Intelligent sales clue scheduling system and method based on intention intensity and historical feedback

Abstract

The invention discloses a sales clue intelligent scheduling system and method based on intention intensity and historical feedback, and relates to the technical field of artificial intelligence and customer relationship management. The system comprises a data acquisition and integration module, a multi-modal feature analysis module, a core intelligent module and a feedback iteration module. After an AI call is ended, a data acquisition and integration module generates a structured clue file, a multi-modal feature analysis module extracts high-dimensional context situation data, a core intelligent module inference intention scores according to the data, plans and executes an optimal action sequence, feeds back a service result fine adjustment model by a feedback iteration module, and forms closed-loop evolution. Therefore, the intelligent scheduling reliability of the sales clues based on the intention intensity and the historical feedback is improved, and the problem that the intelligent scheduling reliability of the sales clues based on the intention intensity and the historical feedback is low due to the fact that full-link closed-loop optimization from data input to intelligent scheduling cannot be formed in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and customer relationship management technology, and in particular to an intelligent sales lead scheduling system and method based on intent strength and historical feedback. Background Technology

[0002] In automated communication scenarios such as AI-powered telemarketing and intelligent customer service, the process begins with real-time speech-to-text technology to synchronously convert the audio of the AI-driven conversation with the user into structured text data. Simultaneously, NLP (Natural Language Processing) entity recognition and keyword extraction technologies are used to extract key information such as user identity and product consultation points from the text. Combined with sentiment analysis, this forms the basic interaction dataset for each lead. Next, the contextual understanding capabilities of a large language model are used for deep semantic analysis of the interactive text. A multi-dimensional weighted scoring algorithm quantifies the intensity of intent, achieving precise stratification of lead intent. Subsequently, user profiles and historical feedback databases are accessed. Data association technology matches the current lead with the user's past call records, consultation records, click behavior, and historical follow-up results. Machine learning regression models are used to analyze the conversion effect data of leads with similar intent scores and similar needs under different follow-up strategies, generating a strategy effectiveness evaluation matrix.

[0003] Finally, based on reinforcement learning algorithms combined with real-time conversion data, the strategy matching logic is dynamically optimized. Leads with high intent and similar historical data that have achieved excellent conversion results under the instant manual outbound calling strategy are automatically assigned to the instant follow-up queue of top sales staff.

[0004] For example, the blockchain-based intelligent scheduling method, system, and medium for the sales product supply chain disclosed in patent application CN117910738A includes: acquiring total supply characteristic data information, intelligently scheduling the total supply characteristic data information based on the estimated supply chain difference characteristics, generating a first intelligent scheduling result, finally acquiring the starting point and ending point of each supply chain in the first intelligent scheduling result, and optimizing the first intelligent scheduling result based on the starting point and ending point of each supply chain to generate an optimized first intelligent scheduling result.

[0005] For example, the patent application announcement CN118350612B discloses a sales resource intelligent scheduling system based on user big data, which includes: a sales resource allocation module, a sales resource information collection module, a sales resource data preprocessing module, a sales data anomaly detection module, a user purchase behavior analysis module, a sales supply and demand balance analysis module, a sales resource comprehensive analysis module, a sales resource management judgment module, a sales resource intelligent scheduling module, an intelligent scheduling execution module, and an AR display terminal module.

[0006] However, in the process of implementing the inventive technical solution in the embodiments of this application, it was found that the above-mentioned technology has at least the following technical problems: In the field of sales lead evaluation and scheduling in automated communication scenarios such as AI telemarketing and intelligent customer service, existing solutions either classify leads based on simple rules such as preset keywords and call duration, or use multiple independent machine learning models to process text, voice and other data separately and then fuse the results through fixed rules. Neither of these solutions achieves unified integration and comprehensive analysis of multi-source heterogeneous data (call text, audio, history, business metadata, etc.), resulting in a one-sided judgment of user intent and an inability to capture the complex relationships between multi-dimensional information. Meanwhile, the system only stays at the pattern matching level, making it difficult to deeply infer real needs and potential intentions. Subsequent follow-up strategies rely on manually preset "IF-THEN" fixed rules, which not only has high maintenance costs and cannot adapt to changing business needs, but also causes the model's classification tasks to become disconnected from the final scheduling business goals. It cannot achieve dynamic optimization and intelligent scheduling, which seriously affects the efficiency of sales lead conversion and human resource utilization. This results in information transmission loss and logical separation between data processing, intention inference and strategy execution, and cannot form a closed-loop optimization from data input to intelligent scheduling. Consequently, there is a problem of low reliability of intelligent scheduling of sales leads based on intention strength and historical feedback. Summary of the Invention

[0007] This application provides a sales lead intelligent scheduling system and method based on intent strength and historical feedback, which solves the problem in the prior art that it is impossible to form a closed-loop optimization of the entire link from data input to intelligent scheduling, resulting in low reliability of sales lead intelligent scheduling based on intent strength and historical feedback, and improves the reliability of sales lead intelligent scheduling based on intent strength and historical feedback.

[0008] On one hand, a sales lead intelligent scheduling system based on intent strength and historical feedback is provided, including: a data collection and integration module, a multimodal feature analysis module, a core intelligence module, and a feedback iteration module. The data collection and integration module automatically collects and integrates sales lead data related to the current sales lead from multiple call task data sources after each AI call task, forming a structured lead profile. The multimodal feature analysis module performs deep feature extraction on the collected sales lead data to obtain contextual data, providing processed, high-dimensional information input. The multimodal feature analysis module includes textual semantics. The system comprises a feature analysis submodule and a speech acoustic feature analysis submodule. The core intelligence module performs end-to-end overall contextual understanding and reasoning based on the structured cue archives and contextual data provided by the multimodal feature analysis module. This generates quantified cue intent strength scores and decision reasons, and autonomously plans the optimal sequence of subsequent actions based on these scores and historical feedback data. Finally, it automatically executes the scheduled tasks by calling a pre-set set of atomic tools. The feedback iteration module collects sales business results feedback after the core intelligence module's scheduling and execution, and feeds this feedback back to the core intelligence module as training data. This data is used to fine-tune and optimize the large-scale language model, forming a closed-loop evolutionary mechanism.

