A smart scoring and differentiated routing system for telemarketing leads that integrates customer behavior profiles

Abstract

This invention discloses an intelligent scoring and differentiated routing scheduling system for telemarketing leads that integrates customer behavior profiling, belonging to the field of telemarketing lead management technology. The system includes modules for data collection and two-way profiling, dynamic scoring, collaborative routing scheduling, real-time assistance, and closed-loop self-optimization. It constructs a heterogeneous two-way knowledge graph for customer and sales, introduces a causal inference model to generate interpretable causal enhancement potential scores, and employs hierarchical multi-agent game decision-making to achieve collaborative scheduling of routing and communication strategies. During calls, it integrates voice and semantic features in real time to perceive customer emotions and dynamically adjust strategies. Through counterfactual attribution and digital twin simulation, it completes full-link closed-loop self-optimization. This invention effectively solves the problems of traditional scoring bias, disconnect between routing and strategy, and high optimization trial-and-error costs, significantly improving lead conversion rates and the utilization rate of top sales resources, while reducing customer complaints and churn risks.

Description

Technical Field

[0001] This invention relates to the field of telemarketing lead scoring and scheduling technology, specifically a telemarketing lead intelligent scoring and differentiated routing scheduling system that integrates customer behavior profiles. Background Technology

[0002] With the deep application of big data and artificial intelligence technologies in customer relationship management, data-driven lead scoring and routing systems for telemarketing have become key technologies for improving marketing efficiency. In existing technologies, lead scoring models are typically based on machine learning or deep learning algorithms, extracting features from dimensions such as customers' historical interaction behavior and demographic characteristics to output a customer's purchase intention probability or conversion potential score. Building on this, some systems further introduce dynamic rating mechanisms, updating the scoring results in real time based on changes in customer behavior data. Regarding routing, existing technologies have evolved from fixed-rule allocation to intelligent matching schemes that consider multi-dimensional factors such as salesperson ability level and availability. Some systems employ reinforcement learning or game theory methods for optimal lead and agent allocation. Furthermore, closed-loop feedback optimization mechanisms have been introduced, using telemarketing conversion result data to iteratively update the scoring model or allocation strategy.

[0003] However, the above-mentioned technical solutions still have significant shortcomings in terms of the causal explanatoryness of the scoring model, the strategic synergy of routing decisions, and the risk control of strategy optimization.

[0004] First, existing scoring models generally belong to the correlation-driven prediction paradigm, which can identify which features are statistically associated with high conversion rates, but cannot distinguish between correlation and causation. For example, the system may misjudge a customer's browsing of the top sales page as a high intention signal, when in fact this behavior is an exposure bias caused by the system's recommendation algorithm, not a true reflection of the customer's active intention. This confusion between correlation and causation leads to insufficient accuracy and interpretability of the scoring results.

[0005] Secondly, although the existing routing and scheduling system has achieved intelligent matching of leads and salespeople, the matching decision and subsequent communication strategy are disconnected. After the system decides which salesperson to call which customer, the salesperson's personal statements during the call rely entirely on personal experience, resulting in the loss of high-value leads due to inappropriate communication strategies.

[0006] Third, the closed-loop optimization of the existing system relies on feedback from real business data. Any strategy adjustment requires trial and error verification in a real environment, which is costly, has a long optimization cycle, and lacks a mechanism for conducting security simulations before the strategy goes live. In addition, the existing system lacks the ability to perceive and dynamically respond to changes in customer emotions and intentions during calls, and cannot form a real-time command closed loop of human-machine collaboration at the communication level. Summary of the Invention

[0007] (a) Technical problems to be solved

[0008] To address the shortcomings of existing technologies, this invention provides an intelligent scoring and differentiated routing scheduling system for telemarketing leads that integrates customer behavior profiles, thus solving the problems mentioned in the background section.

[0009] (II) Technical Solution

[0010] To achieve the above objectives, the present invention is implemented through the following technical solution: a smart scoring and differentiated routing scheduling system for telemarketing leads that integrates customer behavior profiles, including a data collection and bidirectional profiling module, used to collect multimodal customer behavior data and multidimensional salesperson capability data, respectively construct dynamic customer behavior profiles and dynamic salesperson capability profiles, and construct a bidirectional heterogeneous knowledge graph of customers and salespersons based on the two.

[0011] The dynamic scoring module is used to identify customer conversion tendencies and strategy sensitivity based on the customer dynamic behavior profile, and to generate interpretable causal enhancement potential scores.

[0012] The routing and scheduling module is used to perform lead layering and resource isolation based on the causal enhancement potential score, and adopts a hierarchical multi-agent game decision architecture. The high-level agent performs the global optimal matching of leads and salespeople, and the low-level agent pre-generates personalized communication strategies and outbound call timing decisions before making outbound calls, so as to realize differentiated routing and strategy collaborative scheduling.

[0013] The real-time support module is used to analyze changes in customer emotions and intentions in real time during the call, dynamically modify the pre-generated communication strategy and push it to the salesperson's terminal, forming a real-time command mechanism for human-machine collaboration.

[0014] The closed-loop self-optimization module is used to collect telemarketing conversion results and call process data, and feed back to update the model parameters and strategy network of each module to form a continuously self-optimizing closed loop.

[0015] Furthermore, in the data acquisition and bidirectional profiling module:

[0016] The customer dynamic behavior profile includes four sub-profile dimensions: conversion intention time sequence characteristics, decision resistance characteristics, communication preference characteristics, and price sensitivity characteristics; the customer behavior data is collected from online behavior data, offline interaction data, historical telemarketing data, and third-party compliant data.

[0017] The salesperson's dynamic capability profile includes five sub-profile dimensions: industry category proficiency profile, communication style profile, customer compatibility profile, emotional control ability profile, and idle time slot load profile. The customer-sales bidirectional heterogeneous knowledge graph includes three types of nodes: customer nodes, salesperson nodes, and product nodes. The weights of the edges in the graph decay over time to reflect the freshness of the behavior, and the tag weights are adaptively updated hourly based on each telemarketing conversion result.

