Intelligent sales assisting method and system based on AI large model

Abstract

The invention discloses an intelligent sales assistance method and system based on an AI large model, and aims to realize real-time intelligent assistance, communication whole-process redisk and salesperson targeted training in a face-to-face communication process between salespersons and customers, and finally achieve the purposes of improving service efficiency, reducing operation cost and enhancing customer experience. The method comprises the following steps: a preparation stage: a system is connected with a CRM system to obtain customer related information, and a multi-modal large model generates a customer portrait and a personalized reception plan based on the customer information; in the execution stage, the terminal layer collects audio and video data in the sales and customer communication process, preprocesses the audio and video data and transmits the audio and video data to the cloud, a dialogue context is constructed after the audio and video data is processed by the data processing and capability layer, and a multi-modal large model of the core decision layer completes customer intention recognition, emotion analysis, knowledge retrieval and auxiliary content generation based on the dialogue context. The information is fed back to sales personnel through the terminal layer; and in the redisk stage, the large model automatically generates conference summary, action items and personalized improvement suggestions based on complete communication data.

Description

TECHNICAL FIELD

[0001] The application relates to an AI large model-based intelligent sales assistance method and system, which is suitable for enterprise market sales, pre-sale consultation and post-sale follow-up scenes, can realize real-time intelligent assistance in face-to-face communication between sales personnel and customers, communication whole-process review and targeted training of sales personnel, and finally achieves the goal of improving service efficiency, reducing operation cost and enhancing customer experience, and belongs to the technical field of intelligent equipment. BACKGROUND

[0002] In the process of enterprise product sales, monitoring, analyzing and providing assistance to the dialogue process between sales personnel and customers is a key means to improve sales success rate. At present, the industry mainly relies on wearable devices such as intelligent earphones and intelligent glasses, combined with traditional natural language processing (NLP) technology to realize related functions.

[0003] Among them, the intelligent earphone has audio playback, Bluetooth connection and basic AI interaction capabilities, and can realize voice interaction, real-time translation and other functions, but can only provide simple voice assistance and cannot provide accurate support in combination with enterprise business knowledge. Intelligent glasses integrate augmented reality (AR), sensors and display technology, and can realize real-time information display, environmental perception and other functions, and can present basic data in the field of view of sales personnel, but the recognition accuracy and response speed of customer intent are limited. Traditional NLP technology is used for analysis and understanding of dialogue content, but due to the model architecture, the recognition accuracy of customer questions with ambiguous language, logical jumps or irregular grammar is low, the generated reply suggestions lack professionalism and fluency, and cannot efficiently interface with enterprise knowledge base. In addition, although general large language models (LLMs) have shown strong capabilities in natural language understanding and text generation, they can handle complex language scenarios and generate human-like text, but currently there is no complete solution that deeply integrates with intelligent terminal devices and is targeted at enterprise sales scenarios, and they cannot fully realize their value in real-time assistance, knowledge retrieval, emotion analysis and review training, etc., resulting in sales personnel still facing problems such as "high pressure to remember a large amount of product information", "inability to respond to customer objections in a timely manner", and "lack of data support for post-review". SUMMARY

[0004] The purpose of the present application is to overcome the defects of "low customer intention recognition accuracy", "weak real-time assistance capability", "inefficient knowledge docking" and "lack of targeted review training" in existing sales assistance technologies, and to provide an intelligent sales assistance method and system based on an AI large model, an intelligent sales assistance technical solution deeply integrating an AI large model and intelligent terminal equipment (such as smart glasses and smart earphones), which realizes real-time monitoring, accurate analysis, intelligent response of sales dialogues, and automatic review and personalized training afterwards, helps sales personnel to quickly respond to customer needs and grasp the communication rhythm on the premise of maintaining natural communication, and finally assists in improving the sales success rate.

[0005] The purpose of the present application is realized by the following technical solutions:

[0006] 1. An intelligent sales assistance method based on an AI large model, comprising the following steps:

[0007] Step 1: Preparation phase, the system interfaces with a CRM system to obtain customer-related information, and a multi-modal large model generates a customer portrait and a personalized reception plan based on the customer information, providing pre-support for sales communication;

[0008] Step 2: Execution phase, the terminal layer collects audio and video data during the sales and customer communication process and performs preprocessing, the preprocessed data is transmitted to the cloud, and the dialog context is constructed after data processing and capability layer processing, the multi-modal large model of the core decision layer completes customer intention recognition, emotion analysis, knowledge retrieval and auxiliary content generation based on the dialog context, and the generated auxiliary content is fed back to the sales personnel through the terminal layer;

[0009] Step 3: Review phase, the multi-modal large model automatically generates a meeting minutes, action items and personalized improvement suggestions based on complete communication data, the action items and review report are synchronized to the CRM system, and the high-quality sales cases after desensitization processing are included in the enterprise knowledge base for system optimization and team training.

[0010] The purpose of the present application can also be further realized by the following technical measures:

[0011] The aforementioned intelligent sales assistance method based on an AI large model, in step 2, the multi-modal large model can recognize ambiguous, skipping or non-standard grammar customer expressions, and generate smooth and professional response dialogues.

[0012] The aforementioned intelligent sales assistance method based on an AI large model, in step 2, the multi-modal large model analyzes the emotional state through customer tone, pause, speech speed and body language, and generates communication strategy adjustment prompts.

[0013] In step 3 of the aforementioned intelligent sales assistance method based on an AI big model, the debriefing report includes customer needs, objections, communication highlights and shortcomings, and generates personalized training suggestions.

[0014] The aforementioned intelligent sales assistance method based on an AI large model, wherein the multimodal large model is customized and trained according to the needs of the enterprise, and supports model quantization and compression, knowledge distillation technology.

[0015] The aforementioned intelligent sales assistance method based on an AI large model, wherein the multimodal large model supports knowledge augmentation RAG technology, stores enterprise knowledge through a semantic vector database, and connects product attributes, customer needs and success cases with a knowledge graph to achieve dynamic contextual retrieval.

