fields. The source credibility features are determined based on the verifiability of the data source corresponding to the suspected command contamination fragment and the source propagation level. The persistent write features are determined based on the retention stability, replication range, and rollback difficulty after the suspected command contamination fragment is written into the customer management basic data. The persistent contamination propagation chain consists of a set of propagation chain nodes and a set of propagation chain edges. Each propagation chain node includes at least the following: suspected instruction contamination fragment node, customer interaction data node, candidate customer management strategy node, candidate customer management action unit node, interface call request node, interface call result node, target write field node, and data write event node. Each propagation chain node records at least the node identifier, node type, source tracking identifier, customer object identifier, timestamp, and data summary.

[0034] Data write event nodes include already occurred data write event nodes and proposed write event nodes formed based on candidate customer management action units. Proposed write event nodes record the target write field, target customer object, write method, and interface control field. Proposed write event nodes are established before the issuance of the customer management action authorization token and record the corresponding execution control status when the execution request is rejected or frozen.

[0035] A propagation chain edge must include at least a text reference edge, a strategy generation edge, an action formation edge, an interface call edge, and a field writing edge. Each propagation chain edge must record at least the starting node, the ending node, the edge type, the triggering interface, the written field, and the event time.

[0036] Suspected command contamination fragment When there is a data write event node or a planned write event node in the corresponding propagation chain, the persistent state... If 1 is selected; if neither exists, the persistent state is determined. Take 0.

[0037] The persistent contamination propagation chain version digest is generated based on the propagation chain nodes, propagation chain edges, node data digests, edge data digests, and propagation chain state timestamps. The propagation chain state timestamp is the time of the most recent state change in the propagation chain, without considering the digest calculation time. Propagation chain nodes are arranged in ascending order by node type, node identifier, and timestamp. Propagation chain edges are arranged in ascending order by starting node identifier, edge type, and ending node identifier. Node data digests and edge data digests are arranged in ascending order by their corresponding node identifier or corresponding edge identifier. The arrangement results are concatenated with the propagation chain state timestamps sequentially, and a hash operation is performed to generate the persistent contamination propagation chain version digest.

[0038] Targeting suspected command contamination fragments The pollution semantic risk value is denoted as The source credibility risk value is denoted as The risk value for persistent writing is denoted as All three values ​​range from 0 to 1. (Pollution semantic risk value) Take the directive intervention score of the suspected contaminated instruction fragment. The credibility of the source is denoted as... The value ranges from 0 to 1, representing the source credibility risk value. Take 1 and subtract the source credibility Source credibility The data is obtained by weighting customer interaction data based on source type, source authentication result, source tracking field completeness, source propagation level, and historical trusted records. The corresponding weights are saved with each threshold strategy version.

[0039] Persistent write risk value By preserving stability Scope of copying and rollback difficulty Weighted average acquisition:

[0040] in, and For persistent writing of weights, the values ​​range from 0 to 1, and .

[0041] When the written content is located in a temporary cache with an automatic expiration mechanism, the retention stability is 0.25; when it is located in a business temporary storage area that requires manual cleanup, the retention stability is 0.60; when writing to formal fields of the main customer management basic data and without an automatic expiration mechanism, the retention stability is 1.00. When the write scope is a single business record, the replication scope is 0.25; when the write scope is multiple business records of the same customer object, the replication scope is 0.50; when the write scope involves multiple customer objects or multiple business systems, the replication scope is 1.00. When the written record supports automatic transaction rollback, the rollback difficulty is 0.25; when manual approval or manual execution of rollback is required, the rollback difficulty is 0.60; when cross-system compensation is required or there is no direct rollback channel, the rollback difficulty is 1.00.

[0042] Link correction factor Determine using the following formula:

[0043] in, This is the maximum link correction factor, with a value ranging from 1 to 3; This is the normalized value for the propagation chain length; The normalized value for the number of branches in the propagation chain; It is a persistent state; and The link correction weights range from 0 to 1. The normalized propagation chain length is calculated as the ratio of the actual number of edges in the propagation chain to the reference chain length recorded in the system security policy database; if the ratio exceeds 1, it is rounded to 1. The normalized propagation chain branch number is calculated as the ratio of the actual number of branches in the propagation chain to the reference branch number recorded in the system security policy database; if the ratio exceeds 1, it is rounded to 1. Both the reference chain length and the reference branch number are positive integers greater than 0.