[0018] Furthermore, the dynamic scoring module further includes: The causal graph construction unit builds a structural causal model based on historical telemarketing data, explicitly models the causal relationship graph between customer characteristics, salesperson characteristics, communication strategies and conversion results, identifies and eliminates confounding factors caused by exposure bias and selection bias, and obtains bias-free causal effect estimates through backdoor adjustment. The counterfactual reasoning unit calculates the average conditional treatment effect for each clue, quantifies the causal transformation increment of different intervention strategies for that clue, and outputs strategy sensitivity labels. The dual-tower scoring unit adopts a dual-tower deep learning model architecture. One tower predicts the probability of a customer's natural purchase intention without any specific communication strategy intervention, while the other tower predicts the risk of customer churn or complaints. The strategy sensitivity labels are integrated to output a comprehensive causal enhancement potential score, and attribution explanations for the score are also output, indicating the contribution of each feature dimension to the score. The causal enhancement potential score Defined by the following formula: ; In the formula, This represents the probability of a customer's natural purchase intention without intervention. The hyperparameter representing the adjustment of policy gain. This represents the maximum average conditional treatment effect for this customer among all alternative communication strategies.

[0019] Furthermore, the routing coordination scheduling module further includes: The resource isolation unit categorizes leads into three levels—S, A, and B—based on the causal enhancement potential score, and sets up three-level lead resource isolation pools. S-level leads are rigidly locked and only allowed to flow into the exclusive service pool of top salespeople, prohibiting their allocation to ordinary and junior salespeople. At the same time, a maximum service quota of S-level leads is set for each top salesperson, as well as a load limit threshold set based on the number of S-level leads currently being served. When the quota is exhausted or the load threshold is reached, S-level leads are automatically and temporarily cached, and prioritized for allocation when the salesperson is idle. The game-theoretic decision-making unit adopts a multi-agent reinforcement learning architecture with centralized training and distributed execution. The high-level routing strategy network takes the optimization of multiple objectives, including global conversion rate, utilization of top-tier resources, customer satisfaction, and agent load balancing, as its task to find the optimal matching scheme between leads and salespeople. The low-level strategy generation network generates personalized communication strategies and outbound call timing decisions for specific customers and salespeople based on the aforementioned heterogeneous customer-sales bidirectional knowledge graph and customer causal profile. The high-level and low-level networks perform end-to-end collaborative optimization through a joint reward function to ensure that routing decisions and strategy generation are coordinated and consistent under the global objectives.

[0020] Furthermore, the high-level routing policy network adopts the QMIX algorithm, models each salesperson as an independent agent, decomposes the global value function into the local value function of each agent through a hybrid network, achieves global optimal cooperation under the condition that only local observation is required, and introduces an attention mechanism to dynamically adjust the weight coefficients of each constraint. The low-level strategy generation network is built on a pre-trained large language model. It uses the four-dimensional profile of the customer, the five-dimensional profile of the salesperson, and the strategy sensitivity labels output by the counterfactual reasoning unit as prompts and inputs to generate structured communication strategy texts that include suggestions on speaking style, ranking of key selling points, and risk avoidance prompts. The texts are then pushed to the salesperson's terminal along with the task. The optimization objective of the game decision-making unit architecture is to minimize the overall joint loss function. Defined by the following formula: ; In the formula, The loss function of the high-level routing policy network represents the time difference error between the joint action value function calculated based on the QMIX hybrid network and the target value. The target value is calculated using multi-step rewards and discounted cumulative rewards. This represents the loss function of the low-level policy generation network, and the optimization objective for generating communication strategies based on a large language model. This represents a dynamic tradeoff coefficient used to balance routing efficiency and policy quality.

[0021] Furthermore, the real-time assistance module further includes: fusing voice acoustic features and text semantic features in real time during the call to identify the customer's current emotional state and predict the trend of emotional change, and constructing a time curve of the customer's emotions; when a deterioration in the customer's emotions or a drift in intent is detected, updating the communication strategy immediately and providing a silent prompt through the salesperson's terminal; when a clear rejection or complaint signal is detected from the customer, sending a negative feedback signal to the dynamic scoring module to trigger a scoring penalty mechanism and reduce the potential score of the lead.

[0022] Furthermore, the optimization process of the closed-loop self-optimization module includes: after each call ends, analyzing the underlying reasons for the success or failure of the conversion based on the counterfactual inference framework, storing the attribution results in a structured manner and feeding them back to the dynamic scoring module, and updating the edge weights and conditional probability distributions in the structural causal model. When new scoring rules, routing strategies, or quota settings are officially launched, a large-scale Monte Carlo simulation is conducted in a virtual simulation environment built based on historical telemarketing data. The simulation outputs a quantitative evaluation report that includes expected conversion rate, resource utilization rate, and fairness indicators. After verification and manual confirmation, the simulation is automatically deployed to the production environment. At the same time, the deviation between the launch effect and the simulation prediction is monitored in real time. When the deviation exceeds the preset threshold, the strategy rollback is automatically triggered.

[0023] Furthermore, the dynamic scoring module also introduces a key behavioral event triggering mechanism. When a customer triggers a preset key behavioral event, the causal enhancement potential score is immediately recalculated, achieving real-time score updates within seconds. The preset key behavioral events include accessing the pricing page, submitting a demo application, and downloading the product white paper.

[0024] Furthermore, the routing cooperative scheduling module also includes a routing conflict resolution mechanism for abnormal scenarios: When multiple top salespeople are simultaneously available and vying for the same S-level lead, the system automatically arbitrates the case based on a weighted score that considers factors such as historical conversion rate of similar leads, geographic suitability, and available time. The lead is then assigned to the salesperson with the highest weighted overall priority. If a top salesperson who has been assigned an S-level lead suddenly goes offline or loses connection, the system temporarily locks the lead and does not reassign it. The lead will be reassigned as soon as the salesperson becomes online again. If the lead is not reassigned after a preset waiting period, a rerouting is triggered.