[0016] The aforementioned intelligent sales assistance method based on an AI large model, wherein the multimodal large model supports multilingual understanding and generation, enables real-time translation, and the customer's speech is translated and then broadcast through headphones or displayed on an AR screen.

[0017] The aforementioned intelligent sales assistance method based on an AI large model has a full-link response time of less than 500ms for the terminal layer's speech recognition, large model inference, and TTS feedback.

[0018] An intelligent sales support system that implements any of the aforementioned methods includes a terminal layer, a transmission and storage layer, a data processing and capability layer, a core decision-making layer, and an application and feedback layer, with a security and compliance module running through each layer;

[0019] The terminal layer includes AI glasses and / or smart earphones, and is configured with a perception module, a feedback module, and a terminal computing module; the perception module is used to collect voice data, customer facial expressions, body language, and environmental document information; the feedback module is used to transmit auxiliary information through visual display or voice broadcast; the terminal computing module is used for local lightweight preprocessing and network connection management.

[0020] The transmission and storage layer includes a cloud storage and computing platform, which enables secure encrypted data transmission via 5G or Wi-Fi 6, and stores raw data, processed data, and enterprise knowledge base data;

[0021] The data processing and capability layer includes a speech processing unit, a vision processing unit, and a data fusion unit; the speech processing unit converts speech into text and identifies the speaker; the vision processing unit identifies key objects, customer emotions, and body language; the data fusion unit aligns text, visual tags, and timestamps to construct a dialogue context.

[0022] The core decision-making layer includes an enterprise knowledge base, a multimodal large model, and a prompt word engineering module; the enterprise knowledge base stores the enterprise's private knowledge; the multimodal large model has multimodal understanding, intent recognition, sentiment analysis, and content generation capabilities; the prompt word engineering module is configured with prompt word templates specifically for sales scenarios;

[0023] The application and feedback layer includes a real-time support unit, a post-meeting debriefing unit, and a team insight unit; the real-time support unit provides real-time support such as script prompts and objection handling; the post-meeting debriefing unit generates meeting minutes and improvement suggestions; and the team insight unit provides team data analysis.

[0024] The security and compliance module includes a privacy protection unit, an access control unit, and a compliance audit unit, which enables data anonymization, access control, and log auditing.

[0025] The aforementioned intelligent sales support system employs AES+HTTPS encrypted transmission, RBAC access control mechanism, and local desensitization of sensitive information in its security and compliance module.

[0026] Compared with the prior art, the beneficial effects of the present invention are:

[0027] 1. Achieve a shift from "post-event review" to "real-time empowerment": Through the collaboration of smart terminals and large models, provide instant assistance in key real-time scenarios of sales and customer communication, solving the limitation of traditional sales training that can only be analyzed after the fact. Provide rapid support at key points such as customer questions and objections, helping sales personnel to grasp the rhythm of communication.

[0028] 2. Improve the accuracy of intent recognition and response generation: By using a multimodal large model to replace the traditional small model for NLP analysis, it can accurately identify customers' vague, disjointed, or grammatically incorrect expressions of intent. At the same time, the generated response scripts are more fluent, professional, and polite. Combined with knowledge augmentation technology, it ensures that the response content is consistent with the company's knowledge and avoids "illusions".

[0029] 3. Reduce the memory burden on sales staff and realize the function of "enhancing external brain": The system seamlessly connects to the enterprise knowledge base, retrieves product information, competitor comparisons, customer background and other data in real time, and transmits them to sales staff in an invisible way, so that sales staff do not need to memorize massive amounts of information and can focus more on emotional connection and communication with customers.

[0030] 4. Achieve large-scale empowerment of "organizational wisdom": Incorporate the experience, standard scripts, and success stories of top salespeople into the corporate knowledge base, and empower every team member through a large model to ensure the standardization and professionalism of sales communication and narrow the capability gap among team members.

[0031] 5. Multimodal fusion enhances the targeted nature of assistance: By combining multi-dimensional information such as voice, text, customer emotions, and body language, the assistance suggestions provided are more tailored to specific communication scenarios. For example, the emotion warning function can remind sales staff to adjust their communication strategies in a timely manner to improve the customer experience.

[0032] 6. Comprehensive security and compliance safeguards: Through local data anonymization, encrypted transmission, access control, and operation log auditing, we ensure that customer privacy is not compromised and corporate data is secure, complying with relevant laws and regulations and applicable to various sales scenarios.

[0033] 7. Supports cross-language communication and global expansion: Multi-modal large-scale model with multi-language support enables real-time translation, breaks down language barriers, and helps enterprises expand into international markets. Attached Figure Description

[0034] Figure 1 This is a design diagram of the product architecture of the present invention;

[0035] Figure 2 This is a functional block diagram of the present invention;

[0036] Figure 3 This is a typical business process diagram of the present invention. Detailed Implementation

[0037] The intelligent sales assistance method based on an AI large model of the present invention specifically includes the following steps:

[0038] Step 1: Preparation phase. The system connects with the CRM system to obtain relevant customer information. The multimodal big data model generates customer profiles and personalized reception plans based on the customer information, providing upfront support for sales communication.

[0039] The system is deeply integrated with the enterprise CRM system, allowing sales personnel to query customer basic information (company size, industry attributes, contact person's position), company background, historical communication records, past purchase data, and other customer-related data through the terminal layer or related devices (such as mobile phones and tablets), providing comprehensive information support for communication preparation;

[0040] Based on acquired customer information and system-preset analysis rules, the multimodal big data model automatically generates customer profile analysis (e.g., "a medium-sized manufacturing enterprise that focuses on product cost-effectiveness, has consulted competitor A, and has concerns about after-sales response speed"), competitor analysis (competitor's core strengths, weaknesses, and differences from the company's products), and outputs personalized reception plans. These plans include clear visit objectives, appropriate opening remarks, a targeted list of key questions, precise product value propositions, and pre-set responses to common objections, providing sales personnel with a standardized yet personalized communication framework.

