AI Agent marketing system and method based on large model

Abstract

The invention discloses an AI Agent marketing system and method based on a large model, and the method comprises the steps: carrying out the automatic call-out and conversation through a call-out robot according to an intention customer list, and obtaining the contact information of a customer according to the conversation content; starting a chat robot according to the customer contact information, and guiding the customer to activate a voice service by using the chat robot; constructing a customer portrait according to the historical dialogue content, analyzing a customer intention level, matching a recommendation scheme from a marketing scheme database according to the customer portrait and the customer intention level, and obtaining a recommendation verbal skill by using a chat robot; and pushing the recommended verbal skill and the recommended scheme to a worker for manual follow-up. The invention relates to the technical field of intelligent marketing, and solves the problems of inaccurate customer intention recognition, unreasonable marketing resource allocation, poor customer experience and the like in an existing marketing method.

Description

Technical Field

[0001] The present invention belongs to the field of marketing and involves artificial intelligence technology. Specifically, it is an AI Agent marketing system based on a large model. Background Art

[0002] In current insurance business, common marketing methods include manual outbound calls and traditional robot outbound calls. Manual outbound calls rely on staff to manually dial customer phones for business promotion, while traditional robot outbound calls automatically dial phones according to pre-set programs and repeat fixed scripts.

[0003] Nowadays, to protect customer privacy, insurance companies use dynamic virtual number technology and permission isolation design when storing customer contact information in the database. When a customer purchases insurance, the system automatically generates a virtual intermediate number (such as "021-5XX-XXXXX") bound to the real phone number and stores it in the policy data. The outbound robot can call the real number by linking with the operator's virtual number pool through the API interface, while only the virtual number is displayed on the marketer's background interface. In this case, the efficiency of manual outbound calls is extremely low. Facing a large amount of customer data, staff need to input virtual numbers one by one and wait for connection, which is a cumbersome and time-consuming process. Moreover, when dialing a virtual number, it is impossible to directly obtain detailed customer information, and it is necessary to frequently switch systems to search, further reducing the outbound call efficiency. At the same time, it is difficult for manual outbound calls to accurately screen customers. It mainly relies on the experience of staff to judge customer intentions, lacks scientific basis, and cannot comprehensively analyze customer historical data, resulting in the inability of outbound call resources to accurately focus on high-intent customers. Frequent outbound calls to low-intent customers are likely to arouse customer disgust and reduce the customer's favorability towards the insurance company. In addition, manual outbound calls are difficult to adapt to the complex operations under privacy protection requirements. Frequent switching of virtual numbers increases the operation complexity, and the permission isolation design restricts staff from obtaining customer information, making it difficult to provide personalized services and affecting the customer experience.

[0004] Traditional robots lack an intelligent screening mechanism. Usually, they make outbound calls at a fixed frequency and according to fixed rules, and cannot adjust strategies based on real-time customer feedback and potential needs. Facing virtual numbers, they are unable to effectively analyze customer historical data, and it is difficult to distinguish high-intent and low-intent customers. Frequent outbound calls to a large number of low-intent customers cause harassment, leading to customer complaints and damaging the brand image of the insurance company. Moreover, the traditional robot's script is single and cannot be adjusted according to real-time customer feedback and personalized characteristics. When making outbound calls using virtual numbers, it is unable to conduct personalized communication by combining customer historical information, resulting in a low outbound call conversion rate. In addition, the traditional robot's data processing and analysis capabilities are insufficient. After communicating with customers, it is unable to effectively extract key information from the conversation and is difficult to feed the data back into subsequent marketing decisions. In an environment of virtual numbers and permission isolation, data integration and analysis are more difficult, and it is impossible to provide valuable customer insights for insurance companies and achieve precision marketing. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art; for this purpose, the present invention proposes an AI Agent marketing system based on a large model, which is used to solve the problems of inaccurate customer intention recognition, unreasonable marketing resource allocation, poor customer experience, and difficulty in adapting to complex and changeable marketing scenarios in the existing marketing methods.