[0025] (III) Beneficial Effects (1) By adopting causal inference and counterfactual reasoning to reconstruct the clue scoring mechanism, we can accurately identify and eliminate confusing factors such as exposure bias and selection bias, upgrade from correlation prediction to causal assessment, and the scoring has both high accuracy and strong interpretability. At the same time, we output the scoring attribution explanation, intuitively present the contribution of each feature to the score, break through the limitations of traditional models that are easy to confuse correlation and causation and score distortion, and make the quantification of clue value more scientific and the scoring results more forward-looking and credible.

[0026] (2) By constructing a hierarchical multi-agent game routing architecture and combining a customer-sales bidirectional heterogeneous knowledge graph, the global optimal matching of leads and salespeople is achieved. The routing decision and personalized communication strategy are integrated and coordinated to solve the industry pain point of the separation between routing and strategy. With rigid isolation of S-level leads and dual protection mechanisms of quota and load for top salespeople, the high-value leads are strictly guaranteed to be allocated in a targeted manner, maximizing the conversion efficiency of high-potential leads and the utilization rate of high-quality sales resources, while achieving seat load balancing.

[0027] (3) By adding a real-time multimodal emotion perception and dynamic strategy correction mechanism for calls, integrating voice acoustic features and text semantic features, the customer emotion curve is constructed in real time and the trend of emotion change is predicted. Correction strategies are pushed in real time for scenarios such as emotion deterioration and intention drift. When rejection or complaint signals are detected, scoring penalties and lead cooling are triggered to form a real-time command closed loop of human-machine collaboration, effectively reducing the risk of customer complaints and lead loss, and greatly improving the success rate of telemarketing communication and service quality.

[0028] (4) By establishing a closed-loop self-optimization system of counterfactual attribution analysis and digital twin simulation, the core reasons for success or failure of transformation can be located through counterfactual reasoning, and the model and strategy can be accurately iterated. Before the new strategy is launched, a large-scale Monte Carlo simulation verification is completed in the virtual environment, which greatly reduces the trial and error cost of real business. After the launch, the effect deviation is monitored in real time and the rollback is automatically triggered, so as to realize the continuous self-evolution of the scoring model, routing strategy and resource quota, and improve the overall efficiency and conversion effect of telemarketing operations in a long-term and stable manner. Attached Figure Description

[0029] Figure 1 This is an overall system block diagram of the present invention; Figure 2 This is a schematic diagram of the causal enhancement dynamic scoring process of the present invention; Figure 3 This is a schematic diagram of the hierarchical multi-agent game-theoretic routing and scheduling architecture of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Please see Figures 1 to 3 As shown, the embodiments of the present invention provide the following technical solutions: This embodiment discloses a telemarketing lead intelligent scoring and differentiated routing scheduling system that integrates customer behavior profiles, specifically including the following modules: 1. Data Acquisition and Two-Way Profiling Module This is used to collect multimodal customer behavior data and multidimensional salesperson capability data, to construct dynamic customer behavior profiles and dynamic salesperson capability profiles respectively, and to construct a bidirectional heterogeneous knowledge graph of customer and sales based on the two.

[0032] Specifically, the customer dynamic behavior profile includes four sub-profile dimensions: conversion intention time-series characteristics, decision-making resistance characteristics, communication preference characteristics, and price sensitivity characteristics. Conversion intention time-series characteristics are characterized by tracking changes in customer behavior density over time, such as the trend of product page browsing frequency over the past 7 days, the growth slope of average dwell time, and the number of days between browsing and actively inquiring about a price. Decision-making resistance characteristics are quantified by analyzing customer interruptions and repetitive behaviors in the decision-making process, such as the number of times the same page is visited multiple times without submitting a form, the proportion of items added to the shopping cart and then abandoned, and outliers in product comparison frequency. Communication preference characteristics are extracted based on natural language processing results from historical call records, including the distribution of customer preferred communication time periods, differences in responses to open-ended and closed-ended questions, and the weighting of attention to technical details versus business terms. Price sensitivity characteristics are measured by the customer's responsiveness to promotional activities, such as the coupon redemption rate in historical outreach, the slope of the conversion probability curve under different discount levels, and the frequency with which customers actively mention prices.

[0033] The customer behavior data is collected from online behavior data, offline interaction data, historical telemarketing data, and third-party compliant data. Online behavior data includes web browsing logs, clickstream logs, and form interaction logs collected through the event tracking SDK; offline interaction data includes structured extraction results of meeting participation records, event check-in records, and offline negotiation minutes obtained through CRM system integration; historical telemarketing data includes call duration, ASR transcription of call recordings, and call sentiment analysis results obtained through CTI system integration; third-party compliant data includes anonymized credit data and industry report feature extraction results obtained through federated learning nodes. The original information of this part of the data does not leave the local machine; only encrypted gradients or feature embedding vectors are transmitted.

[0034] The salesperson's dynamic competency profile comprises five sub-profile dimensions: industry category proficiency profile, communication style profile, customer fit profile, emotional control profile, and idle time slot load profile. The industry category proficiency profile is derived from industry tags aggregated from the salesperson's historical transaction records, using a weighted TF-IDF method to calculate the proficiency score for each industry category, with weighting factors being the reciprocal of the transaction amount and the transaction cycle. The communication style profile is derived from acoustic and linguistic feature clustering of call recordings, mapping the salesperson's speech rate range, average call length, interruption frequency, and proportion of open-ended questions to a communication style space, including dimensions such as consultative, relational, efficient, and guiding. The emotional control profile is evaluated by statistically analyzing the salesperson's response to customer emotional fluctuations, with specific indicators including the average response time for a customer's emotions to shift from negative to positive, the success rate of intercepting escalating customer complaints, and the net improvement in customer emotions at the end of the call.