[0041] Step 2: Execution phase. The terminal layer collects and preprocesses audio and video data during the communication process between sales and customers. The preprocessed data is transmitted to the cloud. After data processing and capability layer processing, a dialogue context is constructed. The multimodal big model of the core decision layer completes customer intent recognition, sentiment analysis, knowledge retrieval and auxiliary content generation based on the dialogue context. The generated auxiliary content is fed back to the sales personnel through the terminal layer.

[0042] The terminal layer's perception module collects real-time voice data from sales and customers, customer facial expressions, body movements, and document information in the communication environment (such as competitor brochures and customer-provided requirement documents). The terminal computing module performs preprocessing operations such as noise reduction and initial filtering of redundant information on the collected raw data to improve data quality and lay the foundation for subsequent processing.

[0043] The pre-processed audio and video data is stably transmitted to the cloud via a secure encrypted network (5G or Wi-Fi 6) in the transmission and storage layer. The voice processing unit in the data processing and capability layer uses ASR (speech-to-text) technology to convert the voice data into text and complete the speaker identification (distinguishing between sales personnel and customers). The visual processing unit uses CV (computer vision) technology to identify the customer's emotional state (happy, confused, suspicious, averse, etc.) and body language (nodding, shaking, crossing arms, etc.). The data fusion unit accurately aligns the voice-processed text, the visually processed emotion and body language tags, and the corresponding timestamp information to construct a complete and coherent dialogue context.

[0044] The multimodal big model of the core decision-making layer receives the constructed dialogue context and, combined with the sales scenario-specific template of the prompt word engineering module, sequentially completes customer intent recognition, sentiment analysis, knowledge retrieval, and auxiliary content generation.

[0045] Customer intent recognition: accurately captures customers' core needs and potential problems, and can accurately identify their true intent even when faced with customers' ambiguous language, logical leaps, or grammatically incorrect expressions;

[0046] Sentiment analysis: By integrating multi-dimensional information such as customer tone, pause frequency, speech rate changes and body language, it can judge the customer's emotional state (such as hesitation, doubt, satisfaction, aversion, etc.) in real time, and provide data support for adjusting communication strategies.

[0047] Knowledge retrieval: Quickly retrieve factual knowledge related to customer problems and needs from the enterprise knowledge base, such as product information (parameters, functions, application scenarios), competitor comparison data, pricing schemes, success stories, and pre-sales and after-sales FAQs;

[0048] Supporting content generation: Based on intent recognition results, sentiment analysis conclusions, and retrieved knowledge, generate fluent, professional, polite, and scenario-appropriate response scripts, product introduction content, objection rebuttal suggestions, and communication strategy adjustment prompts to ensure the practicality and relevance of the supporting content;

[0049] The application and feedback layer transmits the auxiliary content generated by the multimodal large model to sales personnel through the feedback module of the terminal layer: complex information such as product parameters and competitor comparison charts are presented in visual form through the AR display of AI glasses, while simple information such as keyword prompts and brief sales script suggestions are transmitted in voice broadcast form through bone conduction headphones, achieving discreet and timely assistance without interfering with the natural communication between sales and customers.

[0050] Step 3: In the debriefing phase, the multimodal big model automatically generates meeting minutes, action items, and personalized improvement suggestions based on complete communication data. The action items and debriefing report are synchronized to the CRM system. High-quality sales cases, after being anonymized, are included in the enterprise knowledge base for system optimization and team training.

[0051] After the communication, the multimodal big data model automatically generates structured meeting minutes based on complete dialogue texts, sentiment analysis records, body language tags and other multi-dimensional communication data. It accurately extracts the client's core needs, objections raised, consensus reached by both parties and clear action items (such as "provide a detailed quotation within 3 working days" and "arrange for a technical team to conduct a product demonstration").

[0052] The system quantifies and evaluates communication quality from multiple dimensions, including question quality, listening ratio, objection handling effectiveness, and depth of needs mining. It updates customer profiles (supplementing information such as new needs and potential concerns discovered during communication) and generates personalized improvement suggestions that include communication highlights, shortcomings, and specific improvement directions, forming targeted training guidance.

[0053] The generated review reports and action items are automatically synchronized to the enterprise CRM and workflow system, automatically creating or updating customer follow-up records, and setting reminders for action item completion time to ensure the timeliness and standardization of sales follow-up;

[0054] Successful sales cases and high-quality response scripts, after being anonymized (by removing sensitive customer information and core corporate secrets), are incorporated into the enterprise knowledge base. On the one hand, they are used for the continuous training and optimization of multimodal large models to improve the system's auxiliary capabilities; on the other hand, they serve as training materials for the sales team, enabling the large-scale reuse of high-quality experience.

[0055] To achieve better technical results, the method further employs a multimodal large model that supports knowledge enhancement (RAG, Retrieval-Augmented Generation) technology. This technology stores enterprise knowledge through a semantic vector database and connects product attributes, customer needs, success stories, and other information with a knowledge graph, enabling dynamic contextual retrieval, avoiding the generation of "illusionary" content, and ensuring the accuracy and reliability of auxiliary content.

[0056] Furthermore, the multimodal large model used in the method supports multilingual understanding and generation, enabling real-time translation. After the customer's speech is translated by the large model, the translated text is broadcast to the salesperson in their native language via bone conduction headphones or displayed on the AI ​​glasses screen, supporting cross-language communication and expanding the applicable scenarios of the method.

[0057] Furthermore, the end-to-end response time for speech recognition, large model inference, and TTS (text-to-speech) feedback at the terminal layer in the method is less than 500ms, ensuring that real-time assistance is smooth without significant lag and guaranteeing the fluency of sales and customer communication.

[0058] Furthermore, the multimodal large model used in the method can be customized for training and optimization according to enterprise needs, supports model quantization and compression, knowledge distillation technology, transfers the capabilities of the large model to the lightweight model, adapts to the local computing needs of the terminal layer, and improves the flexibility and adaptability of the method implementation.