[0006] To achieve the above object, the present invention provides an AI Agent marketing system based on a large model, including:

[0007] Intelligent outbound call module: used to automatically make outbound calls and have conversations according to the list of potential customers by using an outbound call robot, and obtain customer contact information according to the conversation content; wherein, the outbound call robot is built based on artificial intelligence algorithms and is used for automatically making outbound calls and having interactive voice interactions with customers;

[0008] Automatic interaction module: used to start a chat robot according to the customer contact information, and use the chat robot to guide the customer to activate the voice service; wherein, the chat robot is built based on artificial intelligence algorithms and is used for parsing text content and having interactive interactions with customers;

[0009] Portrait construction module: used to construct a customer portrait based on the historical conversation content and analyze the customer intention level, match and recommend a plan from the marketing plan database according to the customer portrait and the customer intention level, and obtain recommended conversation scripts by using the chat robot;

[0010] Manual follow-up module: used to push the recommended conversation scripts and recommended plans to the staff for manual follow-up.

[0011] It should be noted that the outbound call robot is an artificial intelligence software that can automatically make phone calls and have voice interactions with customers; the chat robot is an artificial intelligence software that can have conversation interactions with users through natural language, and can actively initiate questions and answers or chats according to the set goals to guide the customer to activate the voice system; and both the outbound call robot and the chat robot are built based on big data and artificial intelligence technologies.

[0012] Furthermore, the automatic outbound call by using the outbound call robot according to the list of potential customers includes:

[0013] A1. Collect historical policy data in the insurance database, screen out users who have not renewed their policies within a preset time period, have a click-through rate greater than a preset threshold, and have no complaint records from the historical policy data to obtain a list of potential customers;

[0014] A2. According to the policy data in the list of potential customers, obtain the difference T between the current time and the policy expiration time 到期 , when T 到期When it is less than the preset expiration threshold, an outbound call is triggered;

[0015] A3, record the customer ID, outbound call time t 外呼 and the connection status as well as the call transcription text to obtain an outbound call record form;

[0016] A4, determine whether the call transcription text contains the customer's contact information; if yes, save the outbound call record form; if not, determine whether to make a second outbound call according to the call transcription text.

[0017] Further, the determination of whether to make a second outbound call according to the call transcription text includes:

[0018] A41, use natural language processing technology to perform word segmentation on the call transcription text to obtain a number of call words;

[0019] A42, calculate the cosine similarity between a number of call words and the words in the insurance vocabulary list, and mark the call words with a cosine similarity greater than the preset similarity threshold as keywords to obtain a number of keywords; among them, the insurance vocabulary list is a custom common vocabulary list related to insurance business;

[0020] A43, according to the formula calculate the second outbound call score Retry_Score; where E dialog represents the information entropy of a number of call words, and is obtained according to the formula P(c i ) represents the word frequency of the call word c i , N key represents the number of a number of keywords, T re represents the time difference from when the outbound call robot asks the customer to the customer's first response;

[0021] A44, save the customers with a second outbound call score Retry_Score greater than the preset score threshold to the second outbound call queue, make a second outbound call after t 外呼 +T, and jump to A3; where T represents the preset outbound call cycle.

[0022] By calculating the information entropy E dialog of the call content, the information value of the conversation is quantified, and high-value conversations where the customer clearly expresses needs (such as asking about premium details) are screened out, avoiding repeated disturbances to invalid calls (such as customers' perfunctory responses); the cosine similarity matching based on keywords can accurately identify insurance-related intentions (such as "renewal" "claim settlement"), ensuring that the outbound call target focuses on customers with potential needs; in addition, through the exponential decay weight (1 - e^(-0.1T re )) of the response time difference, it is ensured to give priority to following up customers with quick responses.

[0023] Further, the method of guiding a customer to activate a voice service by using a chatbot includes:

[0024] B1. Using the chatbot to parse the call transcription text through a first instruction statement to obtain the customer's initial interest;

[0025] B2. Using the chatbot to conduct an interactive conversation based on the customer's initial interest and automatically guide the customer to activate the voice service according to the conversation content;

[0026] B3. Judging whether the voice service is activated; if yes, saving the customer to the manual service queue; if not, analyzing the historical conversation content to obtain a dynamic access period.