[0035] The heterogeneous customer-sales bidirectional knowledge graph includes three node types: customer nodes, salesperson nodes, and product nodes. Customer nodes' attributes include the quantified feature vectors of the aforementioned four-dimensional profile; salesperson nodes' attributes include the quantified feature vectors of the aforementioned five-dimensional profile; and product nodes' attributes include product category, price range, applicable industry tags, and typical decision-making cycle. Edge types between nodes include browsing, inquiry, call, transaction, complaint, similar customers, and competing products. The edge weights in the graph use a negative exponential decay function for time decay, and the behavioral freshness is reflected by the following formula: ; in, As the initial weights, The attenuation coefficient is... The time interval from the current time. Indicates the time interval The weight of subsequent edges. For example, the weight of a customer's browsing behavior 7 days ago decays to approximately 50% of its initial value. When the value is 0.1, the weight of behaviors from 30 days ago is only about 5%, thus ensuring the system's sensitivity to recent behavioral changes.

[0036] The label weights are adaptively updated hourly based on each telemarketing conversion result. The specific update mechanism is as follows: when a salesperson's conversion rate for a certain customer profile is consistently higher than the baseline, the adaptation edge weight between the salesperson node and the customer profile node is increased; when a customer's response rate to a certain communication strategy decreases significantly, the label confidence of the corresponding communication preference attribute of the customer node is decreased, thereby achieving dynamic calibration of the profile.

[0037] 2. Dynamic scoring module Based on the aforementioned customer dynamic behavior profile, a causal inference model is introduced to identify customer conversion tendencies and strategy sensitivity, generating an interpretable causal enhancement potential score. Please see Figure 2 The dynamic scoring module further includes: The causal graph construction unit builds a structural causal model based on historical telemarketing data, explicitly modeling the causal relationship graph between customer characteristics, salesperson characteristics, communication strategies, and conversion results. It identifies and eliminates confounding factors caused by exposure bias and selection bias. Exposure bias refers to customers browsing certain product pages not out of active interest, but passively due to system recommendations. Selection bias refers to the system naturally assigning high-value customers to high-level salespeople, distorting the causal relationship between salesperson ability and conversion results due to customer quality as a confounding factor. The causal effect is adjusted through a backdoor: specifically, for each possible value in the set of confounding factors to be controlled, the conditional conversion probability of applying the target strategy under that value is calculated, and then a weighted sum is performed according to the probability of occurrence of that value, thereby eliminating confounding bias and obtaining a debiased causal effect estimate.

[0038] The counterfactual reasoning unit calculates the conditional average treatment effect for each clue, quantifying the causal conversion increment of different intervention strategies for that clue. The calculation of the conditional average treatment effect is based on heterogeneous treatment effect estimation methods such as causal forest or meta-learner: For target customer A and alternative strategy X, firstly, a sample group with similar characteristics to the customer is selected from historical data, and the average conversion rate when strategy X is applied and the average conversion rate when no strategy is applied are calculated in the group respectively. The difference between the two is the estimated value of the conditional average treatment effect of the customer on the strategy. For example, the system can identify that: Customer A has a significant positive average treatment effect on the limited-time offer strategy, and a positive but weak positive average treatment effect on the authoritative endorsement strategy; Customer B has a near-zero average treatment effect on the limited-time offer strategy, and a significant positive average treatment effect on the authoritative endorsement strategy. Based on this, the system outputs strategy sensitivity labels, classifying customers as discount-sensitive, endorsement-sensitive, or weakly sensitive.

[0039] The dual-tower scoring unit employs a dual-tower deep learning model architecture. One tower predicts the probability of a customer's natural purchase intention without any specific communication strategy intervention. It uses a Transformer encoder to perform time-series modeling of customer behavior sequences and outputs the probability of a customer's natural purchase intention without any specific communication strategy intervention. The other tower predicts the risk of customer churn or complaints. It also uses a Transformer architecture, but the training objective is to perform binary classification prediction of customer churn or complaint risks and outputs the risk probability. The two towers share the bottom embedding layer, but the top layers are trained independently.

[0040] By integrating the aforementioned strategy sensitivity labels, a comprehensive causal enhancement potential score is output. Defined by the following formula: ; In the formula, This represents the probability of a customer's natural purchase intention without intervention, and is output by the first tower. This represents the hyperparameter for adjusting the strategy gain; its value range is determined by the business scenario. This represents the maximum conditional average treatment effect for this customer among all alternative communication strategies, which is the maximum value of the conditional average treatment effect of each strategy from the set of alternative strategies.

[0041] The design philosophy of this formula is that a lead's final score depends not only on its inherent purchase intent but also on the existence of a communication strategy that can effectively persuade it. For example, customer A has a natural purchase intention probability of 0.3, indicating moderate intent, but its maximum conditional average treatment effect on limited-time offers is 0.4, meaning there is significant room for strategy improvement. Adjusting the hyperparameter to 1.0, the calculated score is 0.42. In contrast, customer B, while also having a natural purchase intention probability of 0.4, has a maximum conditional average treatment effect of only 0.02, indicating no significantly effective persuasive strategy. The calculated score is approximately 0.41, lower than customer A's 0.42. This example illustrates that even if a customer has a stronger basic purchase intention, if their strategy is less persuasive, their overall score can be surpassed by a customer with lower strategy sensitivity, demonstrating the corrective effect of strategy sensitivity on lead ranking.

[0042] The output also provides an attribution explanation for the score, illustrating the contribution of each feature dimension to the rating. The attribution explanation is calculated based on Shapley values ​​or integral gradient methods, decomposing the score into the marginal contribution of each input feature. For example, the output might show the attribution for a lead score of 0.42: basic purchase intention contribution of 0.3, and strategy improvement space contribution of 0.12, with the main source of strategy improvement being limited-time offer language.

[0043] Furthermore, the dynamic scoring module introduces a key behavioral event triggering mechanism. When a customer triggers a preset key behavioral event, the causal enhancement potential score is immediately recalculated, achieving real-time score updates within seconds. The preset key behavioral events include accessing the pricing page, submitting a demo request, and downloading the product white paper. The triggering mechanism is implemented based on an event stream processing framework, using a message queue to receive event tracking data in real time. The stream computing engine triggers the score recalculation process upon event arrival.

[0044] 3. Routing Coordination and Scheduling Module This is used to perform lead stratification and resource isolation based on the causal enhancement potential score, and adopts a hierarchical multi-agent game decision architecture. The high-level agent performs the global optimal matching of leads and salespeople, while the low-level agent pre-generates personalized communication strategies and outbound call timing decisions before making outbound calls, thereby realizing differentiated routing and strategy collaborative scheduling.