[0059] In this technical solution, the intelligent sales assistance based on a large model implemented by the above method includes a terminal layer, a transmission and storage layer, a data processing and capability layer, a core decision-making layer, and an application and feedback layer. Furthermore, a security and compliance module is integrated across all layers, and the collaboration of each layer provides support for the implementation of the method. The specific structure and functions are as follows:

[0060] 1. Terminal Layer: Includes AI glasses and / or smart earphones, configured with a perception module, a feedback module, and a terminal computing module; the perception module is used to collect voice data, customer facial expressions, body language, and environmental document information; the feedback module is used to transmit auxiliary information through visual display or voice broadcast; the terminal computing module is used for local lightweight preprocessing and network connection management;

[0061] The perception module consists of a microphone array and a high-definition camera. Its core function is to collect voice data from sales and customers, customer facial expressions and body language, and document information in the communication environment, providing raw input data for method execution and ensuring the comprehensiveness and accuracy of data collection.

[0062] Feedback module: Includes a miniature AR display screen and bone conduction headphones, used to deliver auxiliary information generated by the system to sales personnel in the form of visual display (complex data, charts) or voice broadcast (brief prompts, scripts). The delivery process ensures that it does not interfere with the natural communication between sales and customers, corresponding to the auxiliary information delivery link in the method.

[0063] Terminal computing module: It undertakes local lightweight data processing tasks, including noise reduction and preliminary filtering of raw audio and video data. It is also responsible for caching processed data and managing network connections (5G / Wi-Fi 6) to ensure the stability and timeliness of data transmission and support the data preprocessing and transmission requirements in the method.

[0064] 2. Transmission and storage layer: This includes cloud storage and computing platforms, which enable secure encrypted data transmission via 5G or Wi-Fi 6, and store raw data, processed data, and enterprise knowledge base data;

[0065] Utilizing high-speed 5G or Wi-Fi 6 networks to achieve secure encrypted data transmission (AES+HTTPS) ensures that pre-processed data at the terminal layer is transmitted to the cloud efficiently and securely.

[0066] The cloud storage and computing platform includes object storage and database. Object storage is used to store raw audio and video data, while the database is used to store processed text data, enterprise knowledge base data, customer profile data, etc., providing reliable support for data storage and retrieval at each stage of the method.

[0067] 3. Data Processing and Capability Layer: This layer includes a speech processing unit, a visual processing unit, and a data fusion unit. The speech processing unit converts speech into text and identifies the speaker. The visual processing unit identifies key objects, customer emotions, and body language. The data fusion unit aligns text, visual tags, and timestamps to construct a dialogue context.

[0068] Voice processing unit: It adopts ASR (speech-to-text) technology. Its core function is to convert the collected voice data into text with high precision and complete the speaker identification (distinguishing between sales personnel and customers), clarify the subject of the dialogue, and meet the voice data processing requirements of the corresponding method.

[0069] Visual processing unit: Based on CV (computer vision) technology, it can identify key objects (products, business cards, competitor information, etc.) in the video stream, as well as judge customer emotional state (happy, confused, disgusted, etc.) and body language (nodding, crossed arms, etc.), supporting the emotion and behavior analysis link in the method.

[0070] Data fusion unit: Accurately aligns the text after speech processing, the tags after visual processing, and the corresponding timestamp information to construct a complete and coherent dialogue context, providing unified and standardized input data for the core decision-making layer and ensuring the accuracy of context analysis in the method.

[0071] 4. Core Decision-Making Layer: Includes an enterprise knowledge base, a multimodal large model, and a prompt word engineering module; the enterprise knowledge base stores the enterprise's private knowledge; the multimodal large model has multimodal understanding, intent recognition, sentiment analysis, and content generation capabilities; the prompt word engineering module is configured with prompt word templates specifically for sales scenarios;

[0072] Enterprise knowledge base: Stores proprietary enterprise knowledge such as product manuals, competitor information, standard scripts, success stories, CRM customer information, pre-sales and after-sales FAQs, and pricing plans, providing factual support for multimodal large models, ensuring the accuracy of generated content, and meeting the knowledge retrieval needs of corresponding methods;

[0073] Multimodal large model: Large language models with multimodal understanding capabilities such as DeepSeek, which can be selected, include DeepSeek. Its core capabilities include customer intent recognition based on dialogue context, sentiment analysis, knowledge retrieval and content generation, and support for state memory of multi-turn conversations to ensure the coherence of communication assistance. It is the core of the method's intelligent decision-making.

[0074] The prompt word engineering module configures dedicated prompt word templates for sales scenarios (objection handling, product introduction, demand mining, etc.). Through dynamic context management and dialogue state tracking, combined with few-shot learning and small-shot adaptation techniques, it guides the multimodal large model to generate accurate and professional auxiliary content, thereby improving the quality of auxiliary content in the method.

[0075] 5. Application and Feedback Layer: This includes a real-time support unit, a post-meeting debriefing unit, and a team insight unit. The real-time support unit provides real-time support such as prompts for dialogue and handling of objections. The post-meeting debriefing unit generates meeting minutes and improvement suggestions. The team insight unit provides team data analysis.

[0076] Real-time support unit: Transforms the output of the multimodal big model into specific sales support actions such as script prompts, objection handling, emotion warning, quick customer background check, product information insight, and competitor comparison prompts, providing immediate support for the method execution phase;

[0077] Post-meeting debriefing unit: Enables functions such as automatic generation of structured meeting minutes, extraction of action items, communication quality analysis, updating of customer profiles, and generation of personalized improvement suggestions, supporting various tasks in the methodology debriefing phase;

[0078] Team Insights Unit: Provides managers with a team dashboard, which analyzes data such as common team problems, high-frequency objections, and distribution of best practices to support the optimization of training systems and the adjustment of sales strategies, extending the team empowerment value of the methodology.