[0044] Suspected command contamination fragment Corresponding persistent pollution risk index Determine using the following formula:

[0045] in, and These are risk weights, all ranging from 0 to 1, and Persistent contamination risk index The value range is from 0 to 100.

[0046] The same candidate customer management action unit corresponds to When a suspected instruction contamination fragment is detected, the persistent contamination risk index of the candidate customer management action unit is calculated. Determine using the following formula:

[0047] When a candidate customer management action unit does not correspond to a suspected contaminated instruction fragment, the persistent contamination risk index of that candidate customer management action unit is... Take 0.

[0048] When multiple suspected instruction contamination fragments point to the same write control field or the same interface control field, the number of corresponding fragments is included in the actual number of branches in the propagation chain. The weights, reference chain lengths, and reference branch numbers are obtained based on historical manual review conclusions, rejected execution events, write rollback records, and interface out-of-bounds records, and are saved with the threshold policy version.

[0049] Reference Figure 3 , Figure 4 The token control module is used to authorize and control the management strategies of candidate customers based on the influence radius of the customer relationship graph and the persistent contamination risk index, and to generate authorization tokens for customer management actions.

[0050] Influence radius based on customer relationship graph Persistent contamination risk index of candidate customer management action unit Real-time source trust risk value at the time of token issuance Generate joint risk control value .

[0051] Customer Relationship Map Influence Radius Based on the reference radius of influence Normalization yields Reference radius of influence Recorded by the system security policy library and greater than 0:

[0052] Persistent pollution risk index After normalization to 100, the following was obtained :

[0053] The real-time source credibility at the time of token issuance is denoted as The value ranges from 0 to 1, representing a real-time source reliability risk value. Determine using the following formula:

[0054] Real-time source credibility The real-time source trust risk value is obtained by weighting the current source authentication result, session integrity result, channel trust level, source tracking field integrity, and source propagation level. Source Credibility Risk Value Different sampling times, source credibility risk values Real-time source reliability risk value corresponding to the source status at the time of formation of persistent contamination propagation chain. The current source status at the time the corresponding token was issued.

[0055] Joint risk control value Determine using the following formula:

[0056] in, and These are joint risk weights, all ranging from 0 to 1, and .

[0057] Dynamic authorization thresholds include automatic execution thresholds. Draft generation threshold , manual review threshold and freeze execution threshold All thresholds are between 0 and 100, and increase in the following order: automatic execution threshold, draft generation threshold, manual review threshold, and freeze execution threshold.

[0058] When the combined risk control value is not higher than the automatic execution threshold, the execution level is automatic execution; when the combined risk control value is higher than the automatic execution threshold but not higher than the draft generation threshold, the execution level is draft generation; when the combined risk control value is higher than the draft generation threshold but not higher than the manual review threshold, the execution level is manual review; when the combined risk control value is higher than the manual review threshold but not higher than the freeze execution threshold, the execution level is freeze execution; when the combined risk control value is higher than the freeze execution threshold, the execution level is prohibited execution.

[0059] Threshold strategy version number The sensitivity level is determined by matching the interface sensitivity level, the write field sensitivity level, the customer object level, and the current business status. The interface sensitivity level is obtained from the interface directory configuration, while the customer object level and current business status are obtained from the customer management basic data. When the same candidate customer management action unit involves multiple write fields or multiple interfaces, the highest sensitivity level is used for both the write field sensitivity level and the interface sensitivity level.

[0060] The sensitivity level of the field to be written is denoted as The value can be 1, 2, or 3, corresponding to ordinary fields, important business fields, and authorization status fields, respectively. Normal writes are for writing to formal fields of the main customer management basic data; source-restricted writes are for writing to the pending verification area, the review cache area, or the set of restricted fields, and the written record carries the source tracking field and the review identifier; isolated writes are for writing to an isolated storage area that is logically isolated from the main customer management basic data, and isolated write records do not trigger customer status changes, approval processes, or external interface calls.

[0061] When the execution level is automatic and the sensitivity level of the field to be written is not higher than 2, the write method is normal write. When the execution level is draft generation or manual review, the write method is source-restricted write. When the execution level is frozen or prohibited, the write method is isolated write or no write. When the sensitivity level of the field to be written is 3 and the manual review status is not "review passed", the write method is isolated write or no write; when the manual review status is "review passed", the write method is source-restricted write.

