An industry intelligent circle guest marketing method and system based on compliance portrait knowledge graph and multi-agent strategy arrangement

Abstract

The application discloses an industry intelligent circle guest marketing method and system based on compliance portrait knowledge graph and multi-agent strategy arrangement. The method constructs a compliance portrait knowledge graph with written authorization range, evidence number, sensitive level, disabled matters and reach constraint, extracts a candidate subject local subgraph to generate a candidate circle layer, generates a local permission view for each agent, binds the candidate marketing action with the graph version, the agent identity, the permission view abstract, the reach channel, the reach time and the frequency control quota, generates or rejects a compliance pass token by a compliance review agent, reconfigures the rejected or conflicted action according to the rejection code, generates an audit hash chain, and realizes compliance gating, permission isolation and traceable execution of the marketing action.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, knowledge graphs, data compliance processing, and intelligent marketing technologies, and in particular to an industry-specific intelligent customer acquisition marketing method and system based on a compliance profile knowledge graph and multi-agent strategy orchestration. Background Technology

[0002] With the development of digital operations and the industrial internet, industry marketing is gradually shifting from manual customer screening to customer acquisition based on data modeling, customer profiling, and intelligent recommendations. Existing industry marketing systems typically collect multi-source data, including business registration information, transaction information, bidding information, public opinion, product purchase records, and historical contact records. They then use customer tags, knowledge graphs, or machine learning models to identify potential customers, thereby improving the efficiency of lead generation and conversion rates. Knowledge graphs, due to their ability to represent enterprise entities, supply chain relationships, transaction relationships, product demand relationships, and regional industry relationships, have been widely used for potential customer discovery and industry relationship analysis.

[0003] For example, Chinese patent CN111951052A discloses a method and apparatus for acquiring potential customers based on a knowledge graph. This method obtains enterprise entity data and relationship data between enterprise entities based on enterprise invoice information, constructs an invoice knowledge graph, and identifies potential customers based on this graph. This solution can utilize the invoice relationships between enterprises to mine potential customers, improving the efficiency of potential customer identification and the marketing success rate. This type of solution illustrates that graphing enterprise relationship data and screening potential customers based on the graph has become a common technical approach in intelligent marketing and customer mining.

[0004] However, existing customer acquisition solutions based on knowledge graphs focus more on issues such as whether potential customers match, whether the marketing value is high, and whether the relationship path is reliable. They lack sufficient consideration for constraints such as the scope of authorization, sensitive attributes, no-reach rules, frequency control limits, evidence sources, and compliance audits during the use of marketing data. Especially in multi-source data fusion scenarios, different data sources have different authorization purposes, usage periods, and reach restrictions. If marketing lists are generated directly based solely on the strength of graph relationships or model scores, problems such as unauthorized use of customer data, re-reaching unsubscribed or complained entities, repeated reach to the same entity in different marketing tasks, and intelligent models calling data beyond the necessary scope can easily occur. These problems not only affect the reliable operation of the marketing system but also create high data compliance risks.

[0005] On the other hand, with the development of large language models and intelligent agent technology, marketing systems have begun to introduce multiple intelligent agents to perform tasks such as lead discovery, customer assessment, content generation, channel recommendation, and result verification. Multi-agent orchestration can improve the ability to handle complex tasks, but it also brings new technical problems: different intelligent agents may access different data attributes in the same knowledge graph when performing tasks. If there is a lack of least privilege control for intelligent agent tasks, it is easy for intelligent agents to read sensitive information unrelated to the current task; marketing actions generated by different intelligent agents may conflict in terms of reach time, reach channels, frequency control limits, or compliance rules; if the system only performs manual review at the final output stage, it cannot intercept non-compliant actions in a timely manner during action generation and task flow, making it difficult to form a traceable chain of execution evidence.

[0006] For example, Chinese patent CN121457588B discloses a method for updating and verifying compliance of a financial knowledge graph based on multi-source heterogeneous data. This method constructs a financial knowledge graph using multi-source heterogeneous financial data. When a financial business request is received, it extracts a local subgraph using the business entity and business type as anchor points, performs reasoning on the local subgraph, and outputs an interpretable compliance audit report. This solution focuses on knowledge graph updates and compliance audits in financial business scenarios, demonstrating the technical feasibility of extracting local subgraphs and performing compliance reasoning in specific business requests.

[0007] However, the aforementioned compliance knowledge graph solutions primarily target audit judgments for financial business requests, focusing on compliance reasoning between business requests and regulatory rules. They do not provide action-level execution control for candidate marketing actions during the industry's marketing and customer acquisition process. Specifically, existing technologies do not address the following issues: how to bind candidate marketing actions to local subgraph versions, agent identities, permission view summaries, reach channels, reach times, and frequency control limits; how to ensure that each marketing action obtains a valid compliance pass token before entering subsequent agent task nodes; how to automatically perform downgrading, delay, manual review, or elimination and reconstruction based on the reason for rejection when compliance verification fails; and how to write local permission views, agent calls, compliance pass tokens, rejection codes, and reconstruction results into a traceable audit chain. Summary of the Invention

[0008] The technical objective of this invention is to provide an industry-specific intelligent customer acquisition marketing method and system based on a compliance profile knowledge graph and multi-agent strategy orchestration. By writing authorization scope, evidence number, sensitivity level, prohibited items, and reach constraints into the knowledge graph, and combining it with multi-agent task orchestration, local permission view, compliance access token, and audit hash chain, the system achieves full-process compliance control over the generation, flow, reconstruction, and execution of candidate marketing actions. This solves the problems in existing intelligent customer acquisition marketing such as unclear boundaries of multi-source data use, unauthorized access by agents, unauthorized reach of marketing actions, difficulty in resolving strategy conflicts, and difficulty in tracing the execution process.

[0009] To achieve the objectives of this invention, the following technical solution is adopted:

[0010] An industry-specific intelligent customer acquisition marketing method based on compliance profile knowledge graph and multi-agent strategy orchestration includes the following steps:

[0011] S1 acquires industry data, publicly available enterprise data, authorized customer data, product compatibility data, outreach records and regulatory rules, constructs a compliance profile knowledge graph, and writes data source, authorization scope, evidence number, sensitivity level, prohibited items and outreach constraints into nodes or edges;

[0012] S2 extracts partial subgraphs of candidate subjects using target industries and product tasks as anchor points, and generates candidate layers based on demand relationships, product compatibility relationships, reachability status and compliance risks.

[0013] S3 generates a multi-agent task orchestration diagram based on clue discovery, value assessment, compliance review, channel communication, and verification tasks, and generates a local permission view for each agent containing only the nodes, edges, attributes, and tool interfaces required for its tasks.

[0014] S4 binds each candidate marketing action with the corresponding local subgraph version, agent identity, permission view summary, reach channel, reach time and frequency control limit, and the compliance review agent generates or refuses to generate a compliance pass token.