[0027] Further, the method of analyzing the historical conversation content to obtain a dynamic access period includes:

[0028] B31. Using a predefined sentiment dictionary to match the number of positive words N 正 and the number of negative words N 负 appearing in the historical conversation content, calculating the ratio of N 正 to N 负 to obtain the customer trust score S trust ;

[0029] B32. Collecting the cumulative number N reject of times the customer refuses to activate the voice service, and calculating the dynamic access period T next = T base ×(1 + γ × N reject / (1 + S trust )); where T base represents a preset basic access period, and γ represents a period adjustment coefficient obtained based on historical experience.

[0030] In the calculation formula of the dynamic access period, the customer's emotional tendency is reflected through the trust score, and according to the formula, the follow-up interval for high-trust customers (such as those who mention "good service" many times) is shortened to seize the conversion window period; for low-trust customers (such as those who complain about "high price"), the period is extended to avoid losing customers due to excessive harassment; the adjustment coefficient γ in the formula is determined and optimized through historical data to balance the follow-up frequency and customer experience, improve the reach efficiency of high-intent customers, and effectively reduce the customer complaint rate, achieving a double optimization of resource allocation and customer experience.

[0031] Further, the method of constructing a customer portrait based on the historical conversation content includes:

[0032] C1-1. Collect the historical conversation corpus of several customers, and use the BIO tagging system to label the categories of the corpus to obtain a labeled dataset, and use the labeled dataset to train a BiLSTM-CRF model to obtain an interest extraction model;

[0033] Among them, the labeled categories include price sensitivity, security requirements, service efficiency, and additional benefits, and the input of the keyword extraction model is text data, and the output is several words and the labeled category of each word;

[0034] C1-2. Input the historical conversation content into the keyword extraction model to obtain several keywords and interest types;

[0035] C1-3. Sort the sentences containing keywords in the historical conversation content in chronological order to obtain the sentence position ranking;

[0036] C1-4. According to the formula I k =∑(TF-IDF(w i ,k)×e^(-β×rank(wi))) / N u Calculate the interest intensity I k of the customer for the kth type; where, TF-IDF(w i ,k) represents the TF-IDF value of the keyword w i in the interest type k, rank(w i) represents the sentence position ranking where the keyword w i first appears in the historical conversation content, and N u represents the total number of conversation turns;

[0037] C1-5. Sort according to the interest intensity I k in the order of price sensitivity, security requirements, service efficiency, and additional benefits to obtain the customer portrait P client =[I1, I2, I3, I4].

[0038] The interest extraction model breaks through the traditional rule limitations and can identify the implicit needs of customers (such as the price sensitivity implied when customers repeatedly compare products), while the time decay gives higher weight to early conversations (such as the first mention of "medical security" reflecting the core needs), and the obtained quantified interest intensity results can generate a more accurate customer portrait to guide the matching of marketing plans.

[0039] Further, the method for obtaining the customer intention level includes:

[0040] C2-1. Statistically calculate the time from the first conversation to the consent to activate the voice service to obtain the activation time T 激活 ;

[0041] C2-2, calculate the customer intention level L according to the formula ; where intent ; among them, respectively represent the reciprocal of the activation time T 激活 and the total number of dialogue turns N u , k represents the regression coefficient, and λ represents the interaction weight coefficient.

[0042] Based on the logistic regression model, non-linearly map the activation time and the number of dialogue turns to the intention level of 0-1, ensuring that customers who activate the voice service in a short time and have few rounds of dialogue obtain a higher level and are preferentially assigned artificial seats.

[0043] Furthermore, the matching and recommending the solution from the marketing solution database according to the customer portrait and the customer intention level includes:

[0044] C3-1, extract the marketing portraits P plan of several marketing solutions and the lowest intention level L plan from the marketing solution database;

[0045] C3-2, calculate the matching degree between the customer and the j-th marketing solution according to the hybrid similarity formula where cos() represents the cosine similarity calculation formula;

[0046] C3-3, screen out the marketing solution with the highest matching degree according to to obtain the recommended solution.

[0047] Furthermore, the method of obtaining the recommended conversation script by using the chatbot includes:

[0048] D1, extract the personal information of the customer from the insurance database according to the customer ID; among them, the personal information includes the basic information, insurance history, and claim settlement records of the customer, etc.;

[0049] D2, input the personal information, customer portrait, intention level, and recommended solution of the customer into the chatbot, and use the second instruction statement to guide the chatbot to generate the recommended conversation script of the recommended solution.