[0062] The reference impact radius, dynamic authorization threshold, sensitivity level of written fields, and various weights are maintained by the system security policy library. The system security policy library records the effective time, version number, subject of formulation, and reason for change of the corresponding parameters.

[0063] Taking the scenario of payment period extension application as an example, the candidate text fragments in the customer interaction data contain content that skips the approval process and is directly written into the authorization status data. The calculated semantic pollution risk value is then obtained. This excerpt originates from unverified paraphrased content; the source's credibility is questionable. Source credibility risk value .

[0064] After this segment enters the candidate customer management action unit, it involves writing business status data, maintaining stability. Copy range difficulty of rollback .Pick , , ,but:

[0065] normalized value of propagation chain length Normalized value of the number of branches in the propagation chain Persistent state Maximum link correction factor ,Pick ,but:

[0066] Pick ,but:

[0067] When the candidate client management action unit corresponds to only one suspected instruction contamination fragment, the persistence contamination risk index of the candidate client management action unit is... .

[0068] Take the radius of influence of the customer relationship graph Reference radius of influence ,get ;Pick ,get ; Obtain real-time source credibility Obtain the real-time source credibility risk value. .Pick ,but:

[0069] When the automatic execution threshold, draft generation threshold, manual review threshold, and frozen execution threshold are set to 20, 40, 60, and 80 respectively, the joint risk control value of 76 is higher than the manual review threshold but not higher than the frozen execution threshold. Therefore, the execution level is determined to be the frozen execution level, and the writing method is determined to be isolated writing.

[0070] The above parameters are obtained based on historical manual review samples, interface sensitivity level configurations, field sensitivity level configurations, write rollback records, and system security policy library records, and are updated according to the threshold policy version number. Solidify and preserve.

[0071] The execution control results are determined based on the joint risk control values, and customer management action authorization tokens are generated based on the execution control results. The execution control results include a dynamic authorization threshold set, execution level, and writing method. The execution level is determined from automatic execution level, draft generation level, manual review level, frozen execution level, and prohibited execution level. The writing method is determined from normal writing, source-restricted writing, and isolated writing. The customer management action authorization token carries action constraint information, write constraint information, interface constraint information, risk constraint information, version constraint information, validity verification information, execution level, and manual review status. The action constraint information limits the action execution fields and scope of impact. The write constraint information limits the write control fields and scope of writing. The interface constraint information limits the interface control fields. The risk constraint information limits the customer relationship graph influence radius and persistent contamination risk index. The version constraint information limits the customer relationship graph version summary and persistent contamination propagation chain version summary. The validity verification information is responsible for verifying the validity of the customer management action authorization token. The manual review status includes unreviewed status, reviewed passed status, and reviewed failed status. In one embodiment, the customer management action authorization token records at least the token identifier, candidate customer management action unit identifier, issuing entity identifier, requesting entity identifier, customer object identifier, interface audience, issuance time, effective time, expiration time, one-time use identifier, execution level, manual review status at the time of issuance, action constraint information, write constraint information, interface constraint information, risk constraint information, version constraint information, validity verification information, customer relationship graph version summary, persistent contamination propagation chain version summary, threshold policy version number, action summary, write summary, interface summary, risk summary, version summary, and signature value.

[0072] Action summary is generated from action execution field and action impact scope; write summary is generated from write control field, target write field and write scope; interface summary is generated from interface control field, interface method and interface parameter range; risk summary is generated from customer relationship graph impact radius, persistence contamination risk index and execution level; version summary is generated from customer relationship graph version summary, persistence contamination propagation chain version summary and threshold policy version number.

[0073] Action constraint information is formed by the action execution field and the action's impact scope. Write constraint information is formed by the write control field, target write field, write scope, and write method. Interface constraint information is formed by the interface control field, interface method, interface parameter range, and interface audience. Risk constraint information includes the allowable range of the customer relationship graph's influence radius and the allowable range of the persistent contamination risk index. The allowable range in the risk constraint information is determined based on the dynamic authorization threshold set, execution level, and threshold policy version number applicable when the customer management action authorization token is issued, and is written into the customer management action authorization token. Version constraint information includes the customer relationship graph version summary and the persistent contamination propagation chain version summary. Validity verification information includes issuance time, effective time, expiration time, one-time use identifier, and issuing entity identifier. The signature value is stored as an independent field in the customer management action authorization token, and the signature overlay field does not include the signature value itself.