[0015] S5 only allows candidate marketing actions with valid compliance tokens to enter subsequent task nodes. For rejected or conflicting candidate marketing actions, downgrade, delay, transfer to manual review, or remove and reconstruct them based on the rejection code.

[0016] S6 outputs the target subject list, reach channels, key points of the script, execution order and audit hash chain, and updates the subject demand relationship, product adaptation relationship, channel preference relationship, frequency control status, permission view generation rules and intelligent agent strategy parameters based on the legitimate reach feedback.

[0017] As a further improvement, in step S1, the industry data includes industry classification, position in the industrial chain, procurement scenarios, business events, regional industrial policies, and industry regulatory requirements; the publicly available enterprise data includes business registration information, bidding information, recruitment information, intellectual property information, public opinion information, and abnormal business information; and the authorized customer data includes authorized historical transactions, consultation records, subscription preferences, access permission status, and authorization validity period.

[0018] As a further improvement, the reach constraints include allowed reach channels, prohibited reach channels, allowed reach time periods, prohibited reach time periods, single-entity reach frequency control, same-channel reach frequency control, unsubscribe status, complaint status, and manual review status; when the entity node is in an unsubscribe status, complaint status, or authorization invalidation status, it is prohibited to generate proactive reach compliance tokens for that entity.

[0019] As a further improvement, in step S2, the extraction of candidate subject local subgraphs includes: taking the target industry node, target product node and existing high-quality customer node as the starting node, expanding the path along the upstream and downstream relationships of the industrial chain, similar demand relationships, transaction collaboration relationships, personnel function relationships, regional business circle relationships and public event relationships, and performing authorization filtering, sensitivity level filtering and reachability status filtering on the expansion results.

[0020] As a further improvement, in step S3, generating the local permission view includes: determining the allowed node types, edge types, attribute fields, and tool interfaces based on the agent's task type; performing desensitization, masking, or summarization processing on unnecessary attributes; and binding the local permission view with the agent's identity, task node number, and local subgraph version, so that the agent cannot call graph attributes and external tools that are unrelated to the current task.

[0021] As a further improvement, in step S4, the candidate marketing action includes the target subject, recommended product, reach channel, reach time, reach content type, and execution priority; the compliance pass token includes action number, agent identity, local subgraph version, permission view summary, allowed channel, allowed time, frequency control remaining quota, evidence number, validity period, token status, and prohibited items; the rejection code includes authorization missing rejection code, authorization expired rejection code, sensitive level rejection code, prohibited item rejection code, frequency control over-limit rejection code, channel mismatch rejection code, time mismatch rejection code, evidence missing rejection code, and manual review rejection code; each rejection code is written into the audit hash chain and serves as the trigger condition for subsequent candidate marketing action reconstruction.

[0022] As a further improvement, in step S5, the refactoring of rejected or conflicting candidate marketing actions includes: when the rejection code is a channel mismatch rejection code, replacing it with an allowed channel; when the rejection code is a time mismatch rejection code, delaying it until the allowed time; when the rejection code is a frequency control over-limit rejection code, lowering the execution priority or postponing execution; when the rejection code is a missing evidence rejection code or a manual review rejection code, transferring to manual review; when the rejection code is a missing authorization rejection code, an expired authorization rejection code, a sensitive level rejection code, or a prohibited item rejection code, removing the candidate marketing action.

[0023] As a further improvement, in step S6, the audit hash chain includes graph version records, local permission view records, agent call records, candidate marketing action records, compliance access token records, rejection code records, reconstruction records, manual review records, and outreach feedback records; each subsequent record contains the hash digest of the previous record to form a traceable chain audit structure.

[0024] Another aspect of the present invention provides an industry-specific intelligent customer acquisition marketing system based on a compliance profile knowledge graph and multi-agent strategy orchestration. This system is used to implement the method described above, comprising: a graph construction module for acquiring industry data, publicly available enterprise data, authorized customer data, product compatibility data, reach records, and regulatory rules, and constructing a compliance profile knowledge graph that includes data sources, authorization scope, evidence numbers, sensitivity levels, prohibited items, and reach constraints; a candidate circle generation module for extracting partial subgraphs of candidate subjects using target industry and product tasks as anchor points, and generating candidate circles; and a permission orchestration module for generating a multi-agent task orchestration graph, and generating a section containing only the necessary elements for each agent's task. The system includes: a local permission view for points, edges, attributes, and tool interfaces; a token gating module to bind each candidate marketing action to a local subgraph version, agent identity, permission view summary, reach channel, reach time, and frequency control limit, and to generate or reject the generation of a compliance pass token; a policy reconstruction module to only allow candidate marketing actions carrying valid compliance pass tokens, and to downgrade, delay, transfer to manual review, or remove and reconstruct candidate marketing actions that are rejected or conflict with each other; and an audit update module to output the target subject list, reach channel, key points of the message, execution order, and audit hash chain, and to update the compliance profile knowledge graph, permission view generation rules, and agent policy parameters based on legitimate reach feedback.

[0025] As a further improvement, the token gating module includes an authorization verification unit, a sensitivity level verification unit, a disabled item verification unit, an access constraint verification unit, a frequency control verification unit, an evidence traceability verification unit, and a token generation unit. The token generation unit is used to generate a compliance access token when all verification units pass, and to output a corresponding rejection code when any verification unit fails. The audit update module includes a chain recording unit, a rule change backtracking unit, and a feedback update unit. The chain recording unit is used to write the graph version, local permission view, agent call, compliance access token, rejection code, and reconstruction result into the audit hash chain. The rule change backtracking unit is used to re-trigger the token gating module for candidate marketing actions that have not been executed when regulatory rules or internal enterprise compliance rules change.

[0026] The technical advantages of this invention are as follows: By constructing a compliance profile knowledge graph, industry data, publicly available enterprise data, authorized customer data, product compatibility data, outreach records, and regulatory rules can be expressed in a unified graph data structure, and authorization scope, evidence number, sensitivity level, prohibited items, and outreach constraints can be made into computable and verifiable graph attributes; through candidate subject local subgraph extraction and local permission view generation, different agents can only access the nodes, edges, attributes, and tool interfaces required for their tasks, reducing the risk of sensitive information exposure and unauthorized access during multi-agent collaboration; by linking each candidate marketing action with local subgraph version, agent identity, permission view summary, and outreach channel... The system binds channels, reach times, and frequency control limits, with a compliance review agent generating or rejecting compliance tokens. This ensures that marketing actions are subject to action-level compliance gating before entering subsequent task nodes. Through a rejection code-driven mechanism of downgrading, delaying, transferring to manual review, or eliminating / reconstructing, correctable actions such as channel mismatch, time mismatch, and frequency control exceeding limits can be compliantly corrected, while high-risk actions such as missing authorization, expired authorization, sensitive level triggering, and prohibited event triggering can be promptly blocked. By auditing the hash chain record graph version, partial permission view, agent calls, candidate marketing actions, token status, rejection codes, reconstruction results, and reach feedback, the system achieves complete traceability of the marketing strategy generation and execution process. Therefore, this invention can significantly enhance the compliance, security, and auditability of marketing data processing while improving the automation level and executability of customer acquisition strategies in the industry. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the overall process of the industry intelligent customer acquisition marketing method based on compliance profile knowledge graph and multi-agent strategy orchestration of the present invention;

[0028] Figure 2 This is a schematic diagram of the data structure of the compliance profile knowledge graph of this invention. Detailed Implementation

[0029] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of the present invention.