[0050] The second aspect of the present invention provides an AI Agent marketing method based on a large model, including:

[0051] S1, use the outbound robot to automatically make outbound calls and have conversations according to the list of potential customers, and obtain the customer contact information according to the conversation content;

[0052] S2. Start the chatbot according to the customer's contact information, and use the chatbot to guide the customer to activate the voice service. Among them, the chatbot is constructed based on artificial intelligence algorithms and is used to parse text content and interact with customers interactively;

[0053] S3. Build a customer profile based on the historical conversation content and analyze the customer's intention level. Match and recommend a solution from the marketing solution database according to the customer profile and the customer's intention level, and obtain the recommended sales talk using the chatbot;

[0054] S4. Push the recommended sales talk and the recommended solution to the staff for manual follow-up.

[0055] Compared with the prior art, the beneficial effects of the present invention are:

[0056] Through the intelligent screening and dynamic adjustment mechanism, the present invention significantly improves the friendliness of outreach. In terms of outreach screening, the system accurately identifies high-intent customers by analyzing the information volume and response time of the call content, avoiding repeated interruption of ineffective calls. At the same time, combined with the emotional trust assessment in the customer conversation, the follow-up cycle is dynamically adjusted: for customers with high trust, the cycle is shortened to seize business opportunities, while for customers with low trust or those who repeatedly refuse to provide contact information, the outreach interval is extended to reduce adversarial communication, effectively reducing the possibility of customers having negative emotions due to excessive outreach and improving the overall communication experience;

[0057] The present invention uses natural language processing technology to extract implicit needs from the conversation and combines time decay weights to distinguish core demands from secondary demands. On this basis, by quantifying the customer's intention level, high-intent customers are preferentially assigned to manual services to optimize resource utilization. In addition, the hybrid similarity matching mechanism comprehensively considers the customer's interests and product suitability, avoiding forcibly recommending high-price solutions to low-intent customers, thereby increasing the acceptance rate of the recommended marketing products;

[0058] The present invention supports the real-time iteration of the outbound sales talk and the recommended solution. For example, the outbound call robot automatically updates the keyword library based on historical data to match hot demands, and the chatbot optimizes the guiding sales talk through conversation feedback. At the same time, the full-link data from outbound calls to transactions is automatically synchronized to the customer profile, triggering real-time strategy adjustments. In terms of resource allocation, the system intelligently allocates high-intent customers to the manual seats, and low-intent customers enter the automated cultivation process, significantly reducing ineffective human input. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0060] Figure 1 It is a schematic flow chart of an AI Agent marketing system based on a large model provided by the present invention;

[0061] Figure 2 It is a schematic framework diagram of an AI Agent marketing system based on a large model provided by the present invention;

[0062] Figure 3 It is a schematic work flow diagram of the intelligent outbound call module provided by the present invention. Specific Embodiments

[0063] The following will clearly and completely describe the technical solutions of the present invention in conjunction with the embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0064] Please refer to Figure 1 - Figure 3 , an embodiment of the first aspect of the present invention provides an AI Agent marketing system based on a large model, including:

[0065] Intelligent outbound call module: used to automatically make outbound calls and have conversations according to the list of potential customers by using an outbound call robot, and obtain customer contact information according to the conversation content;

[0066] Automatic interaction module: used to start a chat robot according to the customer contact information, and use the chat robot to guide the customer to activate the voice service; wherein, the chat robot is constructed based on artificial intelligence algorithms and is used to parse text content and interact with the customer interactively;

[0067] Portrait construction module: used to construct a customer portrait based on the historical conversation content and analyze the customer intention level, match and recommend a plan from the marketing plan database according to the customer portrait and the customer intention level, and obtain recommended conversation scripts by using the chat robot;

[0068] Manual follow-up module: used to push the recommended conversation scripts and recommended plans to the staff for manual follow-up.

[0069] Based on the above technical modules, the present invention reduces customer harassment and improves the efficiency of reaching customers through intelligent screening, precise matching and dynamic strategies, thus realizing the transformation from "extensive marketing" to "personalized service". Its core value lies in replacing experience-driven with data-driven, balancing customer experience and business goals through technical means, and providing the industry with a more scientific and sustainable marketing solution.