[0074] The issuing entity performs a signature on the data formed by sequentially concatenating the following elements using its private key: token identifier, candidate customer management action unit identifier, issuing entity identifier, requesting entity identifier, customer object identifier, interface audience, issuance time, effective time, expiration time, one-time use identifier, execution level, manual review status at the time of issuance, action constraint information, write constraint information, interface constraint information, risk constraint information, version constraint information, validity verification information, customer relationship graph version digest, persistent contamination propagation chain version digest, threshold policy version number, action digest, write digest, interface digest, risk digest, and version digest, to obtain the signature value.

[0075] The execution verification module verifies the signature value based on the issuing entity's public key and checks whether the current time is within the valid time range from the effective time to the expiration time. After signature verification, validity time verification, request consistency verification, risk constraint verification, version digest verification, threshold policy version verification, and manual review status verification corresponding to the execution level all pass, the execution verification module performs an atomic query and write to the one-time use identifier before executing the customer management action. If the same one-time use identifier does not exist in the token usage record, the one-time use identifier is written; if the same one-time use identifier already exists or the atomic write of the one-time use identifier fails, the authorization token for the customer management action is determined to be a duplicate submission, and the execution verification module outputs a rejection instruction.

[0076] For example, in a service level adjustment scenario, when the combined risk control value is lower than the automatic execution threshold in the dynamic authorization threshold set, the customer management action authorization token records the automatic execution level and normal write; when the combined risk control value falls into the manual review range, the customer management action authorization token records the manual review level, source restricted write, and unreviewed status; when the combined risk control value exceeds the freeze threshold, the customer management action authorization token records the freeze execution level or prohibit execution level, and restricts the corresponding write range and interface call range.

[0077] The execution verification module is used to perform a second authorization verification on the customer management action authorization token before execution when it receives an execution request from the downstream customer management execution system. If the second authorization verification fails, it outputs an execution control instruction to the downstream customer management execution system. The execution control instruction includes a rejection instruction, a freeze instruction, or a resubmission instruction for manual review.

[0078] Upon receiving an execution request from the downstream customer management execution terminal, a second authorization verification is performed on the customer management action authorization token carried in the execution request before execution. The verification includes whether the validity verification information is complete, whether the action execution field pointed to by the execution request falls within the scope of the action constraint information, whether the data field to be written falls within the scope of the write constraint information, whether the interface called by the request conforms to the interface constraint information, whether the influence radius of the customer relationship graph and the persistent contamination risk index meet the risk constraint information, whether the version summary of the customer relationship graph and the version summary of the persistent contamination propagation chain are consistent with the version status at the current verification time, and whether the execution level and manual review status meet the conditions for this execution. In one embodiment, the execution verification module obtains the action field set, write field set, interface method, interface parameter set, interface audience, customer object identifier, and request subject identifier from the execution request, and obtains the action constraint information, write constraint information, interface constraint information, risk constraint information, version constraint information, customer object identifier, and request subject identifier from the customer management action authorization token.

[0079] The request consistency check passes if all action fields in the execution request fall within the action constraints recorded in the token, all write fields fall within the write constraints recorded in the token, the interface method is consistent with the interface method recorded in the token, the interface parameter set falls within the interface parameter range recorded in the token, the interface audience is consistent with the interface audience recorded in the token, and the execution request is consistent with the client object identifier and request subject identifier recorded in the token. The request consistency check fails if any field exceeds its corresponding constraint range, the interface method is inconsistent, the interface audience is inconsistent, or any identifier is inconsistent. The write method carried in the execution request is not used as the basis for execution; the execution verification module only processes the corresponding customer management action according to the write method recorded in the token.

[0080] The execution verification module re-obtains the customer relationship graph influence radius and persistent contamination risk index based on the threshold policy version recorded in the token; when any current risk value exceeds the corresponding allowable range recorded in the token, the execution verification module outputs a freeze execution command.

[0081] If signature verification fails, token expires, interface audience does not match, one-time use identifier already exists, one-time use identifier atomic write fails, or request consistency verification fails, the execution verification module outputs a rejection instruction.

[0082] If the version summary of the customer relationship graph or the version summary of the persistent contamination propagation chain is inconsistent with the corresponding version summary recalculated at the current verification time, the execution verification module outputs a freeze execution command. If the threshold policy version number recorded in the customer management action authorization token is inconsistent with the currently effective threshold policy version number, the execution verification module outputs a freeze execution command.