[0030] This invention provides an industry-specific intelligent customer acquisition marketing method and system based on a compliance profile knowledge graph and multi-agent strategy orchestration. The method is executed by a server, which can be a single server, a server cluster, a cloud server, a privately deployed enterprise marketing platform, or a computing platform composed of a graph database, a relational database, a rule engine, a multi-agent scheduling service, and an audit and evidence storage service. The core of this invention does not lie in simply generating a marketing list, but in providing computable, interceptable, reconstructable, and traceable technical control over candidate marketing actions generated by multiple agents through a compliance profile knowledge graph, local permission views, action-level compliance access tokens, rejection code-driven reconstruction, and audit hash chains.

[0031] I. Terminology Explanation

[0032] Industry-specific intelligent customer acquisition marketing refers to a computer data processing process in which servers, focusing on a target industry, product, region, or marketing task, identify potential businesses, organizations, or authorized customers from multi-source data and generate contact channels, contact times, key communication points, and execution order. This process includes data fusion, graph modeling, local subgraph extraction, candidate entity screening, multi-agent task orchestration, compliance gating, strategy restructuring, and audit updates.

[0033] A compliance profiling knowledge graph refers to a graph data structure formed by using entities, industries, products, channels, rules, evidence, and actions as graph nodes, and industry chain relationships, demand relationships, product adaptation relationships, authorization relationships, reach relationships, prohibited relationships, evidence source relationships, and rule constraint relationships as graph edges. Unlike ordinary marketing knowledge graphs, the compliance profiling knowledge graph of this invention writes data source, authorization scope, evidence number, sensitivity level, prohibited items, and reach constraints into nodes or edges, enabling the system to simultaneously perform compliance control during graph retrieval, agent invocation, and marketing action generation.

[0034] The candidate subject partial subgraph refers to partial graph data related to this customer acquisition task, extracted from the compliance profile knowledge graph using target industry nodes, target product nodes, target scenario nodes, or existing high-quality customer nodes as anchor points. This candidate subject partial subgraph can include subject nodes, demand relationships, product adaptation relationships, upstream and downstream industry chain relationships, transaction collaboration relationships, public event relationships, channel relationships, rule nodes, and evidence nodes. The extraction of the partial subgraph is not simply a matter of extracting all adjacent nodes; rather, it involves filtering based on authorization status, sensitivity level, prohibited items, and reachability status during the extraction process.

[0035] A candidate circle refers to a set of potential target entities generated based on a local subgraph of the candidate entity. A candidate circle can include information such as candidate entities, demand intensity, relationship paths, product suitability, reachability, compliance risks, and execution priority. The candidate circle reflects both marketing value and compliance feasibility.

[0036] A multi-agent task orchestration graph refers to a graph in which multiple agents are orchestrated into a task graph with dependencies, based on tasks such as clue discovery, value assessment, compliance review, channel communication, and verification. A multi-agent task orchestration graph can adopt a directed acyclic graph structure or a workflow structure that includes conditional branches, rollback nodes, and manual verification nodes.

[0037] A local access control view is a controlled access view formed by the server after pruning nodes, edges, attribute fields, and tool interfaces in the local subgraph of a candidate agent based on the agent's identity and task type. Local access control views can anonymize, mask, or summarize unnecessary attributes, and bind the access control view summary to the agent's identity, task node number, and local subgraph version. Local access control views are used to solve the problem of excessively large data access scope during multi-agent execution.

[0038] Candidate marketing actions refer to marketing actions that the system is prepared to execute on a target audience. Candidate marketing actions may include fields such as target audience, recommended product, reach channel, reach time, reach content type, execution priority, and the source agent. Candidate marketing actions are not allowed to enter subsequent task nodes or be actually reached and executed before obtaining a valid compliance pass token.

[0039] A compliance pass token is an action-level execution credential generated by a compliance review agent based on the scope of authorization, sensitivity level, prohibited items, access constraints, frequency control limits, and evidence tracing results. A compliance pass token is not a regular login token, nor is it simply a data access credential; rather, it is a compliance gating credential used to control whether candidate marketing actions can continue to flow and be executed.

[0040] A rejection code is a structured reason identifier output by the compliance review agent when it refuses to generate a compliance pass token. Rejection codes include those for missing authorization, expired authorization, sensitive level, prohibited items, frequency control exceeding limits, channel mismatch, time mismatch, missing evidence, and manual review. Rejection codes are used both to record the reason for rejection and to trigger the reconstruction of subsequent candidate marketing actions.

[0041] An audit hash chain is a chain-like audit structure that writes graph versions, local permission views, agent calls, candidate marketing actions, compliance access tokens, rejection codes, reconstruction results, manual review results, and outreach feedback in the order of execution. Each subsequent record contains a hash digest of the previous record, thus forming a traceable and tamper-proof execution record.

[0042] II. System Structure

[0043] The industry-specific intelligent customer acquisition marketing system of this invention includes a graph construction module, a candidate circle generation module, a permission orchestration module, a token gating module, a strategy reconstruction module, and an audit update module. Each module can be deployed on the same server or distributed across different servers or containerized services, interacting with data through message queues, service interfaces, or database access interfaces.

[0044] The graph construction module is used to acquire industry data, publicly available enterprise data, authorized customer data, product compatibility data, reach records, and regulatory rules to build a compliance profile knowledge graph. It then writes data sources, authorization scope, evidence numbers, sensitivity levels, prohibited items, and reach constraints into nodes or edges. The graph construction module can include a data access unit, an entity disambiguation unit, a relationship alignment unit, a compliance attribute writing unit, and a graph version management unit. The data access unit accesses business registration information, bidding information, recruitment information, intellectual property information, public opinion information, abnormal business information, historical transaction records, consultation records, subscription preferences, reach permit status, and regulatory rules. The entity disambiguation unit maps different names, historical names, abbreviations, unified social credit codes, or internal customer numbers of the same entity to a unified entity node. The relationship alignment unit maps data from different sources to a unified relationship type. The compliance attribute writing unit writes authorization scope, authorization validity period, sensitivity levels, prohibited items, allowed reach channels, allowed reach time periods, frequency control thresholds, and evidence numbers into nodes or edges. The graph version management unit generates a graph version number for each graph construction and update.