[0070] The outbound calling strategy of traditional marketing systems mostly uses fixed frequency outbound calls or simple rules to reach customers, which is easy to repeatedly harass low-intent customers, not only wasting resources, but also likely to cause customer disgust. Therefore, in the present invention, the potential intention and response characteristics of customers are analyzed through the first call of the outbound calling robot, and the outbound calling strategy is dynamically adjusted;

[0071] Specifically, the system first screens out high-potential customers based on historical policy data, and dynamically triggers outbound calls based on the policy expiration date; in the first call, the system uses natural language processing technology to extract keywords (such as insurance business-related words) in the customer conversation, combines information entropy to quantify the information density of the conversation, and evaluates the communication efficiency based on the customer's response time difference;

[0072] Based on the above multi-dimensional data, the system dynamically screens customers who need to be reached again through the secondary outbound call scoring formula, and only conducts periodic secondary outbound calls to customers with high information volume and high response willingness, while low-scoring customers automatically exit the contact process. Compared with the traditional model, the present invention focuses outbound call resources on high-value customer groups through intelligent decision-making, which not only avoids the interference of invalid outbound calls on low-intent customers, but also improves conversion efficiency through accurate recommendations, achieving a balanced optimization of resource allocation and customer experience.

[0073] In this embodiment, the intelligent outbound calling module first screens out users who have not renewed their insurance within a preset time period, whose click rate is greater than a preset threshold and who have no complaint records from the historical policy data in the insurance database, to form a list of potential customers;

[0074] Then, based on the policy data of the low-intent customer list, calculate the difference T between the current time and the policy expiration time 到期 When T 到期 When the amount is less than the preset expiration threshold, the outbound call robot is activated to automatically make outbound calls;

[0075] When making an outbound call, the outbound call robot will have an interactive conversation with the customer and guide the customer to provide contact information. At the same time, it will record the customer ID, outbound call time, connection status, and call transcription text to generate an outbound call record table;

[0076] Determine whether the call transcription contains the customer contact information based on the text. If yes, directly save the outbound call record table and transmit the customer contact information to the chatbot; if not, determine whether to make another outbound call based on the call transcription text.

[0077] Generally, in order to protect customer privacy, when insurance companies store customer contact information in the database, they will adopt dynamic virtual number technology and permission isolation design. When a customer purchases insurance, the system will automatically generate a virtual middle number (such as "021-5XX-XXXXX") bound to the real phone number and store it in the policy data. The outbound robot is linked with the operator's virtual number pool through the API interface and can directly dial the real number, but only the virtual number is displayed on the marketer's background interface. Therefore, it is necessary to use the outbound robot to automatically complete customer reach and information interaction.

[0078] In one implementation, to determine whether the customer contact information is included in the call transcription text, the following operations can be performed:

[0079] After obtaining the call transcription text, it is necessary to preprocess it first to convert the text into a format that is easier to process;

[0080] Then, according to phone numbers, WeChat IDs, etc., construct the corresponding keyword library; phone number keywords may include "phone", "number", "mobile phone number", etc.; WeChat ID keywords such as "WeChat", "WeChat ID", etc. Use the string matching algorithm to search for the words in the keyword library in the preprocessed text; when a keyword is detected, mark its location and the corresponding keyword type. If "My phone number is" appears in the text, mark the "phone" keyword and its appearance location;

[0081] Next, obtain the contact information according to specific extraction rules:

[0082] For phone number keywords, in the text near the keyword, use regular expressions to match the common formats of phone numbers. For example, domestic mobile phone numbers are generally 11 digits, and the regular expression is "1[3-9]\d{9}";

[0083] For WeChat ID keywords, the extraction rules are relatively flexible. Commonly, they are composed of letters, numbers, and underscores, and the length is between 6 and 20 digits. The regular expression "[a-zA-Z0-9_]{6,20}" can be used to match and extract near the keyword;

[0084] Next, the extracted contact information may be incorrect or incomplete, and it is necessary to perform validity verification; finally, output and store the qualified customer contact information.