[0009] On the other hand, a sales lead intelligent scheduling method based on intent strength and historical feedback is provided, including: after each AI call task ends, automatically collecting and integrating sales lead data related to the current sales lead from multiple call task data sources to form a structured lead profile; performing deep feature extraction on the collected sales lead data to obtain contextual data to provide processed, high-dimensional information input, the multimodal feature analysis module including a text semantic feature analysis submodule and a speech acoustic feature analysis submodule; performing end-to-end overall contextual understanding and reasoning based on the structured lead profile and the contextual data provided by the multimodal feature analysis module to generate a quantified lead intent strength score and decision reasoning, and autonomously planning the optimal subsequent action sequence based on this and historical feedback data, and finally automatically executing the scheduling task by calling a preset atomic toolset; collecting the sales business results feedback after the core intelligent module's scheduling execution, using it as training data to flow back to the core intelligent module for fine-tuning and optimizing the large language model, forming a closed-loop evolution mechanism.

[0010] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. By achieving precision and interpretability in sales lead intent assessment, the system significantly reduces the risk of misjudgment and missed judgment. The data collection and integration module fully integrates call data, historical contact records, and business background data. Combined with the multimodal feature analysis module for deep extraction of text semantics and speech acoustic features, the core intelligent module leverages a pre-trained large-scale language model to achieve cross-dimensional contextual understanding. Based on a multi-dimensional intent assessment system and dynamic weighting algorithm, it generates quantitative scores and outputs structured decision-making reasons containing supporting evidence and potential concerns. This not only solves the problem of biased assessment caused by traditional technologies relying on single data or simple rules, but also improves the accuracy of intent judgment. Furthermore, the interpretable design allows sales personnel to clearly grasp the core characteristics of leads, providing clear guidance for subsequent follow-up and effectively reducing the loss of high-intent leads and the resource consumption of low-intent leads.

[0011] 2. By improving the automation and intelligence of sales lead scheduling, optimizing human resource allocation and conversion efficiency, the core intelligent module filters similar lead sets by associating historical feedback data, plans the optimal action sequence based on a strategy evaluation model trained by reinforcement learning algorithms, and realizes the automatic execution of scheduling tasks in combination with atomic toolsets. At the same time, it matches personalized follow-up strategies and execution parameters for leads with different intention intensities and demand characteristics. This not only breaks through the rigidity of traditional fixed rule scheduling, but also achieves precise matching of intention intensity, historical effect, and scheduling strategy, thereby increasing the immediate follow-up rate of high-intent leads, increasing the number of effective leads followed up per salesperson, improving the response rate of secondary outreach, significantly reducing human resource waste, shortening the lead conversion cycle, and greatly improving marketing conversion efficiency and return on investment.

[0012] 3. By constructing a dynamic closed-loop evolution mechanism, the system ensures long-term adaptability to changes in business scenarios and iterative demands. The feedback iteration module generates a high-quality closed-loop training dataset through comprehensive collection, structured processing, and correlation binding of multi-dimensional feedback data. Combined with a hierarchical fine-tuning strategy, the large language model of the core intelligent module is optimized in a targeted manner. This avoids catastrophic forgetting during model training and accurately optimizes model parameters and decision logic based on feedback from different group data. At the same time, through regular full-scale fine-tuning and effect verification mechanisms, the system can continuously adapt to changes in business scenarios, product updates, and user demand iterations. Core indicators such as the accuracy of intent assessment and the conversion rate of scheduling strategies are continuously optimized with the accumulation of data, achieving long-term evolution of system performance and providing enterprises with stable and continuously improving automated marketing support capabilities.

[0013] 4. By achieving end-to-end integrated analysis, the system innovatively uses all relevant data (call text, audio, historical records, etc.) as a unified input. A core large model performs comprehensive end-to-end analysis and evaluation, breaking down information silos and achieving a global and in-depth understanding of clues. This enables precise inference and quantification of intent. Utilizing the emergent capabilities of the large model, the system is no longer a simple pattern match but can perform deep reasoning about user intent, emotions, and potential needs, outputting a precise and interpretable quantitative intent score. Through intelligent autonomous scheduling, the scheduling process is revolutionaryly upgraded from "rule matching" to "large model calling tools." As an intelligent decision-making brain, the large model can autonomously plan and call the most appropriate engineering tools (such as "transfer to human operator" or "send SMS") to execute the next action based on its evaluation of clues, achieving true automation and intelligence. Changes in business logic no longer require modifying complex engineering code; instead, they can be achieved by providing new instructions or tools to the large model through natural language, greatly improving the system's business agility and scalability. Attached Figure Description

[0014] Figure 1 A schematic diagram of the structure of the intelligent sales lead scheduling system based on intent strength and historical feedback provided in the embodiments of this application; Figure 2 A flowchart of a sales lead intelligent scheduling method based on intent strength and historical feedback provided in an embodiment of this application. Detailed Implementation

[0015] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0016] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, their intended meanings are consistent. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, their intended meanings are consistent.

[0017] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0018] like Figure 1 The diagram shown illustrates the structure of the intelligent sales lead scheduling system based on intent strength and historical feedback provided in this embodiment of the application. This system includes: a data acquisition and integration module, a multimodal feature analysis module, a core intelligence module, and a feedback iteration module. The data acquisition and integration module automatically collects and integrates sales lead data related to the current sales lead from multiple call task data sources after each AI call task ends, forming a structured lead profile. The multimodal feature analysis module performs deep feature extraction on the collected sales lead data to obtain contextual data, providing processed, high-dimensional... The information input includes a text semantic feature analysis submodule and a speech acoustic feature analysis submodule. The core intelligence module performs end-to-end overall contextual understanding and reasoning based on the structured cue archive and the contextual data provided by the multimodal feature analysis module. This generates quantified cue intent intensity scores and decision reasons, and autonomously plans the optimal sequence of subsequent actions based on historical feedback data. Finally, it automatically executes the scheduled task by calling a preset atomic toolset. The feedback iteration module collects the sales business results feedback after the core intelligence module's scheduling and execution, and feeds it back to the core intelligence module as training data for fine-tuning and optimizing the large language model, forming a closed-loop evolution mechanism.