[0045] Please see Figure 3 The routing coordination scheduling module further includes: The resource isolation unit classifies clues into three levels—S, A, and B—based on the causal enhancement potential score. The classification threshold is not a fixed value, but is determined based on the adaptive percentile of the score distribution of all clues in the current clue pool; for example, clues with the current score in the top 10% are classified as S-level, 10%-50% as A-level, and the rest as B-level. The threshold is dynamically adjusted with each score update.

[0046] A three-tiered lead resource isolation pool is set up, in which S-level leads are rigidly locked and only allowed to flow into the exclusive service pool of top salespeople, and are prohibited from being assigned to ordinary and junior salespeople. Rigid locking means that even if ordinary salespeople are completely idle and top salespeople are all busy, S-level leads will not be downgraded to ordinary salespeople, but will enter a temporary cache queue to wait.

[0047] At the same time, a maximum service quota for S-level leads is set for each top salesperson, as well as a load limit threshold set based on the number of S-level leads currently being served. When the quota is exhausted or the load threshold is reached, S-level leads are automatically cached temporarily and allocated to the salesperson when they are available. The quota value is not fixed, but is dynamically adjusted by a closed-loop self-optimization module based on the salesperson's recent S-level lead conversion rate and average processing time. That is, salespeople with high conversion rates and high processing efficiency automatically receive higher quota limits.

[0048] The game-theoretic decision-making unit employs a multi-agent reinforcement learning architecture with centralized training and distributed execution. The high-level routing policy network focuses on multi-objective optimization tasks, including global conversion rate, premium resource utilization, customer satisfaction, and agent load balancing. These multi-objectives are transformed into a single-objective optimization problem through weighted summation. The initial weights of each objective are set by business experts and dynamically adjusted during training using an attention mechanism. Specifically, when the system detects low premium resource utilization, it automatically increases the weight of that objective; similarly, when the customer complaint rate rises, it automatically increases the weight of the customer satisfaction objective.

[0049] Specifically, the high-level routing policy network employs the QMIX algorithm. QMIX is a multi-agent reinforcement learning algorithm based on value function decomposition. Its core idea is to decompose the global joint action value function into the local value functions of each sales agent, and to achieve decomposition and combination through a hybrid network. The key constraint of the hybrid network is to ensure that the joint value function satisfies the monotonicity condition with respect to each local value function. This constraint is achieved by limiting the weights of the hybrid network to non-negative values. The monotonicity constraint ensures that the result of each agent making greedy action selection based on local observations is equivalent to making the optimal joint action selection at the global level, thereby achieving the separation of centralized training and distributed execution. That is, during the training phase, global information is used to optimize the hybrid network and each local value function network, while during the execution phase, each sales agent can independently make the optimal decision based only on its own local observations. The low-level policy generation network is built upon a pre-trained large language model. The base model of the large language model adopts a high-performance dialogue model or a dialogue model fine-tuned for a specific industry vertical domain. On this basis, it is fine-tuned under supervision using historical successful and unsuccessful telemarketing cases, and then further aligned through reinforcement learning based on human feedback, so that the generated policy suggestions conform to the enterprise's sales norms and high conversion characteristics.

[0050] The network uses textual descriptions of the customer's four-dimensional profile, the salesperson's five-dimensional profile, and strategy sensitivity labels output by the counterfactual reasoning unit as input prompts. The prompts are designed as follows: a system instruction text is input into the large language model, stating that the model should act as a senior sales strategy consultant; subsequently, textual summaries of the customer profile, the salesperson profile, and the strategy sensitivity analysis results are provided; finally, the model is required to output structured strategy suggestions. Based on this, the large language model generates structured communication strategy text including suggestions on communication style, key selling point ranking, and risk avoidance tips, outputting it in a structured data format and pushing it to the salesperson's terminal pop-up interface along with the outbound call task.

[0051] The high-level routing strategy network and the low-level strategy generation network perform end-to-end collaborative optimization through a joint reward function. The joint reward function is designed as follows: when a call is successfully closed, a reward signal is simultaneously allocated to both the high-level network (i.e., the routing is correct) and the low-level network (i.e., the strategy is effective); when a call is not closed but the customer's mood has not worsened, only a positive reward is given to the high-level network while a slight penalty is imposed on the low-level network (i.e., the routing may be correct, but the strategy needs optimization); when a call leads to a customer complaint or explicit rejection, a negative penalty is imposed on both the high-level and low-level networks.

[0052] The optimization objective of the game decision-making unit architecture is to minimize the overall joint loss function. Defined by the following formula: ; In the formula, The loss function of the high-level routing policy network represents the time difference error between the joint action value function calculated based on the QMIX hybrid network and the target value. The target value is calculated using multi-step rewards and discounted cumulative rewards. This represents the loss function of the low-level policy generation network, and the optimization objective for generating communication strategies based on a large language model. This represents a dynamic tradeoff coefficient. A larger value is taken in the early stages of training to enable the model to quickly master the basic policy generation capability. As training progresses, the value is gradually reduced to shift the optimization focus towards routing efficiency, thereby balancing routing efficiency and policy quality.

[0053] In addition, the abnormal scenario routing conflict resolution mechanism is as follows: When multiple top salespeople are simultaneously idle and vying for the same S-level lead, the system automatically arbitrates based on a multi-factor weighted score of historical lead conversion rate of similar types, regional suitability, and idle time, allocating the lead to the salesperson with the highest weighted overall priority. When a top salesperson who has been assigned an S-level lead suddenly goes offline or loses connection, the system detects the anomaly within five seconds through a heartbeat detection mechanism, immediately locks the lead without reassignment, and marks the lead's status as locked until recovery in the cache, with a lock timeout of fifteen minutes; once the salesperson returns online, the lead is reassigned first, and the lock mark is automatically cleared; if the lead is not recovered after fifteen minutes, rerouting is triggered, releasing the lead back to the S-level cache queue, where the game decision unit re-matches it with an online top salesperson.