[0085] In one implementation, to determine whether to make a second outbound call according to the call transcription text, the following steps may be included:

[0086] A41, use natural language processing technology to perform word segmentation on the call transcription text to obtain a number of call words;

[0087] A42. Calculate the cosine similarity between several call vocabulary and the vocabulary in the insurance vocabulary list, mark the call vocabulary with a cosine similarity greater than the preset similarity threshold as keywords, and obtain several keywords; among them, the insurance vocabulary list is a custom common vocabulary list related to insurance business;

[0088] A43. According to the formula calculate the retry call score Retry_Score; where E dialog represents the information entropy of several call vocabulary, and is obtained according to the formula P(c i ) represents the word frequency of the call vocabulary c i , N key represents the number of several keywords, and T re represents the time difference from when the outbound robot asks the customer to the customer's first response;

[0089] A44. Save the customers with a retry call score Retry_Score greater than the preset score threshold to the retry call queue, make a retry call after t 通话 + T, and jump to A3; where T represents the preset outbound cycle.

[0090] Suppose in the auto insurance business of an insurance company, the intelligent outbound module screens out customers who have not renewed their auto insurance policies in the previous year, frequently browse auto insurance - related pages recently (click - through rate is greater than the preset threshold) and have no complaint records from the insurance database, and forms a list of potential customers;

[0091] When the time remaining until the policy expiration is less than the preset expiration threshold, the outbound robot makes an automatic outbound call;

[0092] Suppose the outbound robot talks to customer A. After customer A answers the call, the conversation mentions content such as "Has the auto insurance premium increased recently?" and "The previous claim settlement speed was too slow". The outbound robot records the customer ID, outbound time, connection status, and call transcription text. And when the outbound robot asks if the customer can provide contact information for better service, if customer A provides a phone number, save the outbound call record form and transmit it to the automatic interaction module.

[0093] For customer B, who refuses to provide contact information and indicates that they plan to switch to another auto insurance company. At this time, the outbound robot will apologize to the customer and further analyze the call transcription text to determine whether there is a possibility of making a further outbound call to this customer;

[0094] Through word - segmentation processing, keywords such as "claim settlement" and "not cost - effective" are found in the conversation with B, and are determined as keywords after matching with the insurance vocabulary list;

[0095] Suppose after calculation, the information entropy E of this call vocabulary is obtaineddialog =3, number of keywords N key =5, the time difference between the outbound robot asking the question and the customer's first response T re = 3 seconds;

[0096] According to the formula The calculation results show that Retry_Score = [3 / log(1+5)] × [1-e^(-0.1×3)] ≈ 0.9. If the preset score threshold is 0.8, then Retry_Score is greater than the preset score threshold, and customer B is saved in the outbound call queue again. 外呼 Add the preset outbound calling cycle (such as 3 days) and then make another outbound call.

[0097] When the outbound call robot obtains the customer's contact information and call transcription text, the automatic interaction module will automatically add the customer's contact information and then start the chatbot; the chatbot begins to parse the call record text and perform the subsequent task of guiding the customer to activate the voice service; specifically, the workflow of the automatic interaction module includes:

[0098] First, the chatbot will use the first command sentence to parse the call transcript. Through natural language processing technology, it can identify key information in the text and determine the customer's initial interest. If the call record text mentions "I am concerned about the price of auto insurance", the chatbot can parse out that the customer's initial interest is the price of auto insurance;

[0099] In this embodiment, the first instruction statement may be: "Analyze the user's interest in purchasing auto insurance based on the call transcription text";

[0100] Next, based on the customer's initial interest, the chatbot engages in an interactive conversation with the customer, providing targeted information and guidance, and trying to get the customer to activate the voice service. If the customer is interested in the price of auto insurance, the chatbot can introduce the current auto insurance promotions, the price differences of different packages, etc., and guide the customer to activate the voice service during the conversation;

[0101] During the interaction with the customer, the chatbot determines in real time whether the voice service is activated. If the customer successfully activates the voice service, the automatic interaction module will save the customer to the manual service queue so that the manual agent can follow up and provide the customer with more professional services. If the voice service is not activated, the next step is to analyze the historical conversation content to obtain the dynamic access cycle.

[0102] In one implementation, obtaining a dynamic access cycle based on historical conversation content analysis may include the following steps:

[0103] First, use a predefined sentiment dictionary to match the number of positive and negative words that appear in the historical conversation content (including the call transcription text of this outbound call and any previous conversation records).