[0019] It is necessary to understand that Figure 1The system clearly presents the entire chain logic from data input, processing and analysis, intelligent decision-making to execution feedback: The data source on the left includes raw data such as call audio, transcribed text, and historical contact records generated by the AI ​​outbound calling system and AI dispatch system. This data is integrated into a structured clue archive by the data acquisition and integration module within the system boundary. Then, the multimodal feature analysis module generates complete and rich contextual data through LLM text semantic analysis (extracting user emotions and domain-specific intentions) and multimodal LLM audio acoustic analysis (identifying complex emotions such as sarcasm and hesitation). Subsequently, this data is input into the S103 core intelligent module (large model intelligent agent). This module sequentially completes the understanding, reasoning, planning, decision-making, and action process, outputs tool call instructions, and calls API interfaces such as assign_to_sales (sales allocation), send_message (message sending), and update_crm_tag (updating CRM tags) in the atomic toolset to send execution instructions to downstream systems such as the CRM system and SMS / WeChat gateway on the right. At the same time, business result feedback data such as transactions and failures generated by downstream systems are fed back to the system for iterative optimization of the core large model. By integrating multi-source data to break down information silos, leveraging multimodal feature analysis to achieve deep understanding of leads, and relying on large-scale intelligent models to achieve integrated autonomous scheduling of "assessment-decision-execution," and then using business feedback to drive model iteration, this approach not only solves the problems of rigid scheduling and fragmented information in existing technologies, but also significantly improves the accuracy of lead intent judgment and scheduling adaptability, thus building a self-evolving intelligent scheduling closed loop that effectively improves sales conversion efficiency and system business agility.

[0020] In this embodiment, a full-link technology system of "data integration - feature extraction - intelligent scheduling - closed-loop optimization" is constructed through the collaborative linkage of the data acquisition and integration module, the multimodal feature analysis module, the core intelligence module, and the feedback iteration module. The data acquisition and integration module automatically aggregates call data, historical contact records, and business background data after the AI ​​call ends, forming a structured clue archive. This completely breaks down the problem of scattered and disorganized information silos in traditional clue data, providing a complete and standardized data foundation for subsequent accurate analysis. The multimodal feature analysis module, through deep feature extraction from both text semantic and speech acoustic sub-modules, transforms unstructured data into high-dimensional contextual data, compensating for the deficiency of single text analysis in capturing implicit information such as voice emotion and tone, and significantly improving the ability to perceive users' real needs and potential intentions. The core intelligence module, relying on end-to-end overall contextual understanding and reasoning capabilities, generates quantifiable and interpretable clue intention intensity scores, and autonomously plans the optimal subsequent action sequence based on historical feedback data. The automated scheduling through atomic toolsets not only solves the pain points of rigid rule-driven scheduling and high manual intervention costs in traditional methods, but also achieves full-process automation and intelligence from "intent assessment to strategy matching to scheduling execution." This allows high-intent leads to be followed up accurately and efficiently, while medium- and low-intent leads receive personalized nurturing, significantly optimizing human resource allocation and lead conversion efficiency. The feedback iteration module collects multi-dimensional sales business results and feeds them back to the large-scale language model for fine-tuning and optimization, forming a continuously evolving closed-loop mechanism. This enables the system to dynamically adapt to changes in business scenarios, iterative user needs, and product updates, continuously improving the accuracy of intent assessment, the adaptability of scheduling strategies, and conversion effects. Ultimately, this achieves a comprehensive and long-term improvement in marketing conversion efficiency, human resource utilization efficiency, and the return on investment in automated marketing.

[0021] Furthermore, the system collects and integrates sales lead data related to the current sales leads, specifically including: a data set of each call task, a dynamically updated set of historical outreach records, and a set of business background data. The data set of each call task includes the text content of the call after each AI call task, the complete audio recording of the call, and call metadata including call duration and call time. The dynamically updated set of historical outreach records includes the extracted and recorded interaction status and results of sales leads in historical marketing activities. The interaction status and results include at least: the connection status of historical calls, the sending and delivery status of historical SMS messages, and the success or failure of adding social contact information. The set of business background data includes the business scenario in which the call task takes place, including at least: industry classification, the theme of this marketing campaign, and information on the main products being promoted.

[0022] In this embodiment, by precisely defining the scope and specific composition of sales lead data collection, a comprehensive and structured integration of lead-related data is achieved. It clearly defines the collection of metadata such as text content, complete audio recordings, call duration, and call time for each call, ensuring comprehensive retention of current call interaction information and avoiding omissions of key communication details. Simultaneously, dynamically updated historical reach records are integrated, covering core interaction data such as historical call connection status, SMS sending and delivery status, and social contact addition results, fully reconstructing the historical communication trajectory and response characteristics of leads, providing historical evidence for judging lead stickiness and potential intentions. Furthermore, industry classifications and marketing activity data are incorporated. By integrating business background data such as topic and featured product information, lead data is deeply bound to specific business scenarios, avoiding one-sided analysis detached from the context. Ultimately, through the systematic integration of three types of data sets, the limitations of information from a single data dimension are broken, forming a complete structured lead archive that includes current interaction details, historical communication trajectory, and business scenario background. This provides comprehensive and standardized data source support for subsequent multimodal feature extraction and ensures that subsequent intention assessment and scheduling strategies can be carried out based on the entire lead lifecycle information. This improves the accuracy of lead analysis and the adaptability of scheduling decisions from the source, effectively solving the analytical bias problems caused by the fragmentation of traditional lead data collection and weak scenario relevance.