[0045] The candidate circle generation module uses target industries and product tasks as anchors to extract partial subgraphs of candidate subjects from the compliance profile knowledge graph, and generates candidate circles based on demand relationships, product compatibility relationships, reachability status, and compliance risks. The module can include an anchor point identification unit, a path expansion unit, a compliance filtering unit, a circle scoring unit, and a candidate subject output unit. The anchor point identification unit maps target industries, target products, and target regions to nodes in the graph. The path expansion unit expands candidate subjects along upstream and downstream relationships in the industry chain, similar demand relationships, transaction collaboration relationships, personnel function relationships, regional business circle relationships, and public event relationships. The compliance filtering unit removes or marks subjects with missing authorization, expired authorization, unsubscribed, complained about, high sensitivity levels, or prohibited access. The circle scoring unit calculates the comprehensive circle score of the candidate subjects.

[0046] The permission orchestration module generates a multi-agent task orchestration graph based on clue discovery, value assessment, compliance review, channel scripts, and verification tasks. It also generates a partial permission view for each agent, containing only the nodes, edges, attributes, and tool interfaces required for its task. The permission orchestration module can include a task decomposition unit, an agent registration unit, a permission view generation unit, a tool interface whitelist unit, and a view summary generation unit. The task decomposition unit breaks down industry-specific customer acquisition marketing tasks into multiple sub-tasks. The agent registration unit records agent identity, task type, and available tools. The permission view generation unit trims the partial sub-graph based on task type. The tool interface whitelist unit restricts the external tools that agents can call. The view summary generation unit calculates the partial permission view summary and writes it into the subsequent compliance pass token and audit hash chain.

[0047] The token gating module is used to bind each candidate marketing action with its local subgraph version, agent identity, permission view summary, reach channel, reach time, and frequency control limit, and to generate or refuse to generate a compliance pass token. The token gating module may include an authorization verification unit, a sensitivity level verification unit, a prohibited item verification unit, a reach constraint verification unit, a frequency control verification unit, an evidence traceability verification unit, and a token generation unit. The token generation unit generates a valid compliance pass token only when all verification units pass; if any verification unit fails, the token gating module outputs the corresponding rejection code.

[0048] The strategy restructuring module is used to only allow candidate marketing actions carrying valid compliance tokens, and to downgrade, delay, transfer to manual review, or remove and restructure rejected or conflicting candidate marketing actions. The strategy restructuring module may include a rejection code parsing unit, an alternative channel generation unit, a delayed execution unit, a manual review triggering unit, a conflict arbitration unit, and a removal unit. Through the strategy restructuring module, this invention avoids directly deleting all actions that fail the initial compliance verification, but instead performs controlled restructuring based on the reason for rejection.

[0049] The audit update module outputs the target entity list, contact channels, key points of the communication script, execution order, and audit hash chain. It also updates the compliance profile knowledge graph, permission view generation rules, and agent strategy parameters based on legitimate contact feedback. The audit update module can include a chained recording unit, a feedback update unit, a rule change retrospective unit, and a strategy parameter update unit. The chained recording unit records the complete process of each action from generation to execution or removal. The feedback update unit updates the knowledge graph based on customer responses, appointments, sales, no response, cancellations, or complaints. The rule change retrospective unit re-triggers token gating for unexecuted candidate marketing actions when regulatory rules or internal compliance rules change.

[0050] III. Specific Technical Route for Implementing the Method of the Invention

[0051] like Figure 1 As shown, the method of the present invention includes the following steps.

[0052] 1. Step S1: Obtain industry data, publicly available enterprise data, authorized customer data, product compatibility data, outreach records and regulatory rules, construct a compliance profile knowledge graph, and write data source, authorization scope, evidence number, sensitivity level, prohibited items and outreach constraints into nodes or edges.

[0053] Specifically, the server receives data from multiple data sources. Industry data may include industry classifications, position in the supply chain, procurement scenarios, business events, regional industrial policies, and industry regulatory requirements. Publicly available enterprise data may include business registration information, bidding information, recruitment information, intellectual property information, public opinion information, and information on abnormal business operations. Authorized customer data may include authorized historical transactions, consultation records, subscription preferences, access permission status, and authorization validity period. Product compatibility data may include product categories, applicable industries, applicable scenarios, prohibited industries, service areas, and channel restrictions. Access records may include the time of access, results of access, unsubscription records, complaint records, and transaction records from channels such as SMS, email, telephone, WeChat Work, in-site messages, or manual visits. Regulatory rules may include personal information protection rules, advertising compliance rules, industry regulatory rules, and internal corporate compliance rules.

[0054] The server standardizes the aforementioned multi-source data. For entity names, the server can perform entity disambiguation based on the unified social credit code, full company name, historical name, abbreviation, registered address, legal representative, or internal customer number. For product data, the server can normalize based on product name, product number, product category, applicable scenarios, and prohibited scenarios. For channel data, the server can map phone calls, SMS messages, emails, WeChat Work, in-site messages, and offline visits to unified channel nodes. For rule data, the server can parse text-based regulatory rules or internal compliance rules into rule nodes and their constraint edges.

[0055] like Figure 2 As shown, the compliance profile knowledge graph includes at least subject nodes, industry nodes, product nodes, channel nodes, rule nodes, evidence nodes, and action nodes. Subject nodes can represent enterprises, organizational clients, authorized contacts, or client accounts. Industry nodes can represent industry categories, supply chain links, regional industrial clusters, or business scenarios. Product nodes can represent product categories, specific products, service solutions, or marketing task products. Channel nodes can represent telephone, SMS, email, WeChat Work, in-app messages, or manual visits. Rule nodes can represent authorization rules, frequency control rules, prohibited contact rules, sensitivity level rules, and evidence tracing rules. Evidence nodes can represent authorization records, publicly available information sources, contract numbers, subscription records, contact records, or manual review records. Action nodes can represent candidate marketing actions, restructured actions, or executed actions.

[0056] When constructing the compliance profile knowledge graph, the server writes data sources, authorization scope, evidence numbers, sensitivity levels, prohibited items, and access constraints into corresponding nodes or edges. For example, the subject node can write the subject source, authorization status, authorization validity period, sensitivity level, unsubscription status, and complaint status; the channel node can write the channel type, allowed access time period, and channel frequency control threshold; the product node can write the applicable industry, prohibited industries, and product compliance restrictions; the rule node can write the rule source, rule effective time, and rule applicable scope; and the evidence node can write the evidence number, evidence source, evidence validity period, and evidence hash digest. Through the above writing method, compliance constraints are no longer manually created rules independent of the marketing graph, but become graph attributes that can be called, verified, and traced by the server.

[0057] 2. S2: Extract partial subgraphs of candidate subjects using target industries and product tasks as anchor points, and generate candidate layers based on demand relationships, product compatibility relationships, reachability status, and compliance risks.

[0058] Specifically, users or upstream business systems submit target industry, target product, target region, marketing task period, and marketing objectives to the server. The server maps the target industry to industry nodes, the target product to product nodes, and the target region to region attributes or region nodes, and uses these nodes as anchor points for local subgraph extraction.