[0104] Suppose in a conversation, positive words (such as "satisfied" and "good") appear 3 times, and negative words (such as "too expensive" and "troublesome") appear 1 time. Then the customer trust score S trust = 3÷1 = 3;

[0105] Next, collect the cumulative number of times N that the customer has refused to activate the voice service reject ; Suppose the preset basic access period T base is 7 days, the cycle adjustment coefficient γ is 0.5, and the cumulative number of times N that the customer has refused to activate the voice service reject is 2 times. Then, according to the formula T next = T base ×(1 + γ×N reject / (1 + S trust )) to calculate the dynamic access period T next = 7×(1 + 0.5×2÷(1 + 3)) = 7×(1 + 0.25) = 8.75 days, which means that the system will try to interact with the customer again after 8.75 days (which can be rounded up to 9 days).

[0106] In the automatic interaction module, by parsing the call record text to obtain the customer's initial interests, the chatbot can target the conversation, increase the probability of the customer activating the voice service, precisely meet the customer's needs, and enhance the customer experience; and according to the voice service activation status, assign the activated customers to the artificial service queue to ensure that high-intent customers can obtain professional artificial services in a timely manner and optimize resource allocation; finally, analyze the dynamic access period based on the historical conversation content, which can avoid over-disturbing the customer, reduce customer annoyance, and improve the overall marketing effect.

[0107] Traditional marketing relies on static tags or simple behavioral data and it is difficult to deeply understand the customer's true needs. Therefore, in order to accurately grasp the customer's needs, optimize marketing resource allocation, and enhance the customer experience, in the present invention, through the portrait construction module, starting from the customer's historical conversation content, a comprehensive and accurate customer portrait is constructed, and the customer's intention level is analyzed, and based on this, the most suitable recommended plan is matched from the marketing plan database to achieve personalized marketing;

[0108] Specifically, the image construction module first collects the historical conversation corpora of several customers, uses the BIO tagging system to label the categories of the corpora, and obtains the labeled data set; then trains a BiLSTM-CRF model with this data set to obtain an interest extraction model that can accurately extract keywords from text data and label the interest categories to which each keyword belongs (such as price sensitivity, security requirements, service efficiency, and additional benefits, etc.);

[0109] Next, input the historical conversation content of the customer into the interest extraction model to obtain several keywords and their corresponding interest types; then, sort the sentences containing keywords in the historical conversation content in chronological order to determine the position of the sentence where each keyword first appears, and obtain the sentence position ranking, which is used to measure the order and importance of the customer's interests;

[0110] After that, according to the formula I k =∑(TF-IDF(w i ,k)×e^(-β×rank(w i ))) / N u calculate the interest intensity I k of the customer for the k-th type; where, TF-IDF(w i ,k) reflects the importance of the keyword in a specific interest type, rank(w i ) reflects the sentence position ranking where the keyword first appears, and N u is the total number of conversation turns;

[0111] Sort the calculated interest intensity I k in the order of price sensitivity, security requirements, service efficiency, and additional benefits, and finally generate the customer image P client =[I1, I2, I3, I4]. At the same time, by counting the time from the first conversation to the customer's consent to activate the voice service, obtain the activation time T 激活 , and then calculate the customer intention level according to a specific formula.

[0112] After having the customer image and intention level, then extract the marketing image P plan of each marketing plan and the lowest intention level L plan from the marketing plan database. According to the hybrid similarity formula calculate the matching degree of the customer with each marketing plan and screen out the marketing plan with the highest matching degree as the recommended plan. Finally, use the chatbot to generate targeted recommended words based on the customer's basic information, customer image, intention level, and recommended plan to achieve precision marketing; where, L intent represents the customer intention level, and j represents the marketing plan index number.

[0113] In one implementation, the process of obtaining the customer intention level may include:

[0114] C2-1, statistically calculate the time from the first conversation to the consent to activate the voice service, and obtain the activation time T 激活 ;

[0115] C2-2, calculate the customer intention level L according to the formula ; where intent ; among them, respectively represent the reciprocals of the activation time T 激活 and the total number of conversation rounds N u , k represents the regression coefficient, which is obtained by fitting historical data: analyze a number of historical data using a logistic regression model to determine the maximum likelihood estimate value as the final determined value; λ represents the interaction weight coefficient, which is obtained based on historical experience.