[0023] Furthermore, the specific steps for forming a structured lead profile are as follows: Standardize and clean the collected sales lead data, unstructured call text content is segmented and denoised, and audio files of different formats are converted to a unified encoding format and sampling rate; based on a preset timestamp sequence, the call text content, corresponding audio segments, and events during the call are aligned and correlated; the cleaned and aligned data is then integrated with the preset historical outreach record set and business background data set, encapsulated into a structured data object, and output as the structured lead profile to the multimodal feature analysis module.

[0024] In this embodiment, through structured processing steps of standardized cleaning, timestamp alignment and association, and multi-source data fusion, the core technical effect of transforming fragmented sales lead data into standardized and interconnected data is achieved: First, the collected sales lead data is standardized and formatted. Unstructured call text is organized through sentence segmentation and noise reduction. Audio in different formats is converted into a unified encoding format and sampling rate, effectively eliminating data adaptation obstacles caused by text redundancy interference and audio format differences, ensuring the consistency and usability of basic data. Then, based on a preset timestamp sequence, the call text, corresponding audio segments, and call events are precisely aligned and associated, establishing a spatiotemporal correspondence between text content, audio context, and interactive events. This relationship enables subsequent feature extraction to accurately capture communication details and emotional changes at specific time points, avoiding semantic misunderstandings caused by fragmented information. Finally, the cleaned and aligned data is deeply integrated with historical contact records and business background data and encapsulated into structured data objects, forming a logically clear and closely related clue archive. This not only provides a standardized and high-value input data source for the multimodal feature analysis module, ensuring the comprehensiveness and accuracy of feature extraction, but also solves the pain points of traditional clue data being messy in format, loosely related in information, and unable to support in-depth analysis. From the data processing stage, it lays a solid foundation for the accuracy of subsequent intention assessment and the rationality of scheduling decisions, greatly improving the efficiency and reliability of the entire data processing chain.

[0025] Furthermore, the text semantic feature analysis submodule includes a domain pre-training fine-tuning unit, an intent classification unit, and an emotion quantification unit. The domain pre-training fine-tuning unit performs incremental fine-tuning by injecting domain corpus from the AI ​​telephone sales scenario to enable the model to have domain-adaptive semantic understanding capabilities. The intent classification unit constructs domain-specific intent recognition templates, which include at least product function inquiry intent, price inquiry intent, preferential policy inquiry intent, purchase reservation intent, intent to express objections to needs, intent to have no needs at the moment, and intent to repeat inquiries. The emotion quantification unit performs sentiment polarity analysis on the transcribed text and identifies complex emotions in the text to generate multi-dimensional emotion feature vectors. The text semantic feature analysis submodule fuses the intent confidence vector with the multi-dimensional emotion feature vectors to form a text semantic feature set.

[0026] In this embodiment, through the collaborative operation of the domain pre-training fine-tuning unit, the intent classification unit, and the emotion quantification unit, deep and accurate semantic analysis of call text is achieved. The domain pre-training fine-tuning unit performs incremental fine-tuning by injecting domain-specific corpus from AI telemarketing scenarios, enabling the model to overcome the limitations of general semantic understanding and accurately adapt to industry-specific scripts, product terminology, and communication scenarios, significantly improving the recognition accuracy of complex semantics within the domain. The intent classification unit, using domain-specific intent recognition templates, comprehensively covers core interactive intents such as product function inquiries, price consultations, purchase reservations, and objection expressions, achieving accurate classification and confidence quantification of user communication purposes, avoiding the one-sidedness and ambiguity of traditional intent recognition. The emotion quantification unit can not only perform basic analysis of text sentiment polarity but also accurately identify hesitation, entanglement, and longing. The system identifies complex emotions such as doubt and generates multi-dimensional emotion feature vectors to fully capture users' implicit emotional tendencies. Finally, by deeply fusing the intent confidence vector with the multi-dimensional emotion feature vectors, a comprehensive and high-dimensional text semantic feature set is formed. This provides rich and accurate semantic input for subsequent multimodal feature fusion and contextual reasoning of core intelligent modules. It also solves the pain points of traditional text analysis, which only focuses on surface information and lacks domain adaptability and deep emotion capture capabilities. From a semantic perspective, it provides solid support for the assessment of clue intent intensity and the formulation of subsequent scheduling strategies, significantly improving the accuracy of the overall system's intelligent decision-making foundation.

[0027] Furthermore, the speech acoustic feature analysis submodule includes an acoustic feature extraction unit, a multimodal emotion recognition unit, and a feature fusion unit. The acoustic feature extraction unit preprocesses the call audio data to extract core acoustic features such as Mel frequency cepstral coefficients, fundamental frequency, speech rate, energy entropy, speech pause duration, and pause frequency to form a high-dimensional acoustic feature matrix. The multimodal emotion recognition unit uses a multimodal large model (such as the CLIP derivative model or BLIP-2) to perform cross-modal fusion of the high-dimensional acoustic feature matrix with the semantic features of the ASR transcribed text. It captures the correlation between speech intonation and text content through an attention mechanism and outputs the recognition confidence of each complex emotion and the timestamp information of the emotion trigger segment. The feature fusion unit normalizes the output recognition confidence of each complex emotion with the high-dimensional acoustic feature matrix to generate a speech acoustic feature set. The speech acoustic feature analysis submodule concatenates and aligns the dimensions of the speech acoustic feature set with the text semantic feature set output by the text semantic feature analysis submodule to form complete high-dimensional contextual data.