[0079] 6. Security and Compliance Module: Includes a privacy protection unit, an access control unit, and a compliance audit unit, enabling data anonymization, access control, and log auditing;

[0080] Privacy Protection Unit: Sensitive information such as faces, customer names, contact information, and corporate secrets are anonymized and obfuscated before data processing to protect customer privacy and corporate data security, in accordance with data compliance requirements in the implementation of the method.

[0081] Access Control Unit: Employs RBAC (Role-Based Access Control) mechanism to strictly control data access permissions. Sales personnel can only access their own relevant customer data and communication records, while managers can access the team's overall data. Administrators have system configuration and full data viewing permissions to prevent data leakage.

[0082] Compliance Audit Unit: Records all data access, operation, and system interaction logs, with a log storage period of ≥1 year, to ensure that the implementation of system processes and methods complies with relevant laws and regulations such as GDPR and the Personal Information Protection Act, facilitating audit traceability.

[0083] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0084] Figure 1 This is a system architecture diagram of the present invention, showing the hierarchical relationship of "intelligent terminal layer - transmission and storage layer - data processing layer - core decision layer - application feedback layer - security and compliance layer" and the core modules of each layer;

[0085] Figure 2 This is a functional block diagram of the present invention, showing the specific sub-modules of the four core functions of "real-time intelligent coach", "scenario-enhanced memory", "full-process navigator" and "intelligent review assistant" and their interaction with the large model and enterprise knowledge base;

[0086] Figure 3 This is a typical business process diagram of the present invention. Taking "customer raises price objection" as an example, it shows the complete process of "data collection - transmission - processing - big model decision - feedback - sales response".

[0087] like Figure 1 As shown:

[0088] (I) Product architecture design, the core components of the system are described as follows:

[0089] 1. Terminal Layer (AI Glasses & Headphones)

[0090] Perception module: responsible for collecting first-person perspective video (customer expressions, environment, documents) and audio (mutual dialogue).

[0091] Feedback module: Information is delivered discreetly to salespeople through AR visual cues (text, charts) and headphone audio (voice prompts) without interrupting natural communication.

[0092] Computing module: Performs local lightweight processing (such as noise reduction and preliminary filtering) and manages network connectivity.

[0093] 2. Transmission and Storage Layer

[0094] Secure transmission: Encrypted data is transmitted stably and with low latency to the cloud via high-speed networks such as 5G / Wi-Fi 6.

[0095] Cloud platform: Provides scalable storage (for storing raw audio and video data and processed text) and computing resources (GPU servers for model inference).

[0096] 3. Data Processing and Capability Layer

[0097] Automatic Speech Processing (ASR): Converts speech conversations into text with high accuracy and identifies the speaker.

[0098] Visual processing (CV): Analyze video streams to identify key objects (products, business cards, competitor information), customer emotions (happy, confused, averse), and body language (nodding, crossed arms).

[0099] Data fusion: Aligning information such as text, visual tags, and timestamps to construct a complete "dialogue context" that can be understood by large models.

[0100] 4. Core Decision-Making Layer (Large Model Engine)

[0101] Enterprise Knowledge Base: This is the system's "memory," storing product manuals, competitor information, standard scripts, success stories, CRM customer information, and more. The large model retrieves information from here to ensure the accuracy of its responses.

[0102] Multimodal large model: the "brain" of the system. It receives and processes multimodal information, combines it with a knowledge base to perform deep reasoning, analysis, and content generation (e.g., generating coping strategies, answering questions, and summarizing key points).

[0103] Prompt engineering: Guide the large model to complete a specific task (such as "analyze the current stage of the dialogue and generate a phrase to facilitate the transaction") through carefully designed prompts, ensuring that the output is accurate and useful.

[0104] 5. Application and Feedback Layer

[0105] Real-time assistance: Translate the decisions of the large model into specific sales actions, such as real-time script prompts, process navigation, risk warnings (e.g., negative customer sentiment), and information inquiries (e.g., "Let me introduce the XX function").

[0106] Post-meeting debriefing: Automatically generates meeting minutes, communication quality scores, and personalized improvement suggestions to help salespeople grow continuously.

[0107] Team Insights: Provides managers with a team dashboard to analyze common team problems and best practices, which can be used to optimize training systems and sales strategies.

[0108] 6. Safety and compliance (overall)

[0109] Privacy protection: Sensitive information (such as faces and customer names) is anonymized and obscured before data processing.

[0110] Access control: Strictly control data access permissions (e.g., salespeople can only see their own data, and managers can see team data).

[0111] Compliance audit: Record all data access and operation logs to ensure that processes comply with laws and regulations (such as GDPR and personal information protection laws).

[0112] This architecture demonstrates a complete, implementable, and highly intelligent sales support system. Its core value lies in seamlessly combining the cognitive capabilities of large models with the immersive experience of AI smart terminals, empowering frontline sales personnel.

[0113] (II) Core Functions

[0114] like Figure 2 The diagram shown is a functional module diagram. The main functions are as follows:

[0115] 1. Real-time intelligent coach

[0116] Functional objective: To provide timely strategic guidance during the conversation to help sales cope with complex situations.

[0117] Application scenarios:

[0118] Objection handling: When a customer says "the price is too high", generate a value reiteration script in real time.

[0119] Script Tips: Best practice scripts are provided for specific scenarios (such as product introductions).

[0120] Emotional alert: By analyzing customer tone and micro-expressions, sales staff are reminded to adjust their communication strategies.

[0121] Terminal presentation: Voice prompts are delivered discreetly and promptly via bone conduction headphones.

[0122] 2. Contextual Enhancement of Memory

[0123] Functional goal: To expand sales' "memory capacity" and provide accurate background information when needed.

[0124] Application scenarios:

[0125] Quick Customer Background Check: Simply ask "What's this customer's recent news?" to get an instant summary.

[0126] Product Information Perspective: When looking at a product, the smart glasses automatically display key parameters and differentiating advantages.

[0127] Competitive comparison prompt: When a customer mentions a competitor, core comparative data will be automatically pushed to them.

[0128] Terminal presentation: Complex information is displayed via AR through smart glasses, while simple information is read aloud via voice.