[0054] 4. Real-time assistance module This mechanism is used to analyze changes in customer emotions and intentions in real time during a call, dynamically modify pre-generated communication strategies, and push them to the salesperson's terminal, forming a real-time command mechanism for human-machine collaboration.

[0055] The process of the real-time assistance module is as follows:

[0056] During the call, the audio stream and real-time speech-to-text stream are acquired in real time via a long connection. The audio stream is input to a speech emotion recognition model, which is fine-tuned based on a pre-trained audio model and trained on a public emotion dataset overlaid with labeled data from a telemarketing scenario. The output is a frame-level probability distribution of emotion categories, including calm, interest, hesitation, impatience, and anger. The text stream is input to a text emotion classification model fine-tuned based on a pre-trained language model, which outputs the emotional tendency at the sentence level. The two emotion signals are merged through an attention fusion layer to generate the fused emotional state at the current moment.

[0057] Based on a sequence of consecutive timestamps of emotional states, a time curve of customer emotions is constructed. Simultaneously, an emotion trend prediction sub-model is introduced, which models the emotional sequence of the past thirty seconds using a temporal neural network to predict the emotional change trend for the next ten to thirty seconds, and outputs the trend prediction confidence score.

[0058] When a customer's emotional state is detected to be deteriorating, such as exhibiting impatience or anger for five consecutive seconds, the dynamic strategy adjustment unit is immediately triggered. Based on a pre-generated library of strategies for handling deteriorating emotions and the current call context, a silent text prompt is pushed to the salesperson's terminal. The prompt includes: adjustment suggestions, such as suggesting a slower speaking speed and switching to empathetic language; strategy switching, such as immediately switching to a risk mitigation strategy and stopping the push for a sale; and risk warnings, such as if the customer is about to express refusal, please confirm their needs before quoting a price.

[0059] When a shift in customer intent is detected, such as a sudden change from focusing on price to questioning product reliability, the communication strategy is updated in real time. The lower-level strategy generation network is invoked again to generate new strategy recommendations and push them out.

[0060] When a customer sends a clear rejection signal, such as saying they are not interested, asking not to call again, or filing a complaint, a negative feedback signal is sent to the dynamic scoring module, triggering the scoring penalty mechanism. The scoring penalty mechanism uses a multiplicative penalty method: the current score is multiplied by a penalty coefficient, with a penalty of 0.5 for the first rejection, 0.2 for the second rejection, and 0 for a complaint. Simultaneously, a cooling-off period mechanism is triggered for the lead, automatically removing it from the active allocation pool during the preset cooling-off period.

[0061] When a customer's mood is detected to be stabilizing or turning positive, a positive prompt is sent to the salesperson's terminal, forming a complete closed loop of emotion perception and strategy adjustment.

[0062] 5. Closed-loop self-optimization module

[0063] It is used to collect telemarketing conversion results and call process data, and feeds back to update the model parameters and strategy network of each module, forming a closed loop of continuous self-optimization.

[0064] Specifically, after each call, for those that resulted in a sale, the key drivers of customer conversion are analyzed: was it accurate routing, appropriate strategy selection, or a strong enough customer intention to buy? For calls that did not result in a sale, the main reasons for failure are analyzed: was it routing error, strategy failure, or insufficient customer willingness? The specific method of attribution analysis is as follows: within the framework of a structural causal model, the actual outcome is compared with the counterfactual prediction. For example, counterfactual reasoning is performed on a closing call, assuming the lead was assigned to a different salesperson with a different communication strategy, predicting the conversion probability; if the counterfactual prediction of the conversion probability is significantly lower than the actual outcome, it indicates that the routing and strategy selection were effective, and vice versa.

[0065] The attribution results are stored in a structured manner and fed back to the dynamic scoring module to update the edge weights and conditional probability distributions in the structural causal model. Specifically, when a large number of attribution results point to a causal edge with a significant estimation bias, such as the causal effect of a limited-time discount strategy on conversion results, the parameters of the structural equation are refitted to ensure that the causal graph is consistent with the latest actual data distribution.

[0066] When new scoring rules, routing strategies, or quota settings are officially launched, large-scale Monte Carlo simulations are conducted in a digital twin environment. The digital twin environment is constructed by extracting statistical features such as customer arrival time distribution, customer characteristic experience distribution, salesperson service time distribution, and conversion probability conditional distribution from historical telemarketing data over the past ninety days. This creates a virtual telemarketing center with statistical characteristics consistent with the real environment. This virtual environment supports accelerated simulations: simulating the operations of the next week at 100x speed, completing the simulation within minutes. During the simulation, both the current and new strategies are run, each undergoing one thousand Monte Carlo simulations. Indicators such as total conversions, resource utilization, number of customer complaints, and agent load distribution are recorded for each simulation. The final output includes a quantitative evaluation report containing the expected change in conversion rate, the change in resource utilization, and fairness indicators.

[0067] After deployment, the system monitors the deviation between the deployment results and simulation predictions in real time. Monitoring metrics include the difference between the actual and predicted conversion rates, the difference between the actual and predicted resource utilization rates, and changes in the actual customer complaint rate. When the deviation between the actual and predicted values ​​exceeds a preset threshold, a policy rollback is automatically triggered, reverting to the previous version of the policy, and an alert is sent to operations personnel for manual investigation.

[0068] The following is an overall description of the workflow of this system.

[0069] Step 1: Data Collection and Profile Building. The data collection and bidirectional profile building module continuously collects customer behavior data and salesperson capability data from multiple sources, including online tracking, CRM systems, CTI systems, and federated learning nodes, to build a four-dimensional dynamic behavioral profile of customers and a five-dimensional dynamic capability profile of salespeople. Simultaneously, it constructs a bidirectional heterogeneous knowledge graph of customer, salesperson, and product nodes and their multi-dimensional interactions. The weights of edges in the graph decay over time to reflect the freshness of behavior, and the tag weights are adaptively updated based on each conversion result. Through multimodal data fusion and knowledge graph construction, a comprehensive, dynamic, and relational representation of customers and salespeople is achieved, providing a richer and more real-time data foundation for subsequent scoring and routing than traditional static tags.