[0059] Starting from the anchor point, the server expands its path along the relationships between upstream and downstream industries, similar needs, transaction collaborations, personnel functions, regional business circles, and public events. Upstream and downstream industry relationships are used to identify suppliers, distributors, customers, or collaborating enterprises in the target industry; similar needs relationships are used to identify entities with similar product needs to existing customers; transaction collaboration relationships are used to identify entities with collaborative relationships in procurement, sales, invoicing, or contracts; personnel functions relationships are used to identify entities with publicly disclosed procurement, operation and maintenance, IT, or financial functions; regional business circle relationships are used to identify entities within the same regional industrial cluster; and public event relationships are used to identify entities that have recently engaged in bidding, capacity expansion, recruitment, financing, equipment procurement, or digital transformation events.

[0060] After path expansion, the server performs filtering and scoring on candidate entities. For entities with missing authorization, expired authorization, unsubscribed, complained about, or triggering prohibited actions, the system does not directly generate proactive marketing actions. For entities with incomplete evidence but potential demand, the system marks them as requiring manual review. For entities with normal reachability and low compliance risk, the system further calculates demand intensity, product suitability, and relationship strength.

[0061] In one alternative implementation, the comprehensive layer score of the candidate entity is calculated using the following formula:

[0062] ;

[0063] in, Indicates the first A comprehensive score for each candidate entity across different circles; Indicates the first The demand intensity of each candidate entity; Indicates the first The degree of fit between each candidate entity and the target product; Indicates the first The strength of the relationship between each candidate entity and the target industry, target product, or existing high-quality customers; Indicates the first Compliance risks of individual candidate entities; , , , These represent the weights of demand intensity, product suitability, relationship strength, and compliance risk, respectively.

[0064] Demand intensity can be determined based on public procurement events, bidding information, job postings, access records, consultation records, and historical transaction similarity; product suitability can be determined based on the product's applicable industry, applicable scenarios, company size, business events, and historical demand tags; relationship strength can be determined based on path length, relationship type, relationship confidence, and number of paths; compliance risk can be determined based on authorization status, sensitivity level, unsubscribe status, complaint status, evidence completeness, and reach restrictions. It should be noted that the above formula is only one implementation method. The server can also use rule-based models, classification models, ranking models, or graph neural network models to obtain candidate layer scores. However, regardless of the method used, reachability status and compliance risk must be used as constraints for candidate layer generation.

[0065] 3. S3 generates a multi-agent task orchestration diagram based on clue discovery, value assessment, compliance review, channel communication, and verification tasks, and generates a local permission view for each agent containing only the nodes, edges, attributes, and tool interfaces required for its tasks.

[0066] The server generates a multi-agent task orchestration diagram based on marketing tasks. The lead discovery agent, located at the front end of the task, identifies candidate entities, relationship paths, and demand leads. The value assessment agent assesses product suitability and marketing value for candidate entities. The compliance review agent verifies authorized scope, sensitivity levels, prohibited items, reach constraints, frequency control limits, and evidence traceability. The channel script agent generates reach channels and key script points within the scope permitted by the compliance pass token. The review agent verifies whether the final output list of target entities, reach channels, key script points, and execution order are consistent with the token status, rejection code, and audit records.

[0067] Instead of providing the complete compliance profile knowledge graph to all agents, the server generates a local permission view for each agent. The generation process for this local permission view includes: first, the server determines the accessible node types, edge types, attribute fields, and tool interfaces based on the agent's identity and task type; second, the server extracts graph fragments related to the agent's task from the candidate agent's local subgraph; third, unnecessary attributes are anonymized, masked, or summarized; finally, a local permission view summary is generated, and this summary is bound to the agent's identity, task node number, and local subgraph version.

[0068] For example, the clue discovery agent can access subject nodes, industry nodes, product nodes, and public relationship paths, but cannot access the original authorized contact number, the full text of the complaint, or the original text of sensitive tags. The value assessment agent can access product compatibility relationships, demand tags, public event summaries, and historical transaction statistics summaries, but cannot access personal contact information beyond the scope of the current task. The compliance review agent can access the authorized scope, evidence number, reach constraints, frequency control limits, and prohibited items. The channel communication agent can only access channels, product selling points, content types, and allowed reach times that have passed compliance pass token verification, and cannot read unauthorized subject attributes. The review agent can read audit summaries, token status, rejection codes, and reconstruction results, but cannot directly read original sensitive information.

[0069] In one implementation, the server can generate a permission vector for each agent to represent its accessible scope:

[0070] Vm={Nm,Em,Am,Tm};

[0071] in, Indicates the first A local permission view for each agent; Indicates the first The set of node types that an agent is allowed to access; Indicates the first The set of edge types that an agent is allowed to access; Indicates the first The set of attribute fields that an agent is allowed to access; Indicates the first A collection of tool interfaces that an agent is allowed to call.

[0072] Through the aforementioned partial permission view, the system enables agents to access only the necessary data when performing their respective tasks, reducing the risk of unauthorized access and allowing subsequent compliance tokens to be bound to the permission view summary, ensuring that the generation process of candidate marketing actions is traceable.

[0073] 4. S4: Bind each candidate marketing action to the corresponding local sub-graph version, agent identity, permission view summary, reach channel, reach time and frequency control limit, and have the compliance review agent generate or refuse to generate a compliance pass token.

[0074] Candidate marketing actions are generated by lead discovery agents, value assessment agents, or channel-based sales script agents during task execution. Each candidate marketing action includes at least the target audience, recommended product, reach channel, reach time, reach content type, and execution priority. Upon receiving a candidate marketing action, the server does not allow it to directly proceed to subsequent task nodes. Instead, it first binds the action with the local subgraph version, agent identity, permission view summary, reach channel, reach time, and frequency control limit to form an action object to be reviewed.

[0075] The compliance review agent performs multiple checks on the objects to be reviewed. The authorization scope check determines whether the target entity and its related data are permitted for use in the current product task and marketing objectives. The sensitivity level check determines whether the candidate action triggers a sensitivity level requiring blocking, review, or prohibition. The prohibited items check determines whether the entity, product, channel, or content type falls under prohibited items. The reach constraint check determines whether the reach channel and time meet the requirements of allowed reach channels, prohibited reach channels, allowed reach time periods, and prohibited reach time periods. The frequency control check determines whether the number of reach attempts by the same entity, the same channel, or the same marketing task cycle exceeds the threshold. The evidence traceability check determines whether the data relied upon by the candidate marketing action can be traced back to a valid evidence number.

[0076] In one implementation, the token generation condition can be represented by the following logical expression:

[0077] ;

[0078] in, Indicates candidate marketing actions Does it meet the requirements for generating a compliant access token? Indicates candidate marketing actions Whether the authorization scope verification has been passed; Indicates candidate marketing actions Whether it passes the sensitivity level check; Indicates candidate marketing actions Has the disabled action not been triggered? Indicates candidate marketing actions Whether the channels for reaching out meet the reach constraints; Indicates candidate marketing actions Does the time of arrival satisfy the arrival constraint? Indicates candidate marketing actions Does it have sufficient frequency control quota? Indicates candidate marketing actions Does it have a traceable evidence number? Represents the logical AND operation.