[0129] 3. End-to-end Navigator

[0130] Functional objective: To ensure the integrity and standardization of the sales process and avoid omitting key steps.

[0131] Application scenarios:

[0132] Stage guidance: Identify the current stage of the conversation (needs exploration → solution demonstration → sales closure).

[0133] Next step suggestion: After completing the current stage, intelligently recommend the next action.

[0134] Key Question Reminder: Remind sales staff to ask those easily overlooked but crucial questions.

[0135] Terminal presentation: The process is displayed through voice prompts via headphones and an AR progress bar on the glasses.

[0136] 4. Intelligent Review Assistant

[0137] Functional goal: To automatically complete post-meeting data collection and analysis, helping sales to continuously improve.

[0138] Application scenarios:

[0139] Automatic meeting minutes: Automatically generate structured meeting minutes based on complete conversations.

[0140] Communication quality analysis: Evaluate communication effectiveness from multiple dimensions (such as question quality and listening ratio).

[0141] Action item extraction: Automatically identifies and lists the next to-do items, and synchronizes them to the CRM.

[0142] Terminal presentation: Detailed reports can be viewed via mobile app or website backend after the meeting.

[0143] 5. Core AI Engine

[0144] Functional goal: To provide intelligent support for all upper-level functions.

[0145] Components:

[0146] Multimodal large model: responsible for understanding, reasoning, decision-making and content generation, it is the "brain" of the system.

[0147] Enterprise knowledge base: Stores product information, customer data, successful sales scripts, sales processes, etc., ensuring the accuracy and professionalism of all suggestions.

[0148] 6. Smart Terminal Interaction Layer

[0149] Functional objective: To serve as the interaction interface between the system and sales personnel, enabling the natural and efficient transmission of information.

[0150] Equipment division of labor:

[0151] Smart glasses: Primarily responsible for presenting visual information, suitable for displaying complex content such as charts, data, and documents.

[0152] Bone conduction headphones: primarily responsible for transmitting voice information, suitable for real-time, discreet conversation assistance.

[0153] based on Figure 3 The typical business process diagram shown is illustrated in the following embodiment of the present invention:

[0154] A large-scale model-based intelligent sales assistance method is used to improve sales efficiency. The specific implementation steps are as follows:

[0155] 1. Preparation stage:

[0156] Before visiting a customer, sales staff can log in to the system via a mobile app. The system automatically synchronizes the customer's basic information (e.g., a medium-sized manufacturing company that mainly processes auto parts, contact person: Mr. Zhang, purchasing manager, who inquired about our company's product A six months ago but did not complete the transaction due to price issues, and recently browsed the official website of competitor B), historical communication records, and related needs and pain points from the CRM.

[0157] The multimodal big model generates a customer profile based on the above customer information: "The customer is a medium-sized manufacturing enterprise with product upgrade needs, pays attention to cost-effectiveness and after-sales response speed, is price-sensitive, and has had contact with competitor B." At the same time, it generates a competitor analysis: "Competitor B's price is lower than our company's product, but its after-sales network coverage is insufficient, its core functions meet basic production needs, and it has no customized services."

[0158] The system outputs a personalized reception plan: The visit objective is to "introduce the core advantages of our company's Model C product (cost-effective version), dispel price concerns, emphasize after-sales guarantee, and encourage product demonstration intentions"; the opening remarks are: "Mr. Zhang, thank you for taking the time to receive me! I understand that your company has recently been optimizing its production line. Six months ago, you consulted us about our Model A product. Today, I would like to recommend Model C, which is more suitable for the needs of small and medium-sized manufacturing enterprises. It has targeted optimizations in cost-effectiveness and after-sales response"; the list of key questions includes "What is your company's current daily production capacity target?" "Are there any pain points in the after-sales response of the equipment currently used?"; the pre-set objection response script includes: "Regarding the price, although the unit price of Model C is 5% higher than that of competitor B, its energy consumption is 12% lower, annual maintenance costs are reduced by 18%, and the overall operating cost is lower. In addition, we have 3 after-sales service outlets in the local area, and the response time is no more than 4 hours."

[0159] 2. Execution Phase:

[0160] Sales staff wearing AI glasses and bone conduction headphones communicate with Mr. Zhang. The microphone array collects the voice of the conversation between the two parties in real time, and the high-definition camera captures Mr. Zhang's facial expressions (frowning, nodding) and the brochure of competitor B on the table.

[0161] The terminal computing module performs noise reduction processing on the voice data and transmits it to the cloud via 5G network with encryption. The voice processing unit converts the voice into text (Mr. Zhang: "Your C model product is still more expensive than the competitor's B product, and can your after-sales service really respond within 4 hours?"). The visual processing unit recognizes that Mr. Zhang's expression is slightly suspicious and his body language is that his arms are crossed.

[0162] The data fusion unit aligns text, "doubt" emotion tags, "crossed arms" body language tags, and timestamps to construct a dialogue context and input it into a multimodal large model;

[0163] The large model uses knowledge augmentation (RAG) technology to retrieve comprehensive cost comparison data between product C and competitor B from the enterprise knowledge base, local after-sales service network distribution, and response time commitment letter. Combined with prompt word templates, it generates a response script: "Mr. Zhang, your concerns about price and after-sales service are very important! First, regarding price, we have done the calculations, and the comprehensive operating cost (procurement + energy consumption + maintenance) of product C is 8% lower than that of competitor B. Here is the detailed calculation table (simultaneously displayed on the AR screen of the AI ​​glasses); second, regarding after-sales service, we have spare parts readily available at our three local after-sales service outlets, and we promise on-site service within 4 hours. Here is our after-sales response commitment letter (which can be sent to you after the meeting)." This is then transmitted to the sales staff via bone conduction headphones.