[0070] Step Two: Dynamic Scoring with Causal Enhancement. After acquiring a dynamic customer behavior profile, the dynamic scoring module uses a causal graph construction unit to build a structural causal model based on historical data, identifying and eliminating confounding factors. The counterfactual reasoning unit calculates the conditional average treatment effect of each alternative strategy and outputs strategy sensitivity labels. The dual-tower scoring unit predicts organic purchase intention and churn risk separately, integrates the strategy sensitivity labels, and calculates the causal enhancement potential score. When a customer triggers key behavioral events such as visiting the pricing page or submitting a demo request, the score is immediately recalculated, achieving real-time updates within seconds. This fundamentally breaks through the limitations of traditional scoring models that only capture correlation. By introducing a causal inference framework, the score not only reflects how much a customer wants to buy but also quantifies the degree to which the optimal strategy can persuade them to buy, significantly improving the score's forward-looking predictive ability for conversion results.

[0071] Step 3: Lead Hierarchy and Resource Isolation. The resource isolation unit of the routing and collaborative scheduling module adaptively divides leads into three levels (S, A, and B) based on causal enhancement potential scores, establishing a three-level resource isolation pool. S-level leads are rigidly locked, flowing only into the exclusive service pool of top salespeople, and are subject to both quota and load limits. Through this physical isolation mechanism, the possibility of high-value leads being inefficiently consumed is eliminated from the system design, providing stronger protection and higher certainty compared to existing priority allocation schemes.

[0072] Step 4: Hierarchical Game Theory Decision Making and Strategy Pre-generation. The high-level routing strategy network of the game decision-making unit adopts the QMIX multi-agent reinforcement learning algorithm, aiming at a global optimum with multiple objectives including conversion rate, resource utilization, customer satisfaction, and load balancing, to find the optimal matching scheme between leads and top salespeople. After matching, the low-level strategy generation network, based on a large language model, takes customer profiles, salesperson profiles, and strategy sensitivity labels as input to pre-generate structured communication strategy text containing suggestions on conversation style, ranking of key selling points, and risk avoidance tips before outbound calls. This text is then pushed to the salesperson's terminal along with the outbound call task. Unlike existing technologies that separate routing and strategy, this system uses a hierarchical multi-agent architecture for end-to-end collaborative optimization, ensuring that routing decisions and communication strategies are coordinated and consistent under a unified global objective, maximizing the conversion potential of high-value leads.

[0073] Step 5: Real-time Assistance During the Call. After the call is established, the real-time assistance module fuses speech acoustic features and text semantic features in real time using a multimodal emotion perception model to construct a customer emotion time curve and predict emotion trends. When a deterioration in emotion or intention drift is detected, a correction strategy is generated immediately and pushed to the salesperson's terminal via a silent prompt; when an explicit rejection or complaint is detected, negative feedback is sent to the dynamic scoring module to trigger scoring penalties and lead cooling. This forms a real-time closed loop of human-machine collaboration, effectively reducing customer churn and complaint risks caused by excessive harassment through the systematic utilization of negative feedback signals.

[0074] Step 6: End-to-End Closed-Loop Self-Optimization. After each call, the closed-loop self-optimization module automatically analyzes the key driving factors for successful or failed conversions based on a counterfactual attribution framework, and feeds the attribution results back to update the causal model parameters. Before a new strategy goes live, a large-scale Monte Carlo simulation is conducted in a digital twin environment, outputting a quantitative evaluation report. After successful verification, the strategy is automatically deployed and performance deviations are continuously monitored. The benefits of this step are: establishing a complete self-evolving closed loop, enabling the system to continuously iterate itself, migrating strategy trial and error from real business to the virtual environment, and reducing innovation risks and optimization costs.

[0075] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0076] The various embodiments of this application have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A telemarketing lead intelligent scoring and differentiated routing scheduling system that integrates customer behavior profiles, characterized in that, include: The data collection and bidirectional profiling module is used to collect multimodal customer behavior data and salesperson multidimensional capability data, respectively construct dynamic customer behavior profiles and dynamic salesperson capability profiles, and build a bidirectional heterogeneous knowledge graph of customer and sales based on the two. The dynamic scoring module is used to identify customer conversion tendencies and strategy sensitivity based on the dynamic customer behavior profile, and to generate interpretable causal enhancement potential scores. The routing and scheduling module is used to perform lead layering and resource isolation based on the causal enhancement potential score, and adopts a hierarchical multi-agent game decision architecture. The high-level agent performs the global optimal matching of leads and salespeople, and the low-level agent pre-generates personalized communication strategies and outbound call timing decisions before making outbound calls, so as to realize differentiated routing and strategy collaborative scheduling. The real-time support module is used to analyze changes in customer emotions and intentions in real time during the call, dynamically modify the pre-generated communication strategy and push it to the salesperson's terminal, forming a real-time command mechanism for human-machine collaboration. The closed-loop self-optimization module is used to collect telemarketing conversion results and call process data, and feed back to update the model parameters and strategy network of each module to form a continuously self-optimizing closed loop.

2. The system according to claim 1, characterized in that, In the data acquisition and two-way profiling module: The customer dynamic behavior profile includes four sub-profile dimensions: conversion intention time sequence characteristics, decision resistance characteristics, communication preference characteristics, and price sensitivity characteristics; the customer behavior data is collected from online behavior data, offline interaction data, historical telemarketing data, and third-party compliant data. The salesperson's dynamic capability profile includes five sub-profile dimensions: industry category proficiency profile, communication style profile, customer compatibility profile, emotional control ability profile, and idle time slot load profile. The customer-sales bidirectional heterogeneous knowledge graph includes three types of nodes: customer nodes, salesperson nodes, and product nodes. The weights of the edges in the graph decay over time to reflect the freshness of the behavior, and the tag weights are adaptively updated hourly based on each telemarketing conversion result.