[0083] When the execution level is automatic and all other validation items pass, the execution validation module performs the corresponding customer management action according to the writing method and interface constraint information recorded in the token. When the execution level is draft generation and all other validation items pass, the execution validation module generates a customer management draft to be confirmed or writes the corresponding content to the review cache. Before confirmation, the writing of formal fields of the main customer management basic data and external interface calls are not triggered.

[0084] The manual review status recorded on the customer management action authorization token is the status at the time of issuance. The manual review status during the execution verification phase is obtained from the manual review record, which is stored correspondingly with the token identifier, candidate customer management action unit identifier, review subject identifier, review time, and review result. If no corresponding manual review record is found, the manual review status is recorded as "not reviewed." The manual review status carried in the execution request is not used as the basis for execution. When the execution level is manual review and the manual review status is "not reviewed," the execution verification module outputs a resubmission manual review instruction; when the manual review status is "review failed," an execution rejection instruction is output; when the manual review status is "review passed" and all other verification items pass, the corresponding customer management action is executed according to the writing method and interface constraint information recorded on the token.

[0085] When the execution level is frozen, the execution verification module outputs a freeze execution command; when the execution level is prohibited, it outputs a refuse execution command.

[0086] For example, if the execution level recorded on the customer management action authorization token is manual review, and the manual review status is still unreviewed, but the execution request requires directly writing customer authorization status data and calling the service level change interface, then it is determined that the pre-execution secondary authorization verification failed, and a resubmission manual review instruction is output. Another example is if the customer relationship graph version summary is inconsistent with the summary recalculated from the current customer relationship graph, indicating that the customer object, business record, status constraint item, or management process item has changed, then a freeze execution instruction is output. Finally, if the interface call scope in the execution request exceeds the interface constraint information limit, or the writing scope exceeds the writing constraint information limit, then a refuse execution instruction is output.

[0087] Example 2: A customer management data processing method based on artificial intelligence, comprising: Step 1: Using an AI model, candidate customer management strategies are generated based on customer interaction data and basic customer management data. Corresponding candidate customer management action units are then identified. Customer interaction data reflects requests, feedback, changes in intent, and interaction context during customer communication. Basic customer management data reflects customer object data, business-related data, business status data, and authorization status data. The AI ​​model generates candidate customer management strategies under the combined constraints of both, avoiding management decisions based solely on customer interaction content. The candidate customer management action units correspond to the candidate customer management strategies and can contain action execution fields, write control fields, interface control fields, and source tracking fields, facilitating subsequent consistency processing of action boundaries, write scope, interface calls, and data sources.

[0088] Step 2: Perform graph impact simulation on candidate customer management action units to generate the customer relationship graph impact radius and customer relationship graph version summary. The customer relationship graph contains graph structure data corresponding to customer management objects and the relationships between them. When customer management actions involve customer objects, business records, state constraints, and management process items, an action impact subgraph will be formed in the customer relationship graph. The customer relationship graph impact radius reflects the impact strength of the action propagation path in the action impact subgraph. The customer relationship graph version summary records the graph nodes, relationship edges, node attributes, and edge attribute states of the customer relationship graph during simulation, facilitating version consistency verification before execution.

[0089] Step 3: Identify suspected command contamination fragments based on customer interaction data and candidate customer management action units. This involves identifying content in customer interaction data that may interfere with candidate customer management strategies or action units. Suspected command contamination fragments are not limited to explicit command statements, but also include content embedded in customer interaction data, affecting action execution fields, writing to control fields, or determining interface control field results. By analyzing the correspondence between customer interaction data and candidate customer management action units, fragments that abnormally guide customer management strategies, exhibit unauthorized writing tendencies, or deviate from interface call settings can be located.

[0090] Step 4: Generate a persistent contamination propagation chain for suspected instruction contamination fragments and determine contamination assessment features, including source credibility characteristics. The persistent contamination propagation chain records the propagation path of the suspected instruction contamination fragment among customer interaction data, basic customer management data, candidate customer management strategies, and candidate customer management action units. The contamination assessment features include at least source credibility characteristics, which are determined by combining the verifiability of the data source corresponding to the suspected instruction contamination fragment with the source propagation level. This processing method can distinguish between customer information with a clear source and a well-defined path and abnormal content with a complex source and a deeper propagation level.