[0164] During the communication, Mr. Zhang mentioned his "concern about product compatibility with existing production lines," and his tone was hesitant. The big data model recognized the customer's concerns through sentiment analysis and generated a prompt in real time: "The customer has compatibility concerns. You can emphasize the product's compatibility and free adaptation and debugging services." Based on this, the sales staff added: "Mr. Zhang, please rest assured that the C model product supports interface adaptation for mainstream production lines. We will arrange a technical team to come to your location free of charge for adaptation and debugging to ensure normal operation."

[0165] 3. Post-mortem review phase:

[0166] After the communication, the system automatically generated meeting minutes: the core requirement was "to purchase products with high cost-performance ratio, fast after-sales response, and compatibility with existing production lines"; the objections were "the price is higher than competitor B, there are doubts about the speed of after-sales response, and there are concerns about product compatibility"; the consensus reached was "to provide a comprehensive cost calculation sheet and after-sales commitment letter within 3 working days, and to arrange for the technical team to conduct an on-site assessment of compatibility next week"; the action items were "sales personnel: send the calculation sheet and commitment letter within 3 days; technical team: conduct an on-site assessment next Tuesday";

[0167] System communication quality assessment: The highlight was "timely response to the customer's core objections and supporting the viewpoint with data"; the shortcoming was "failure to proactively inquire about the customer's procurement budget and timeline"; the improvement suggestion is "in future communications, we can increase the inquiry into key information such as procurement budget and timeline, such as 'What is the approximate budget range for this procurement? When do you plan to complete the equipment upgrade?'"

[0168] Action items and review reports are automatically synchronized to the CRM system, and a reminder to "send the data within 3 days" is set.

[0169] The high-quality responses from this communication (such as responses to comprehensive cost comparisons and after-sales guarantee instructions) have been anonymized and incorporated into the company's knowledge base for use in subsequent large-scale model training and new employee training.

[0170] Based on the foregoing technical solutions and embodiments, the core technology and technical advantages of the present invention are as follows:

[0171] 1. Core Technology: Intelligent and Low-Latency Collaboration on the Terminal Side

[0172] Traditional sales support systems primarily utilize marketing call center agents to identify customer intent by converting the agent's voice messages into text and displaying supplementary suggestions via pop-up windows on the agent's computer for reference. The new design uses smart terminals to collect customer voice messages, converting them into text via edge computing (Edge AI) on local devices such as mobile phones or tablets and uploading it to the server. The system's response text is then converted into voice or image locally and transmitted to the smart terminal, reducing transmission latency.

[0173] Key technologies: Utilizing the built-in voice and image acquisition functions of smart terminals, a lightweight model (ASR / TTS) is deployed on the terminal side for real-time preprocessing.

[0174] Technical advantages: The entire process of voice recognition, large model inference, and TTS feedback is optimized, with a response time of less than 500ms, ensuring timely response on the sales site without any obvious lag or delay, and maintaining a natural communication experience.

[0175] 2. Core Technology: Large Model Semantic Understanding and Generation (LLM Core)

[0176] Traditional sales support systems use customized small models for NLP semantic analysis, primarily employing word segmentation, keyword detection, and word vector analysis to parse customer intent and provide relevant response suggestions. The new solution uses large language model technology to replace traditional models like BERT for NLP analysis. This allows for a more accurate understanding of customer intent, extraction of key dialogue points, and the simultaneous generation of intelligent responses and sales suggestions. This enables functions such as supplementary knowledge-based Q&A, competitor analysis, and sales script recommendations.

[0177] Compared to traditional NLP, the introduction of large models offers the following technical advantages: The latest large models, such as GPT, DeepSeek, and Qianwen, can be adopted on demand and customized for training and optimization based on enterprise needs, ensuring accurate multi-intent recognition and low latency to support real-time dialogue. It supports dialogue context management: maintaining state memory across multiple turns of conversation, enabling the system to understand the evolution of customer intent. It can utilize knowledge augmentation (RAG, Retrieval-Augmented Generation): combining with the enterprise's proprietary knowledge base, such as CRM, FAQs, and product documents, to provide factual support for LLM and avoid "illusions." It supports multilingual understanding and generation, enabling global sales support.

[0178] Core technical points:

[0179] Multimodal large models: Supports text, speech, and visual multimodal understanding capabilities, with contextual long-term memory and reasoning abilities. Supports real-time streaming processing and response.

[0180] Prompt Engineering: Dedicated prompt templates optimized for sales scenarios, dynamic context management and dialogue state tracking, and few-shot learning and few-shot adaptation techniques.

[0181] Model optimization techniques: model quantization and compression to ensure real-time response; knowledge distillation to transfer the capabilities of large models to lightweight models; continuous learning and online fine-tuning mechanisms.

[0182] Technological advantages: The large model technology can understand customer intent in real time. Based on the RAG engine, the knowledge enhancement results in accurate and consistent answers, and can generate precise answers, rebuttals, questions and recommendations, providing sales with high-quality intelligent scripts.

[0183] 3. Core Technologies: Knowledge Augmentation and Enterprise Knowledge Integration

[0184] Traditional sales knowledge support systems primarily retrieve relevant knowledge points by matching keywords to entries in the company's FAQ knowledge base, presenting multiple retrieved entries simultaneously to agents for them to choose the appropriate one. This model is unsuitable for face-to-face sales support. The new solution requires building a semantic vector database integrated with a large-scale model to store company knowledge, using a knowledge graph to connect information such as product attributes, customer needs, and success stories. This allows for real-time retrieval of knowledge from the company's CRM, ERP, and FAQ systems, and supports dynamic contextual retrieval.

[0185] Core technical points:

[0186] Knowledge graph / structured index: Construct a semantic network of products, customers, prices, and cases to enable semantic retrieval.

[0187] Integrate AIAgent with your enterprise CRM: Read customer history, order status, and preference tags to generate personalized sales recommendations.

[0188] Develop an AI Agent for post-sales processing: After the sales process is completed, use a large model to automatically generate meeting minutes annotations, summarize customer concerns and objections, uncover potential needs, provide optimization directions, archive relevant records, and generate corporate sales training materials.