[0028] In this embodiment, through the progressive processing of the acoustic feature extraction unit, the multimodal emotion recognition unit, and the feature fusion unit, in-depth mining and cross-modal information fusion of call audio data are achieved. After preprocessing, the acoustic feature extraction unit accurately extracts core acoustic features such as Mel-frequency cepstral coefficients, fundamental frequency, and speech rate, forming a high-dimensional acoustic feature matrix. This comprehensively captures the underlying acoustic information such as pitch, rhythm, and energy changes in speech, compensating for the speech dimension details that cannot be perceived by relying solely on text. The multimodal emotion recognition unit uses CLIP derived models, BLIP-2, and other large multimodal models to integrate the high-dimensional acoustic feature matrix with ASR (Automatic Speech Recognition). The semantic features of the transcribed text are fused across modally, and the attention mechanism is used to accurately capture the relationship between speech intonation and text content. This not only identifies the confidence level of complex emotions, but also locates the timestamps of emotion-triggered segments, solving the problem that single-modal analysis is difficult to accurately judge the user's true emotions, making emotion recognition more context-specific. The feature fusion unit generates a speech acoustic feature set after normalization, and then splices and aligns it with the text semantic feature set to form a complete high-dimensional contextual data. This not only achieves the complementary effect of speech and text information, but also provides a more comprehensive and three-dimensional contextual input for the core intelligent module. It completely changes the limitation of traditional analysis that only focuses on a single data type, and greatly improves the accuracy of subsequent contextual understanding and intention assessment, laying a solid multimodal feature foundation for the accurate formulation of scheduling strategies.

[0029] Furthermore, end-to-end overall contextual understanding and reasoning are performed to generate quantified clue intent intensity scores and decision reasons. Specifically, this includes: the core intelligent module includes a contextual understanding Prompt template, which contains multimodal feature association guidance, historical feedback fusion rules, and domain demand reasoning logic. After receiving user historical contact records and business domain metadata from the structured clue archive, as well as text semantic feature sets and speech acoustic feature sets output by the multimodal feature analysis module, the multi-source information is structured and injected into a pre-trained large-scale language model to achieve cross-dimensional association understanding of text semantics, speech emotion, historical behavior, and business scenarios. Based on the results of cross-dimensional association understanding, hierarchical reasoning is performed through a preset intent evaluation dimension system, and a weighted summation algorithm is used to calculate the scores of each dimension. The weight values ​​are obtained from historical feedback data through a logistic regression model and support dynamic updates. Based on the output clue intent intensity score, structured decision reasons are generated simultaneously. These decision reasons include scores for each first-level dimension and key supporting evidence (such as "demand clarity 85"). Score, supporting evidence: "User inquired about product delivery cycle and installation process 3 times", potential concerns (such as "Difficulty in resolving objections 60 points, potential concern: User mentioned the price advantage of competing products"), to ensure that the intention assessment results are explainable.

[0030] In this embodiment, multimodal features, historical reach records, and business domain metadata are structurally injected through a contextual understanding Prompt template. A pre-trained large-scale language model is used to achieve cross-dimensional understanding of text semantics, voice emotion, historical behavior, and business scenarios, completely breaking the limitations of traditional single-dimensional analysis and avoiding misjudgments of intent caused by information fragmentation. Based on a pre-defined multi-dimensional intent assessment system, hierarchical reasoning is performed. Weights trained on a logistic regression model using historical feedback data and supporting dynamic updates are employed. A weighted summation algorithm accurately calculates scores for each dimension, generating quantified clue intent strength scores. This ensures that intent assessment aligns with business realities and can be continuously optimized with data accumulation. Simultaneously, a structured decision-making rationale is output, including scores for each primary dimension, key supporting evidence, and potential concerns. This transforms intent assessment results from "black-box judgment" to "interpretable reasoning," addressing the pain points of low accuracy and lack of evidence in traditional intent assessments. It also provides clear decision support for subsequent scheduling strategy formulation, significantly improving the accuracy, reliability, and practicality of clue intent assessment and laying a solid foundation for precise scheduling.

[0031] Furthermore, the specific process for autonomously planning the optimal sequence of follow-up actions based on historical feedback data is as follows: A set of similar leads is obtained by calling the historical feedback database through a data association interface. The follow-up strategy type, execution sequence, tool call combination, and final conversion result data corresponding to each historical lead in this set are extracted. The historical data of the similar lead set is used for training, and the expected benefit score of each candidate follow-up strategy (including immediate human outbound calls, AI secondary outreach, graphic and text material push, multi-round nurturing, etc.) is output. Based on the ranking results of the expected benefit scores, the optimal action sequence is autonomously planned. This optimal action sequence includes the tool call type, execution order, triggering conditions, and timeout handling mechanism (e.g., "prioritize calling the sales personnel allocation tool and assigning it to the top sales queue; if the outbound call is not completed within 1 hour, automatically trigger the SMS push tool to send product core advantage information"). Personalized execution parameters (such as outbound call script focus and push content keywords) are matched to the action sequence based on the decision-making reasons for the current lead to ensure that the scheduling strategy accurately adapts to the lead characteristics.

[0032] In this embodiment, complete follow-up data for similar leads is obtained by associating with a historical feedback database. After targeted training, the expected benefits of each candidate strategy are quantified. Based on this, an optimal action sequence is automatically generated, including tool call type, execution order, triggering conditions, and timeout handling mechanism. Personalized execution parameters are matched with the current lead characteristics. This not only achieves intelligent optimization and dynamic adaptation of follow-up strategies by leveraging historical data accumulation, avoiding redundant and ineffective actions, and significantly improving the accuracy and efficiency of lead follow-up; but also ensures the continuity, stability, and relevance of strategy execution through clear timing rules, fallback mechanisms, and personalized parameter configuration, effectively reducing the cost of manual decision-making, maximizing the potential for lead conversion, and realizing a closed loop of intelligent, efficient, and precise operation from lead follow-up to conversion.