[0079] when When true, the compliance review agent generates a compliance pass token. The compliance pass token includes the action number, agent identity, local subgraph version, permission view summary, permitted channel, permitted time, remaining frequency control quota, evidence number, validity period, token status, and prohibited items. The token status can include valid, expired, revoked, and pending review. If the verification is false, the compliance review agent refuses to generate a compliance pass token and outputs the corresponding rejection code. If multiple verification items fail simultaneously, multiple rejection codes can be output, and the primary rejection code will be determined according to a preset priority.

[0080] 5. S5: Only candidate marketing actions carrying valid compliance tokens are allowed to enter subsequent task nodes. For rejected or conflicting candidate marketing actions, downgrade, delay, transfer to manual review, or remove and reconstruct based on the rejection code.

[0081] Before a candidate marketing action enters the channel script intelligent agent, review intelligent agent, or execution interface, the server checks whether it carries a valid compliance pass token. Candidate marketing actions that do not carry a token, have an expired token, have a revoked token, or whose token does not match the current action parameters cannot enter subsequent task nodes.

[0082] For rejected candidate marketing actions, the strategy refactoring module performs different actions based on the rejection code. When the rejection code is a channel mismatch rejection code, the server selects an alternative channel from the set of allowed channels and resubmits it for compliance review by the intelligent agent. For example, if the original action intended to reach via phone, but the entity only authorized email, the system will replace the channel with email. If the rejection code is a time mismatch rejection code, the server will delay the action to the allowed reach period and update the execution order. If the rejection code is a frequency control limit exceeding rejection code, the server will lower the action's execution priority or postpone it to the next frequency control cycle. If the rejection code is a missing evidence rejection code or a manual review rejection code, the server will transfer the action to the manual review queue and prohibit automatic execution. If the rejection code is a missing authorization rejection code, expired authorization rejection code, sensitive level rejection code, or prohibited item rejection code, the server will remove the candidate marketing action and write the reason for removal into the audit hash chain.

[0083] When multiple candidate marketing actions conflict, the strategy reconstruction module arbitrates based on compliance risk, evidence completeness, frequency control limits, and marketing value. For example, if the same entity has only one reach quota remaining in the same period, but different agents generate email and telephone reach actions respectively, the system prioritizes retaining actions with low compliance risk, complete evidence, clear channel authorization, and high value, while the remaining actions are delayed or eliminated. Through this mechanism, this invention can not only intercept non-compliant actions but also reconstruct some repairable actions, thereby reducing the loss of executable leads caused by simple deletion.

[0084] 6. S6 outputs the target entity list, contact channels, key points of the communication script, execution order and audit hash chain, and updates the compliance profile knowledge graph, permission view generation rules and intelligent agent strategy parameters based on the feedback of legitimate contact.

[0085] The audit update module generates a chain of audit logs for each candidate marketing action. These logs include graph version records, local permission view records, agent call records, candidate marketing action records, compliance token records, rejection code records, refactoring records, manual review records, and outreach feedback records. Each subsequent record contains a hash digest of the preceding record, allowing auditors to trace back from the final output to the graph version, permission view, agent identity, candidate action source, token status, rejection reason, and refactoring process.

[0086] In one implementation, the audit hash chain is... The hash digest of a record can be generated as follows:

[0087] ;

[0088] in, Indicates the first A hash digest of each audit record; Indicates the first A hash digest of each audit record; Indicates the first The contents of the audit record; Indicates the first The time when the audit record was generated; Indicates the generation or triggering of the first The identity of the agent in each audit record; This represents a hash operation.

[0089] After the outreach is executed, the server updates the compliance profile knowledge graph, permission view generation rules, and agent strategy parameters based on the legitimate outreach feedback. Legitimate outreach feedback includes customer proactive responses, appointments, transactions, no response, cancellations, and complaint results. If the entity proactively responds or completes a transaction, the server can enhance the entity's demand relationship and product compatibility relationship; if the entity does not respond, the server can reduce the corresponding channel preference weight; if the entity cancels or complains, the server writes the cancellation or complaint status to the entity node or channel edge and restricts or prohibits proactive outreach in subsequent tasks. If regulatory rules or internal enterprise compliance rules change, the server retrospectively identifies affected unexecuted candidate marketing actions and re-executes compliance pass token verification. Through feedback updates and rule change retrospectives, this invention forms a closed loop from action generation, token gating, strategy reconstruction, actual outreach to graph updates.

[0090] IV. Specific Implementation Examples, Comparative Examples, and Experimental Data

[0091] To further illustrate the technical effects of this invention, the applicant built an industry-specific intelligent customer acquisition and marketing verification system in a laboratory-isolated environment. This verification system includes an application server, a graph database server, a relational database server, a multi-agent scheduling service, and an audit logging service. The data used in the experiment consisted of anonymized enterprise entity data, publicly available industry data, product compatibility data, authorization status data, outreach record data, and compliance rule data. No real outbound calls, real SMS messages, or real emails were sent; all outreach executions were recorded using simulated execution interfaces to avoid impacting real customers. This experiment successfully reproduced the technical effects of this invention in candidate entity identification, agent orchestration, permission view control, compliance access token gating, rejection code-driven reconstruction, and audit traceability.

[0092] This experiment follows Figure 1 The overall process is executed as shown. First, the system follows... Figure 2 The compliance profiling knowledge graph structure shown incorporates subject nodes, industry nodes, product nodes, channel nodes, rule nodes, evidence nodes, and action nodes into a graph database. Data source, authorization scope, evidence number, sensitivity level, prohibited items, and reach constraints are written into nodes or edges. Next, the system extracts candidate subject subgraphs using target industry nodes and target product nodes as anchor points, generating candidate layers. Then, the system's multi-agent task orchestration graph sequentially calls the clue discovery agent, value assessment agent, compliance review agent, channel communication agent, and verification agent. Next, the system generates local permission views for different agents, restricting each agent to access only the nodes, edges, attributes, and tool interfaces required for its tasks. Subsequently, the system binds candidate marketing actions to the local subgraph version, agent identity, permission view summary, reach channel, reach time, and frequency control limit, with the compliance review agent generating or rejecting compliance access tokens. For candidate marketing actions that fail verification or conflict with each other, the system uses rejection codes to downgrade, delay, transfer to manual review, or remove and reconstruct them. Finally, the system generates an audit hash chain according to the shown method, and updates the compliance profile knowledge graph, permission view generation rules, and agent policy parameters according to the shown method based on the feedback and rule changes.

[0093] 1. Experimental Dataset

[0094] The experimental dataset includes 2000 candidate entities, 6000 candidate marketing actions, 12800 entity relationship edges, 3200 product adaptation relationships, 4200 historical outreach records, and 950 compliance rules. Of the 6000 candidate marketing actions, 1250 were determined to be interceptable after manual annotation and rule review, while the remaining 4750 were either allowed to execute or could be executed after compliance restructuring. The classification of interceptable actions is shown in Table 1 below.