[0189] Technological advantages: Based on the RAG engine, knowledge enhancement ensures accurate and highly consistent answers across enterprises. This allows AI to go beyond simply "talking," truly understanding the company's product and customer logic. Every question asked in the sales field is supported by data.

[0190] 4. Core Technologies: Security and Compliance Technologies

[0191] Traditional agent support systems are typically deployed privately in client rooms, with centralized agent management and lower security requirements. Newer sales support systems, however, have higher security and compliance requirements because the smart terminals worn by sales personnel are located in public spaces and connect wirelessly or via Bluetooth. This solution requires end-to-end encryption and local anonymization of data: recordings, voice texturization, and sentiment analysis must all be anonymized beforehand. Data isolation and access control are implemented: different sales roles access different levels of information. An auditing mechanism is also established to trace the source and reasoning logic of AI prompts, facilitating auditing and trust building.

[0192] Core technical points:

[0193] Local anonymization and encryption of voice and text data during transmission (AES+HTTPS).

[0194] Access control (RBAC) and log tracing mechanisms are adopted.

[0195] Provides explainable tracking to ensure corporate compliance and customer privacy.

[0196] Technological advantages: Local anonymization and enterprise private cloud deployment comply with data compliance requirements (GDPR / Chinese cryptographic standards). It can provide enterprises with AI solutions to ensure that sales data and customer information are not leaked.

[0197] In addition to the above embodiments, the present invention may have other implementation methods. All technical solutions formed by equivalent substitution or equivalent transformation fall within the protection scope claimed by the present invention.

Claims

1. A smart sales assistance method based on an AI large-scale model, characterized in that, Includes the following steps: Step 1: Preparation phase. The system connects with the CRM system to obtain relevant customer information. The multimodal big data model generates customer profiles and personalized reception plans based on the customer information, providing advance support for sales communication. Step 2: Execution phase. The terminal layer collects and preprocesses audio and video data during the communication process between sales and customers. The preprocessed data is transmitted to the cloud. After data processing and capability layer processing, a dialogue context is constructed. The multimodal big model of the core decision layer completes customer intent recognition, sentiment analysis, knowledge retrieval and auxiliary content generation based on the dialogue context. The generated auxiliary content is fed back to the sales personnel through the terminal layer. Step 3: In the debriefing phase, the multimodal big model automatically generates meeting minutes, action items, and personalized improvement suggestions based on complete communication data. The action items and debriefing report are synchronized to the CRM system, and high-quality sales cases after anonymization are included in the enterprise knowledge base for system optimization and team training.

2. The intelligent sales assistance method based on an AI large model according to claim 1, characterized in that, In step 2, the multimodal large model can identify customer expressions that are vague, disjointed, or grammatically incorrect, and generate fluent and professional response scripts.

3. The intelligent sales assistance method based on an AI large model according to claim 1, characterized in that, In step 2, the multimodal big model analyzes the customer's emotional state through tone of voice, pauses, speech rate and body language, and generates communication strategy adjustment prompts.

4. The intelligent sales assistance method based on an AI large model according to claim 1, characterized in that, In step 3, the debriefing report includes customer needs, objections, communication highlights and shortcomings, and generates personalized training suggestions.

5. The intelligent sales assistance method based on an AI large model according to claim 1, characterized in that, The multimodal large model is customized for training according to enterprise needs and supports model quantization and compression, as well as knowledge distillation techniques.

6. The intelligent sales assistance method based on an AI large model according to claim 1, characterized in that, The multimodal large model supports knowledge-enhanced RAG technology, which stores enterprise knowledge through a semantic vector database and connects product attributes, customer needs, and success stories with a knowledge graph to achieve dynamic contextual retrieval.

7. The intelligent sales assistance method based on an AI large model according to claim 1, characterized in that, The multimodal large model supports multilingual understanding and generation, enabling real-time translation. The client's speech is translated and then broadcast through headphones or displayed on an AR screen.

8. The intelligent sales assistance method based on an AI large model according to claim 1, characterized in that, The end-to-end response time for speech recognition, large model inference, and TTS feedback at the terminal layer is less than 500ms.

9. An intelligent sales assistance system implementing the method of any one of claims 1-8, characterized in that, The system includes a terminal layer, a transmission and storage layer, a data processing and capability layer, a core decision-making layer, and an application and feedback layer, with security and compliance modules running through all layers. The terminal layer includes AI glasses and / or smart earphones, and is configured with a perception module, a feedback module and a terminal computing module; the perception module is used to collect voice data, customer facial expressions, body language and environmental document information. The feedback module is used to deliver auxiliary information through visual display or voice broadcast; the terminal computing module is used for local lightweight preprocessing and network connection management. The transmission and storage layer includes a cloud storage and computing platform, which enables secure encrypted data transmission via 5G or Wi-Fi 6, and stores raw data, processed data, and enterprise knowledge base data; The data processing and capability layer includes a speech processing unit, a vision processing unit, and a data fusion unit; the speech processing unit converts speech into text and identifies the speaker; the vision processing unit identifies key objects, customer emotions, and body language; the data fusion unit aligns text, visual tags, and timestamps to construct a dialogue context. The core decision-making layer includes an enterprise knowledge base, a multimodal large model, and a prompt word engineering module; the enterprise knowledge base stores the enterprise's private knowledge; the multimodal large model has multimodal understanding, intent recognition, sentiment analysis, and content generation capabilities; the prompt word engineering module is configured with prompt word templates specifically for sales scenarios; The application and feedback layer includes a real-time support unit, a post-meeting debriefing unit, and a team insight unit; the real-time support unit provides real-time support such as script prompts and objection handling; the post-meeting debriefing unit generates meeting minutes and improvement suggestions; and the team insight unit provides team data analysis. The security and compliance module includes a privacy protection unit, an access control unit, and a compliance audit unit, which enables data anonymization, access control, and log auditing.

10. The intelligent sales assistance system according to claim 9, characterized in that, The security and compliance module uses AES+HTTPS encrypted transmission, RBAC access control mechanism, and local desensitization of sensitive information.

Images