[0033] Furthermore, the system collects sales performance feedback after the core intelligent module's scheduling and execution. Specifically, this includes: acquiring multi-dimensional sales performance feedback data, including core conversion result data (sold, unsold, decision postponed, churn), process feedback data (follow-up duration, communication rounds, user secondary objections, lead status change trajectory), and strategy suitability feedback data (sales personnel's rating of the rationality of the scheduling strategy, user response behavior data to the reached content); the collected feedback data is structured, with core conversion result data mapped to binary labels (sold = 1, unsold = 0) or multi-category labels (sold / unsold / deferred decision / churn corresponding to 1 / 2 / 3 / 4 respectively), process feedback data having key information extracted and converted into structured fields (such as secondary objection type, response behavior frequency), and strategy suitability feedback data quantified into suitability scores; and establishing a linking mechanism between feedback data and original leads, binding structured feedback data with corresponding structured lead profiles, intent intensity scores, and scheduling action sequences through unique lead identifiers to generate a "lead-scheduling-feedback" sequence. The three-in-one closed-loop training dataset simultaneously performs outlier removal (such as erroneous operation feedback and invalid data) and normalization to ensure the quality of the training data.

[0034] In this embodiment, by comprehensively collecting multi-dimensional sales business feedback data including core conversion results, process feedback, and strategy adaptability, non-standardized data is transformed into binary / multi-variable labels, structured fields, and adaptability scores through structured processing. A strong correlation mechanism of "lead-scheduling-feedback" is established through unique lead identifiers. Combined with outlier removal and normalization processing, a high-quality closed-loop training dataset is constructed. On the one hand, this achieves full-dimensional capture and standardized accumulation of sales business feedback data, breaking down data silos and fully restoring the full-link correlation from lead characteristics, scheduling strategies to business results. On the other hand, it provides accurate, comprehensive, and high-quality training data support for the core intelligent module, enabling the module to continuously optimize the strategy decision model based on real business feedback, continuously improve the adaptation accuracy of scheduling strategies and lead characteristics, and help uncover hidden user needs and objections in process feedback. This provides data basis for subsequent strategy iteration, tool optimization, and personalized parameter configuration, promoting the formation of a virtuous cycle of "data collection-model training-strategy optimization-effect feedback" in the sales lead follow-up scheduling system, and continuously improving the overall sales conversion efficiency and business adaptability.

[0035] Further, the specific steps for fine-tuning and optimizing the large-scale language model are as follows: A hierarchical fine-tuning strategy is adopted, grouping the closed-loop training dataset according to intent intensity range (high / medium / low), lead conversion result (sold / unsold), and strategy fit (high / medium / low). Differentiated fine-tuning targets are set for different groups of data. For feedback data with high-intent but unsold leads, the focus is on optimizing the weight parameters of the intent evaluation dimension and the accuracy of potential concern identification. For feedback data with low strategy fit, the focus is on optimizing the decision logic of action sequence planning. This is achieved by combining a domain-adaptive loss function (fusing classification loss and regression loss, with weights based on historical data). The fine-tuning effect is dynamically adjusted. The parameters of the top-level Transformer layer of the large language model are updated, while the parameters of the bottom-level pre-trained layer are kept fixed to avoid forgetting. A fine-tuning effect evaluation and iteration mechanism is established to verify the effect of the fine-tuned model. If the index amplitude is lower than the preset index amplitude threshold, the amount of training data for the corresponding group is automatically expanded and the fine-tuning process is re-executed. At the same time, the model is fine-tuned with full data at preset intervals. If not, the fine-tuning process is not performed, forming a closed-loop evolution mechanism of "data collection-model fine-tuning-effect verification-iterative optimization" to continuously improve the inference accuracy and scheduling adaptability of the core intelligent module.

[0036] In this embodiment, a hierarchical strategy is used to precisely group closed-loop training data according to key dimensions and set differentiated optimization objectives. Combined with a domain-adaptive dynamic weight loss function, the top-level parameters of the model are updated in a targeted manner. Coupled with an iterative mechanism of effect evaluation and periodic full-scale fine-tuning, this achieves precise focus in model optimization. It improves the accuracy of intention evaluation and the rationality of decision-making logic in key scenarios such as high intentions without conversion and low strategy adaptability, avoiding resource waste and effect generalization caused by indiscriminate fine-tuning. Furthermore, by fixing the underlying pre-training parameters, the model is prevented from forgetting its general capabilities. The dynamic loss function and iterative verification mechanism ensure the stability and effectiveness of fine-tuning effects. At the same time, periodic full-scale fine-tuning consolidates the optimization results. Ultimately, this drives the core intelligent module to form a continuously evolving closed loop, continuously improving the accuracy of inference decisions, scenario adaptability, and sales business fit. This makes the scheduling strategy more suitable for the characteristics and needs of different types of leads, further amplifying lead conversion efficiency and business value.

[0037] like Figure 2 The diagram shows a flowchart of the intelligent sales lead scheduling method based on intent strength and historical feedback provided in this application embodiment. The method includes: automatically collecting and integrating sales lead data related to the current sales lead from multiple call task data sources after each AI call task ends, forming a structured lead profile; performing deep feature extraction on the collected sales lead data to obtain contextual data, providing processed, high-dimensional information input; the multimodal feature analysis module includes a text semantic feature analysis submodule and a speech acoustic feature analysis submodule; performing end-to-end overall contextual understanding and reasoning based on the structured lead profile and the contextual data provided by the multimodal feature analysis module to generate a quantified lead intent strength score and decision rationale, and autonomously planning the optimal subsequent action sequence based on this and historical feedback data; finally, automatically executing the scheduling task by calling a preset atomic toolset; collecting sales business results feedback after the core intelligent module's scheduling execution, using it as training data to feed back to the core intelligent module for fine-tuning and optimizing a large language model, forming a closed-loop evolution mechanism.