[0091] Step 5: Calculate the persistent pollution risk index based on the persistent pollution propagation chain and pollution assessment characteristics, and generate a persistent pollution propagation chain version summary based on the persistent pollution propagation chain. The persistent pollution propagation chain is combined with the pollution assessment characteristics to form the persistent pollution risk index. The pollution semantic features, source credibility features, and persistent writing features in the pollution assessment characteristics correspond to the degree of directive intervention, the reliability of the data source, and the retention stability, replication scope, and rollback difficulty after being written into the customer management basic data, respectively. The persistent pollution propagation chain version summary records the version state of the propagation chain during calculation, avoiding the reference of changed propagation chain results during the execution phase.

[0092] Step Six: Authorize and control candidate customer management strategies based on the customer relationship graph's influence radius and persistent contamination risk index, and generate customer management action authorization tokens. Integrate the graph influence projection results and contamination risk assessment results into the same authorization control process. The customer relationship graph's influence radius reflects the scope of influence of candidate customer management action units on the customer relationship graph, and the persistent contamination risk index reflects the contamination risk of suspected instruction contamination fragments on candidate customer management strategies. The authorization control results are linked to the customer management action authorization token, which carries action constraint information, write constraint information, interface constraint information, risk constraint information, version constraint information, validity verification information, execution level, and manual review status.

[0093] Step 7: Receive the execution request sent by the downstream customer management execution system, and perform a pre-execution secondary authorization verification on the customer management action authorization token. If the verification fails, an execution control instruction is output. This step occurs when the downstream customer management execution system submits the execution request, performing a pre-execution secondary authorization verification on the customer management action authorization token. The verification content includes action constraint information, write constraint information, interface constraint information, risk constraint information, version constraint information, validity verification information, execution level, and manual review status. If the pre-execution secondary authorization verification fails, an execution control instruction is output to the downstream customer management execution system. The execution control instruction includes a rejection instruction, a freeze instruction, or a resubmission instruction for manual review.

[0094] 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 scope of the technology 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.

Claims

1. A customer management data processing system based on artificial intelligence, characterized in that, include: The strategy generation module is used to generate candidate customer management strategies based on customer interaction data and basic customer management data through an artificial intelligence model, and to determine the candidate customer management action units corresponding to the candidate customer management strategies. The graph inference module is used to infer the graph impact of candidate customer management action units and generate the customer relationship graph impact radius and customer relationship graph version summary. The contamination simulation module is used to identify suspected contaminated instruction fragments based on customer interaction data and candidate customer management action units; generate persistent contamination propagation chains for suspected contaminated instruction fragments and determine contamination assessment characteristics; A persistent pollution risk index is calculated based on the persistent pollution propagation chain and pollution assessment characteristics, and a persistent pollution propagation chain version summary is generated based on the persistent pollution propagation chain; wherein, the pollution assessment characteristics include source credibility characteristics; The token control module is used to authorize and control the management strategies of candidate customers based on the influence radius of the customer relationship graph and the persistent pollution risk index, and to generate authorization tokens for customer management actions. The execution verification module is used to perform a second authorization verification on the customer management action authorization token before execution when it receives an execution request from the downstream customer management execution system. If the second authorization verification fails, it outputs an execution control instruction to the downstream customer management execution system. The execution control instruction includes a rejection instruction, a freeze instruction, or a resubmission instruction for manual review.

2. The customer management data processing system based on artificial intelligence according to claim 1, characterized in that, Customer management basic data includes customer object data, business association data, business status data, and authorization status data. Candidate customer management strategies include at least one candidate customer management action unit, which carries an action execution field, a write control field, an interface control field, and a source tracking field.

3. The customer management data processing system based on artificial intelligence according to claim 2, characterized in that, The generation methods for the customer relationship graph influence radius and customer relationship graph version summary include: constructing a customer relationship graph based on basic customer management data; generating an action influence subgraph in the customer relationship graph based on candidate customer management action units; calculating the customer relationship graph influence radius based on the action influence subgraph; obtaining graph nodes, relationship edges, node attributes, and edge attributes from the customer relationship graph; sorting the graph nodes, relationship edges, node attributes, and edge attributes according to a preset sorting rule; generating a graph state sequence; and performing a hash operation on the graph state sequence to obtain the customer relationship graph version summary.