[0095] Table 1 Classification of Actions to be Intercepted

[0096]

[0097] As shown in Table 1, the sum of 650, 180, 120, 150, 80, and 70 is 1250, which is consistent with the total number of actions to be intercepted as pre-labeled in the experiment. An additional 360 sets of policy conflict actions were set up in the experiment to verify the effectiveness of the rejection code-driven reconstruction and conflict resolution.

[0098] 2. Evaluation Indicators

[0099] To evaluate the technical effects of this invention, the following indicators are set.

[0100] The compliance interception accuracy rate is used to evaluate the system's ability to identify and intercept candidate marketing actions that should not enter subsequent task nodes. It is calculated as follows:

[0101] ;

[0102] in, Indicates the accuracy rate of compliance interception; Indicates the number of actions that should have been intercepted but were correctly blocked; This indicates the total number of actions that should be intercepted in the test set.

[0103] The "Missed Outreach" metric indicates the number of candidate marketing actions that should have been blocked but were not. A lower metric indicates better action-level compliance gating.

[0104] The number of unauthorized attribute accesses indicates the number of times an agent reads node attributes, edge attributes, or tool interfaces beyond its task's authorized scope during task execution. This metric is used to verify the technical effectiveness of the illustrated local permission view.

[0105] The automatic conflict resolution rate is used to evaluate the system's ability to automatically resolve conflicts related to outreach channels, outreach times, frequency control limits, and repeated outreach. Its calculation method is as follows:

[0106] ;

[0107] in, Indicates the automatic resolution rate of strategy conflicts; This indicates the number of policy conflicts that were automatically resolved by the system. This represents the total number of policy conflicts in the test set.

[0108] The number of executable actions retained represents the number of actions that are ultimately retained as executable by the system out of 4750 actions that are allowed to be executed or executed after compliant refactoring. This metric illustrates that the present invention does not simply delete all risky actions, but rather retains repairable actions through rejection code-driven refactoring.

[0109] Audit traceability completeness rate is used to evaluate whether the system can completely trace the graph version, local permission view, agent call, token status, rejection code, reconstruction results, and feedback records of candidate marketing actions. Its calculation method is as follows:

[0110] ;

[0111] in, Indicates the audit traceability completeness rate; This indicates the number of candidate marketing actions with a complete audit chain record; This indicates the total number of candidate marketing actions.

[0112] 3. Examples

[0113] Example 1 adopts the complete technical solution of the present invention. Specifically, the system follows... Figure 2 Construct a compliance profile knowledge graph, writing authorization scope, evidence number, sensitivity level, prohibited items, and reach constraints into nodes or edges; extract partial subgraphs of candidate subjects and generate candidate layers; generate task orchestration diagrams for clue discovery, value assessment, compliance review, channel scripts, and review agents; generate partial permission views for each agent; generate or reject compliance access tokens for each candidate marketing action; perform downgrading, delay, transfer to manual review, or elimination and reconstruction for rejected or conflicting actions; generate an audit hash chain for feedback updates and rule change backtracking.

[0114] In Example 1, if a candidate marketing action is rejected due to channel mismatch, the system does not directly delete the action, but instead selects a new channel based on the target entity's allowed channels and resubmits it for verification by the compliance review agent. If a candidate marketing action is rejected due to time mismatch, the system delays it to the allowed time period. If a candidate marketing action is rejected due to frequency control exceeding limits, the system postpones it to the next frequency control cycle or lowers its execution priority. If a candidate marketing action is rejected due to missing authorization, expired authorization, triggering of sensitive levels, or triggering of prohibited items, the system directly removes the action and writes it into the audit hash chain.

[0115] 4. Comparative Example

[0116] Comparative Example 1 uses a standard knowledge graph-based customer acquisition solution. This solution generates a candidate subject list based solely on subject relationships, industry relationships, and product compatibility relationships. It does not generate a partial permission view or a compliance pass token, and only performs rule filtering once before the final list is output. This comparative example illustrates that a standard knowledge graph-based customer acquisition solution cannot perform action-level compliance gating during the action generation and task flow stages.

[0117] Comparative Example 2 employs a knowledge graph and multi-agent orchestration scheme. This scheme uses a similar structure with a clue discovery agent, a value assessment agent, and a channel-based communication agent, but it does not generate a local permission view, nor does it bind a compliance access token to each candidate marketing action. This comparative example illustrates that multi-agent orchestration alone cannot solve the problems of unauthorized agent access and action-level compliance gating.

[0118] Comparative Example 3 employs a compliance token gating scheme, but does not use the rejection code-driven reconstruction mechanism shown. This scheme directly rejects any actions that fail the compliance review, without channel replacement, time delay, frequency control cycle adjustment, or manual review. This comparative example illustrates that while simple blocking can improve compliance, it can result in the loss of executable actions.

[0119] Comparative Example 4 employs a token gating and rejection code reconstruction scheme, but does not use the audit hash chain shown; it only stores ordinary operation logs. This comparative example illustrates that ordinary log records are insufficient to fully trace graph versions, local permission views, agent calls, token states, and reconstruction results.

[0120] 5. Experimental Results

[0121] Table 2 Comparison of Key Performance of Different Solutions

[0122]

[0123] The original counts for each percentage in Table 2 are as follows: Example 1: 1228 / 1250, 338 / 360, 6000 / 6000; Comparative Example 1: 780 / 1250, 126 / 360, 4272 / 6000; Comparative Example 2: 1035 / 1250, 216 / 360, 4716 / 6000; Comparative Example 3: 1240 / 1250, 0 / 360, 5790 / 6000; Comparative Example 4: 1226 / 1250, 335 / 360, 4944 / 6000

[0124] As shown in Table 2, Example 1 demonstrates comprehensive advantages in compliance interception accuracy, unauthorized access control, automatic policy conflict resolution, retention of executable actions, and audit traceability completeness. Specifically, Example 1 achieves a compliance interception accuracy of 98.24%, with only 22 missed violations, lower than Comparative Example 1 and Comparative Example 2; the number of unauthorized accesses is 0, indicating that the local permission view effectively limits the agent's access scope; the automatic policy conflict resolution rate is 93.89%, significantly higher than Comparative Example 1 and Comparative Example 2; the number of retained executable actions is 4686, higher than Comparative Example 3, indicating that rejection code-driven reconstruction avoids the loss of executable actions caused by simple removal; and the audit traceability completeness rate is 100.00%, higher than Comparative Example 4, indicating that the audit hash chain can completely record the graph version, permission view, agent call, token state, rejection code, and reconstruction result.

[0125] The foregoing description of embodiments of the present invention, through which those skilled in the art are able to implement or use the present invention, will be readily apparent to those skilled in the art. Various modifications to these embodiments will be readily apparent to those skilled in the art. The general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novelty disclosed herein.