[0038] It should be noted that, Figure 2The entire execution logic from the end of the AI ​​call to the completion of the lead processing is clearly presented. Specifically, starting from the end of the AI ​​call, the process first involves S101, a multi-source heterogeneous data collection and integration step, which collects call audio, transcribes the text, extracts historical contact records, and integrates business metadata to form structured lead data. Then, in the S102 multimodal feature analysis step, text semantic analysis is performed through LLM, and voice sentiment analysis is performed through MM LLM to generate a complete lead context. Finally, in the S103 integrated evaluation and autonomous scheduling step (executed by a large model intelligent agent), the overall context understanding and reasoning are completed in sequence, quantitative intention scores and decision reasons are generated, action sequences are planned, and atomic tools are called. Finally, specific actions such as transferring to human agent, sending SMS, updating CRM, and scheduling follow-up visits are executed through API calls until the lead processing is completed. This process breaks down information silos in existing technologies by integrating multi-source data, achieves a deep understanding of leads through multimodal feature analysis, and solves the pain points of rigid scheduling and disconnect between models and business processes in traditional solutions by relying on end-to-end integrated processing of large-scale intelligent agents. It not only significantly improves the accuracy of lead intent judgment and the adaptability of scheduling strategies, but also realizes full automation and intelligence of sales lead processing, reduces manual maintenance costs, enhances the system's business agility and scalability, and effectively improves sales conversion efficiency.

[0039] In this embodiment, structured lead profiles are automatically generated by collecting and integrating multi-source data after an AI call. A multimodal feature analysis module extracts high-dimensional contextual data, and combined with end-to-end overall contextual understanding and reasoning, a quantitative intention strength score and decision rationale are generated. The system then autonomously plans and executes the optimal sequence of subsequent actions. Simultaneously, sales results are fed back for model fine-tuning and optimization. This achieves a fully intelligent closed loop for the entire sales lead process, from data collection, feature mining, and contextual understanding to strategy execution and feedback iteration. Structured profiles and multimodal deep feature extraction break down data fragmentation barriers, providing comprehensive and high-value information support for accurate intention assessment and decision-making, making subsequent action sequence planning more aligned with the lead's real needs and contextual characteristics. Furthermore, the model fine-tuning mechanism driven by business result feedback promotes the continuous evolution of core intelligent modules, constantly improving the accuracy of intention judgment, strategy adaptability, and execution efficiency, effectively reducing manual intervention costs, and achieving precise, efficient, and adaptive optimization of lead follow-up, maximizing the conversion potential and business value of sales leads.

[0040] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0041] This invention is described with reference to flowchart illustrations and / or block diagrams of systems, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0042] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0043] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0044] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0045] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A sales lead intelligent dispatch system based on intent strength and historical feedback, characterized in that, It includes a data acquisition and integration module, a multimodal feature analysis module, a core intelligence module, and a feedback and iteration module: The data collection and integration module is used to automatically collect and integrate sales lead data related to the current sales lead from multiple call task data sources after each AI call task ends, forming a structured lead file. The multimodal feature analysis module is used to perform deep feature extraction on the collected sales lead data to obtain contextual data, so as to provide processed, high-dimensional information input. The multimodal feature analysis module includes a text semantic feature analysis submodule and a speech acoustic feature analysis submodule. The core intelligent module is used to perform end-to-end overall contextual understanding and reasoning based on the contextual data provided by the structured clue archive and the multimodal feature analysis module, so as to generate a quantitative clue intention intensity score and decision reason, and autonomously plan the optimal subsequent action sequence based on this and historical feedback data, and finally automatically execute the scheduling task by calling the preset atomic toolset. The feedback iteration module is used to collect the sales business results feedback after the core intelligent module schedules and executes them, and use it as training data to flow back to the core intelligent module for fine-tuning and optimizing the large language model, forming a closed-loop evolution mechanism.

2. The intelligent sales lead scheduling system based on intent strength and historical feedback as described in claim 1, characterized in that, The collection and integration of sales lead data related to the current sales lead specifically includes: a data set of each call task, a dynamically updated set of historical outreach records, and a set of business background data. The data set for each call task includes the text content of the call after each AI call task ends, the complete audio recording of the call, and the call metadata containing the call duration and call time. The dynamically updated historical outreach record set includes extracting and recording the interaction status and results of sales leads in historical marketing activities. The interaction status and results include at least: the connection status of historical phone calls, the sending and delivery status of historical text messages, and the success or failure status of adding social contact information. The business background data set includes the business scenario in which the call task takes place, and includes at least: industry classification, the theme of this marketing campaign, and information on the main products being promoted.

3. The intelligent sales lead scheduling system based on intent strength and historical feedback as described in claim 1, characterized in that, The specific steps for creating a structured clue file are as follows: By standardizing and cleaning the collected sales lead data and unifying its format, the unstructured call text content is segmented and denoised, and audio files of different formats are converted into a unified encoding format and sampling rate. Based on a preset timestamp sequence, the call text content, corresponding audio segments, and events during the call are aligned and associated. The cleaned and aligned data is merged with the preset historical contact record set and business background data set, and encapsulated into a structured data object, which is then output to the multimodal feature analysis module as the structured clue file.

4. The intelligent sales lead scheduling system based on intent strength and historical feedback as described in claim 1, characterized in that, The text semantic feature analysis submodule includes a domain pre-training fine-tuning unit, an intent classification unit, and an emotion quantification unit. The domain pre-training fine-tuning unit performs incremental fine-tuning by injecting domain corpus from the AI ​​telephone sales scenario, so that the model has domain-adaptive semantic understanding capabilities. The intent classification unit constructs a domain-specific intent recognition template, which includes at least the following intents: product function inquiry intent, price inquiry intent, preferential policy inquiry intent, purchase reservation intent, intent to express objections to needs, intent to have no needs at the moment, and intent to repeat inquiries. The emotion quantification unit performs sentiment polarity analysis on the transcribed text and identifies complex emotions in the text to generate a multi-dimensional emotion feature vector. The text semantic feature analysis submodule fuses the intent confidence vector with the multi-dimensional emotion feature vector to form a text semantic feature set.