3. The system according to claim 1, characterized in that, The dynamic scoring module further includes: The causal graph construction unit builds a structural causal model based on historical telemarketing data, explicitly models the causal relationship graph between customer characteristics, salesperson characteristics, communication strategies and conversion results, identifies and eliminates confounding factors caused by exposure bias and selection bias, and obtains the bias-free causal effect through backdoor adjustment. The counterfactual reasoning unit calculates the average conditional treatment effect for each clue, quantifies the causal transformation increment of different intervention strategies for that clue, and outputs strategy sensitivity labels. The dual-tower scoring unit adopts a dual-tower deep learning model architecture. One tower predicts the probability of a customer's natural purchase intention without any specific communication strategy intervention, while the other tower predicts the risk of customer churn or complaints. The strategy sensitivity labels are integrated to output a comprehensive causal enhancement potential score, and attribution explanations for the score are also output, indicating the contribution of each feature dimension to the score. The causal enhancement potential score Defined by the following formula: ； In the formula, This represents the probability of a customer's natural purchase intention without intervention. The hyperparameter representing the adjustment of policy gain. This represents the maximum average conditional treatment effect for this customer among all alternative communication strategies.

4. The system according to claim 1, characterized in that, The routing coordination scheduling module further includes: The resource isolation unit categorizes leads into three levels—S, A, and B—based on the causal enhancement potential score, and sets up three-level lead resource isolation pools. S-level leads are rigidly locked and only allowed to flow into the exclusive service pool of top salespeople, prohibiting their allocation to ordinary and junior salespeople. At the same time, a maximum service quota of S-level leads is set for each top salesperson, as well as a load limit threshold set based on the number of S-level leads currently being served. When the quota is exhausted or the load threshold is reached, S-level leads are automatically and temporarily cached, and prioritized for allocation when the salesperson is idle. The game-theoretic decision-making unit adopts a multi-agent reinforcement learning architecture with centralized training and distributed execution. The high-level routing strategy network takes the optimization of multiple objectives, including global conversion rate, utilization of top-tier resources, customer satisfaction, and agent load balancing, as its task to find the optimal matching scheme between leads and salespeople. The low-level strategy generation network generates personalized communication strategies and outbound call timing decisions for specific customers and salespeople based on the aforementioned heterogeneous customer-sales bidirectional knowledge graph and customer causal profile. The high-level and low-level networks perform end-to-end collaborative optimization through a joint reward function to ensure that routing decisions and strategy generation are coordinated and consistent under the global objectives.

5. The system according to claim 4, characterized in that, In the game decision-making unit: The high-level routing strategy network adopts the QMIX algorithm, which models each salesperson as an independent agent. The global value function is decomposed into the local value function of each agent through a hybrid network, achieving global optimal cooperation under the condition that only local observation is required. At the same time, an attention mechanism is introduced to dynamically adjust the weight coefficients of each constraint. The low-level strategy generation network is built on a pre-trained large language model. It uses the four-dimensional profile of the customer, the five-dimensional profile of the salesperson, and the strategy sensitivity labels output by the counterfactual reasoning unit as prompts and inputs to generate structured communication strategy texts that include suggestions on speaking style, ranking of key selling points, and risk avoidance prompts. The texts are then pushed to the salesperson's terminal along with the task. The optimization objective of the game decision-making unit architecture is to minimize the overall joint loss function. Defined by the following formula: ； In the formula, The loss function of the high-level routing policy network represents the time difference error between the joint action value function calculated based on the QMIX hybrid network and the target value. The target value is calculated using multi-step rewards and discounted cumulative rewards. This represents the loss function of the low-level policy generation network, and the optimization objective for generating communication strategies based on a large language model. This represents a dynamic tradeoff coefficient used to balance routing efficiency and policy quality.

6. The system according to claim 1, characterized in that, The real-time assistance module further includes: fusing voice acoustic features and text semantic features in real time during the call to identify the customer's current emotional state and predict the trend of emotional change, and constructing a time curve of the customer's emotions; when the deterioration of the customer's emotions or the drift of intent is detected, the communication strategy is updated immediately and a silent prompt is given through the salesperson's terminal; when the customer gives a clear rejection or complaint signal, a negative feedback signal is sent to the dynamic scoring module to trigger the scoring penalty mechanism and reduce the potential score of the lead.

7. The system according to claim 1, characterized in that, The optimization process of the closed-loop self-optimization module includes: after each call ends, analyzing the underlying reasons for the success or failure of the conversion based on the counterfactual inference framework, storing the attribution results in a structured manner and feeding them back to the dynamic scoring module, and updating the edge weights and conditional probability distributions in the structural causal model. When new scoring rules, routing strategies, or quota settings are officially launched, a large-scale Monte Carlo simulation is conducted in a virtual simulation environment built based on historical telemarketing data. The simulation outputs a quantitative evaluation report that includes expected conversion rate, resource utilization rate, and fairness indicators. After verification and manual confirmation, the simulation is automatically deployed to the production environment. At the same time, the deviation between the launch effect and the simulation prediction is monitored in real time. When the deviation exceeds the preset threshold, the strategy rollback is automatically triggered.

8. The system according to claim 1, characterized in that, The dynamic scoring module also introduces a key behavioral event triggering mechanism. When a customer triggers a preset key behavioral event, the causal enhancement potential score is immediately recalculated, achieving real-time score updates within seconds. The preset key behavioral events include accessing the pricing page, submitting a demo request, and downloading the product white paper.

9. The system according to claim 4, characterized in that, The routing coordination scheduling module also includes a routing conflict resolution mechanism for abnormal scenarios: When multiple top salespeople are simultaneously available and vying for the same S-level lead, the system automatically arbitrates the case based on a weighted score that considers factors such as historical conversion rate of similar leads, geographic suitability, and available time. The lead is then assigned to the salesperson with the highest weighted overall priority. If a top salesperson who has been assigned an S-level lead suddenly goes offline or loses connection, the system temporarily locks the lead and does not reassign it. The lead will be reassigned as soon as the salesperson becomes online again. If the lead is not reassigned after a preset waiting period, a rerouting is triggered.

Images