4. The customer management data processing system based on artificial intelligence according to claim 3, characterized in that, The calculation method for the influence radius of the customer relationship graph includes: determining the path level, risk weight, and time decay coefficient of each action propagation path in the action influence subgraph, calculating the weighted influence value of each action propagation path, and determining the maximum weighted influence value as the influence radius of the customer relationship graph.

5. The customer management data processing system based on artificial intelligence according to claim 3, characterized in that, The customer relationship graph includes the graph structure data corresponding to the customer management objects and the relationships between them. The customer management objects include customer objects, business records, status constraints, and management process items. The action impact subgraph includes the set of action reachable nodes, the set of action propagation paths, the set of risk weights, the set of change rollback costs, and the set of time-decrease coefficients.

6. The customer management data processing system based on artificial intelligence according to claim 1, characterized in that, The contamination assessment features include contamination semantic features, source credibility features, and persistent write features. Among them, the contamination semantic features are determined based on the degree of directive intervention of suspected instruction contamination fragments on customer interaction data, candidate customer management strategies, and the interface call results of downstream customer management execution systems limited by interface control fields. The source credibility features are determined based on the verifiability of the data source corresponding to the suspected instruction contamination fragments and the source propagation level. The persistent write features are determined based on the retention stability, replication scope, and rollback difficulty after the suspected instruction contamination fragments are written into the customer management basic data.

7. A customer management data processing system based on artificial intelligence according to claim 6, characterized in that, The calculation method of the persistent contamination risk index includes: determining the contamination semantic risk value based on the contamination semantic features, determining the source credibility risk value based on the source credibility features, determining the persistent write risk value based on the persistent write features, and determining the link correction coefficient based on the persistent contamination propagation chain; weighting and summing the contamination semantic risk value, source credibility risk value, and persistent write risk value according to preset weights, and correcting the weighted summation result according to the link correction coefficient to obtain the persistent contamination risk index corresponding to the suspected instruction contamination fragment.

8. The customer management data processing system based on artificial intelligence according to claim 7, characterized in that, The token control module generates customer management action authorization tokens in the following way: it generates a joint risk control value based on the customer relationship graph influence radius, persistent contamination risk index and source credibility characteristics, determines the execution control result based on the joint risk control value, and generates a customer management action authorization token based on the execution control result. The execution control results include a set of dynamic authorization thresholds, execution levels, and write methods. Execution levels include automatic execution level, draft generation level, manual review level, frozen execution level, and prohibited execution level. Write methods include normal write, source-restricted write, and isolated write.

9. A customer management data processing system based on artificial intelligence according to claim 8, characterized in that, The customer management action authorization token carries action constraint information, write constraint information, interface constraint information, risk constraint information, version constraint information, validity verification information, execution level, and manual review status; Among them, action constraint information limits the action execution fields and scope of influence; write constraint information limits the write control fields and scope of write; interface constraint information limits the interface control fields; risk constraint information limits the influence radius of the customer relationship graph and the persistent contamination risk index; version constraint information limits the version summary of the customer relationship graph and the version summary of the persistent contamination propagation chain; validity verification information is used to verify the validity of the customer management action authorization token; and manual review status includes unreviewed status, reviewed passed status, and reviewed failed status.

10. A customer management data processing method based on artificial intelligence, applied to the customer management data processing system based on artificial intelligence as described in any one of claims 1-9, characterized in that, include: Step 1: Generate candidate customer management strategies based on customer interaction data and basic customer management data using an artificial intelligence model, and determine the candidate customer management action units corresponding to the candidate customer management strategies; Step 2: Perform graph impact analysis on candidate customer management action units to generate customer relationship graph impact radius and customer relationship graph version summary; Step 3: Identify suspected instruction contamination fragments based on customer interaction data and candidate customer management action units; Step 4: Generate a persistent contamination propagation chain for suspected instruction contamination fragments and determine contamination assessment features, including reliable source characteristics; Step 5: Calculate the persistent pollution risk index based on the persistent pollution propagation chain and pollution assessment characteristics, and generate a persistent pollution propagation chain version summary based on the persistent pollution propagation chain; Step 6: Based on the influence radius of the customer relationship graph and the persistent contamination risk index, authorize and control the management strategies for candidate customers, and generate authorization tokens for customer management actions; Step 7: Receive the execution request sent by the downstream customer management execution system, perform a second authorization verification on the customer management action authorization token before execution, and output the execution control command if the verification fails. Among them, execution control instructions include instructions to refuse execution, freeze execution instructions, or resubmit for manual review.

Images