5. The intelligent sales lead scheduling system based on intent strength and historical feedback as described in claim 1, characterized in that, The speech acoustic feature analysis submodule includes an acoustic feature extraction unit, a multimodal emotion recognition unit, and a feature fusion unit; The acoustic feature extraction unit preprocesses the call audio data to extract core acoustic features, thereby forming a high-dimensional acoustic feature matrix. The multimodal emotion recognition unit adopts a multimodal large model, which performs cross-modal fusion of the high-dimensional acoustic feature matrix and the semantic features of the ASR transcribed text, captures the correlation between speech intonation and text content through an attention mechanism, and outputs the recognition confidence of each complex emotion and the timestamp information of the emotion triggering segment. The feature fusion unit generates a speech acoustic feature set by normalizing the recognition confidence of each complex emotion with the high-dimensional acoustic feature matrix. The speech acoustic feature analysis submodule concatenates and aligns the speech acoustic feature set with the text semantic feature set output by the text semantic feature analysis submodule to form complete high-dimensional contextual data.

6. The intelligent sales lead scheduling system based on intent strength and historical feedback as described in claim 1, characterized in that, The process of performing end-to-end holistic contextual understanding and reasoning to generate quantified cue intention intensity scores and decision rationales specifically includes: The core intelligent module includes a context understanding Prompt template, which contains multimodal feature association guidance, historical feedback fusion rules, and domain requirement reasoning logic. After receiving user historical contact records and business domain meta-information from the structured clue archive, as well as the text semantic feature set and speech acoustic feature set output by the multimodal feature analysis module, the multi-source information is structured and injected into a pre-trained large-scale language model to achieve cross-dimensional association understanding of text semantics, speech emotion, historical behavior, and business scenarios. Based on the results of cross-dimensional correlation understanding, hierarchical reasoning is performed through a pre-set intention assessment dimension system. A weighted summation algorithm is used to calculate the scores of each dimension. The weight values ​​are obtained by training a logistic regression model from historical feedback data and support dynamic updates. Based on the output clue intention intensity score, a structured decision reason is generated, which includes scores for each first-level dimension, key supporting evidence, and potential concerns, ensuring that the intention assessment results are interpretable.

7. The intelligent sales lead scheduling system based on intent strength and historical feedback as described in claim 1, characterized in that, The specific process for autonomously planning the optimal sequence of subsequent actions by combining historical feedback data is as follows: By calling the historical feedback database through the data association interface, a set of similar clues is obtained, and the follow-up strategy type, execution sequence, tool call combination and final conversion result data corresponding to each historical clue in the set are extracted; By training on historical data of similar clue sets, the expected return score of each candidate follow-up strategy is output. Based on the ranking results of expected return scores, the optimal action sequence is planned autonomously. The optimal action sequence includes tool call type, execution order, triggering conditions and timeout handling mechanism. According to the decision reason of the current clue, personalized execution parameters are matched for the action sequence to ensure that the scheduling strategy is accurately adapted to the clue characteristics.

8. The intelligent sales lead scheduling system based on intent strength and historical feedback as described in claim 1, characterized in that, The feedback of sales business results after the core intelligent module schedules and executes the data collection specifically includes: Obtain multi-dimensional sales business result feedback data, including core conversion result data, process feedback data, and strategy adaptability feedback data; The collected feedback data is structured, with core conversion result data mapped to binary labels, process feedback data having key information extracted and transformed into structured fields, and strategy adaptability feedback data quantified into adaptability scores. Establish a mechanism to link feedback data with original clues. By using unique clue identifiers, structured feedback data is bound to corresponding structured clue profiles, intent intensity scores, and scheduling action sequences. At the same time, outlier removal and normalization are performed on the data to ensure the quality of training data.

9. The intelligent sales lead scheduling system based on intent strength and historical feedback as described in claim 1, characterized in that, The specific steps for fine-tuning and optimizing a large language model are as follows: By adopting a hierarchical fine-tuning strategy, the closed-loop training dataset is grouped according to the intention intensity range, the lead conversion result, and the strategy fit. Different fine-tuning targets are set for different groups of data. For feedback data of high-intention but unsuccessful leads, the focus is on optimizing the weight parameters of the intention evaluation dimension and the accuracy of potential concern identification. For feedback data of low strategy fit, the focus is on optimizing the decision logic of action sequence planning. By combining a domain-adaptive loss function, the parameters of the top-level Transformer layer of a large language model are updated, while the parameters of the bottom-level pre-trained layers remain fixed to avoid forgetting. Establish a fine-tuning effect evaluation and iteration mechanism to verify the effect of the fine-tuned model. If the index amplitude is lower than the preset index amplitude threshold, the training data volume of the corresponding group will be automatically expanded and the fine-tuning process will be re-executed. At the same time, the model will be fine-tuned with full data every preset period. Otherwise, the fine-tuning process will not be performed.

10. A sales lead intelligent scheduling method based on intent strength and historical feedback, wherein the sales lead intelligent scheduling method based on intent strength and historical feedback implements the sales lead intelligent scheduling system based on intent strength and historical feedback as described in any one of claims 1-9, characterized in that, The method includes: After each AI call task is completed, sales lead data related to the current sales lead is automatically collected from multiple call task data sources and integrated to form a structured lead profile. The collected sales lead data is subjected to deep feature extraction to obtain contextual data, so as to provide processed, high-dimensional information input. The multimodal feature analysis module includes a text semantic feature analysis submodule and a speech acoustic feature analysis submodule. Based on the contextual data provided by the structured clue archive and the multimodal feature analysis module, end-to-end overall contextual understanding and reasoning are performed to generate quantified clue intention intensity scores and decision reasons. Based on this, combined with historical feedback data, the system autonomously plans the optimal sequence of subsequent actions and finally automatically executes the scheduling task by calling the preset atomic toolset. The sales results feedback after the core intelligent module schedules and executes are collected and fed back to the core intelligent module as training data. This data is then used to fine-tune and optimize the large language model, forming a closed-loop evolution mechanism.

Images