Claims

1. An industry-specific intelligent customer acquisition marketing method based on compliance profile knowledge graph and multi-agent strategy orchestration, characterized in that, Includes the following steps: S1 acquires industry data, publicly available enterprise data, authorized customer data, product compatibility data, outreach records and regulatory rules, constructs a compliance profile knowledge graph, and writes data source, authorization scope, evidence number, sensitivity level, prohibited items and outreach constraints into nodes or edges; S2 extracts partial subgraphs of candidate subjects using target industries and product tasks as anchor points, and generates candidate layers based on demand relationships, product compatibility relationships, reachability status and compliance risks. S3 generates a multi-agent task orchestration diagram based on clue discovery, value assessment, compliance review, channel communication, and verification tasks, and generates a local permission view for each agent containing only the nodes, edges, attributes, and tool interfaces required for its tasks. S4 binds each candidate marketing action with the corresponding local subgraph version, agent identity, permission view summary, reach channel, reach time and frequency control limit, and the compliance review agent generates or refuses to generate a compliance pass token. S5 allows only candidate marketing actions carrying valid compliance tokens to enter subsequent task nodes. For rejected or conflicting candidate marketing actions, downgrade, delay, transfer to manual review, or elimination and reconstruction are performed based on the rejection code. The rejection code refers to the structured reason identifier output by the compliance review agent when it refuses to generate a compliance token. S6 outputs the target subject list, reach channels, key points of the script, execution order and audit hash chain, and updates the subject demand relationship, product adaptation relationship, channel preference relationship, frequency control status, permission view generation rules and intelligent agent strategy parameters based on the legitimate reach feedback.

2. The method according to claim 1, characterized in that, In step S1, the industry data includes industry classification, position in the industrial chain, procurement scenarios, business events, regional industrial policies, and industry regulatory requirements; the publicly available enterprise data includes business registration information, bidding information, recruitment information, intellectual property information, public opinion information, and abnormal business information; and the authorized customer data includes authorized historical transactions, consultation records, subscription preferences, access permission status, and authorization validity period.

3. The method according to claim 1, characterized in that, The reach constraints include allowed reach channels, prohibited reach channels, allowed reach time periods, prohibited reach time periods, single entity reach frequency control, same channel reach frequency control, unsubscribe status, complaint status, and manual review status; when the entity node is in unsubscribe status, complaint status, or authorization invalidation status, it is prohibited to generate proactive reach compliance tokens for that entity.

4. The method according to claim 1, characterized in that, In step S2, the extraction of candidate subject local subgraphs includes: starting with target industry nodes, target product nodes, and existing high-quality customer nodes, expanding the path along the upstream and downstream relationships of the industrial chain, similar demand relationships, transaction collaboration relationships, personnel function relationships, regional business circle relationships, and public event relationships, and performing authorization filtering, sensitivity level filtering, and reachability status filtering on the expansion results.

5. The method according to claim 1, characterized in that, In step S3, generating the local permission view includes: determining the allowed node types, edge types, attribute fields, and tool interfaces based on the agent's task type; performing desensitization, masking, or summarization on unnecessary attributes; and binding the local permission view with the agent's identity, task node number, and local subgraph version, so that the agent cannot call graph attributes and external tools that are unrelated to the current task.

6. The method according to claim 1, characterized in that, In step S4, the candidate marketing action includes the target subject, recommended product, reach channel, reach time, reach content type, and execution priority; the compliance pass token includes action number, agent identity, local subgraph version, permission view summary, allowed channel, allowed time, frequency control remaining quota, evidence number, validity period, token status, and prohibited items; the rejection code includes authorization missing rejection code, authorization expired rejection code, sensitive level rejection code, prohibited item rejection code, frequency control over-limit rejection code, channel mismatch rejection code, time mismatch rejection code, evidence missing rejection code, and manual review rejection code; each rejection code is written into the audit hash chain and serves as the trigger condition for subsequent candidate marketing action reconstruction.

7. The method according to claim 1, characterized in that, In step S5, the refactoring of rejected or conflicting candidate marketing actions includes: when the rejection code is a channel mismatch rejection code, replacing it with an allowed channel; when the rejection code is a time mismatch rejection code, delaying it until the allowed time; when the rejection code is a frequency control over-limit rejection code, lowering the execution priority or postponing execution; when the rejection code is a missing evidence rejection code or a manual review rejection code, transferring it to manual review; when the rejection code is a missing authorization rejection code, an expired authorization rejection code, a sensitive level rejection code, or a prohibited item rejection code, removing the candidate marketing action.

8. The method according to claim 1, characterized in that, In step S6, the audit hash chain includes graph version records, local permission view records, agent call records, candidate marketing action records, compliance access token records, rejection code records, reconstruction records, manual review records, and outreach feedback records; each subsequent record contains the hash digest of the previous record to form a traceable chain audit structure.

9. An industry-specific intelligent customer acquisition and marketing system based on compliance profile knowledge graph and multi-agent strategy orchestration, the system being used to implement the method described in any one of claims 1-8, characterized in that, include: The graph construction module is used to acquire industry data, publicly available enterprise data, authorized customer data, product adaptation data, outreach records, and regulatory rules to build a compliance profile knowledge graph that includes data sources, authorization scope, evidence numbers, sensitivity levels, prohibited items, and outreach constraints. The candidate layer generation module extracts local subgraphs of candidate subjects using target industry and product tasks as anchor points and generates candidate layers. The permission orchestration module generates a multi-agent task orchestration graph and creates a local permission view for each agent, containing only the nodes, edges, attributes, and tool interfaces required for its task. The token gating module is used to gate each agent... A candidate marketing action is bound to a local subgraph version, agent identity, permission view summary, reach channel, reach time, and frequency control limit, and generates or rejects the generation of a compliance pass token; the strategy reconstruction module is used to only allow candidate marketing actions carrying valid compliance pass tokens, and to downgrade, delay, transfer to manual review, or remove and reconstruct candidate marketing actions that are rejected or conflict with each other; the audit update module is used to output the target subject list, reach channel, key points of the message, execution order, and audit hash chain, and to update the compliance profile knowledge graph, permission view generation rules, and agent strategy parameters based on the feedback from legitimate reach.

10. The system according to claim 9, characterized in that, The token gating module includes an authorization verification unit, a sensitivity level verification unit, a prohibited item verification unit, an access constraint verification unit, a frequency control verification unit, an evidence traceability verification unit, and a token generation unit. The token generation unit generates a compliance access token when all verification units pass, and outputs a corresponding rejection code when any verification unit fails. The audit update module includes a chain recording unit, a rule change backtracking unit, and a feedback update unit. The chain recording unit writes the graph version, local permission view, agent call, compliance access token, rejection code, and reconstruction result into the audit hash chain. The rule change backtracking unit re-triggers the token gating module for candidate marketing actions that have not been executed when regulatory rules or internal enterprise compliance rules change.

Images