[0116] The portrait construction module also includes obtaining recommended conversation scripts using a chatbot to facilitate the staff

[0117] Specifically, when a customer enters the manual follow-up queue, the system first extracts personal information from the insurance database according to the customer ID, including basic information (such as age, occupation), insurance history (policy type, term), and claim records (number of times, amount);

[0118] Then, input the personal information, customer portrait, intention level, and recommended plan into the chatbot, and generate conversation scripts through the second instruction statement; the instruction template clearly requires combining customer characteristics with product advantages to output colloquial and compliant recommended conversation scripts, and ensure compliance through sensitive word filtering. For example, the second instruction statement can be: Please analyze the reasons why [recommended plan] is suitable for the customer based on the [personal information], [customer portrait], and [intention level] of the customer, and give a user-friendly recommended conversation script to promote successful signing;

[0119] Next, in the manual follow-up module, the generated conversation scripts and recommended plans will be automatically pushed to the workbench interface of the manual agent, sorted according to the customer intention level and associated with key data (such as browsing records). After the customer service dials the phone, the system will push the customer portrait and conversation script key points in real time. For example, the interface shows "Ms. Li, has browsed critical illness insurance 10 times, recommend female lifelong critical illness insurance". During the conversation, if the customer asks specific questions, the system will automatically retrieve the knowledge base and supplement the conversation script details. If the customer refuses, it will be marked as "requiring secondary follow-up" and the intention level will be updated; if the intention is reached, it will directly enter the signing process.

[0120] An embodiment of the second aspect of the present invention provides an AI Agent marketing system based on a large model, including:

[0121] S1. Use an outbound robot to make automatic outbound calls and conduct conversations according to the list of potential customers, and obtain customer contact information based on the conversation content;

[0122] S2. Start a chatbot based on the customer contact information, and use the chatbot to guide the customer to activate the voice service; among them, the chatbot is constructed based on artificial intelligence algorithms and is used to parse text content and interact with customers interactively;

[0123] S3. Build a customer portrait based on the historical conversation content and analyze the customer's intention level. Match and recommend a plan from the marketing plan database according to the customer portrait and the customer's intention level, and obtain recommended conversation scripts using the chatbot;

[0124] S4. Push the recommended conversation scripts and recommended plans to the staff for manual follow-up.

[0125] Some of the data in the above formula is calculated by removing the dimension and taking its numerical value. The formula is obtained by software simulation of a large amount of collected data to get a formula closest to the actual situation; the preset parameters and preset thresholds in the formula are set by those skilled in the art according to the actual situation or obtained through simulation of a large amount of data.

[0126] The working principle of the present invention:

[0127] First, screen high-potential customers based on historical policy data and trigger automatic outbound calls. After obtaining contact information through voice interaction, use information entropy analysis and response time evaluation to dynamically screen high-value customers for secondary follow-up;

[0128] Subsequently, through multi-modal interaction guidance, the chatbot analyzes the customer's interests and dynamically adjusts the strategy, marks customers who activate the voice service with a high intention level, and generates a personalized access cycle for customers who do not activate through emotional trust analysis;

[0129] Then build a dynamic customer portrait, use a neural network model to extract demand features, combine the time decay algorithm to quantify the interest intensity, and convert the activation time and the number of conversation rounds into a precise intention level through a logistic regression model; then use a hybrid similarity model to screen and recommend plans in real time;

[0130] Finally, achieve efficient conversion through human-machine collaboration, automatically generate compliant conversation scripts embedded with customer historical data, the manual agent follows up with the help of a real-time recommendation dashboard, and the system automatically updates the customer status according to the feedback and triggers a secondary reminder.

[0131] The above embodiments are only used to illustrate the technical method of the present invention and not to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical method of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical method of the present invention.

Claims

1. An AI Agent marketing system based on a large model, characterized in that, Including: Intelligent outbound call module: used to automatically make outbound calls and have conversations according to the list of potential customers by using an outbound call robot, and obtain the real contact information of customers based on the conversation content; wherein, the outbound call robot is built based on artificial intelligence algorithms, used for automatically making outbound calls and having interactive voice interactions with customers, and the customer list includes the virtual contact information of customers. Automatic interaction module: used to start a chatbot according to the real contact information of customers, and use the chatbot to guide customers to activate voice services; wherein, the chatbot is built based on artificial intelligence algorithms, used for parsing text content and having interactive interactions with customers. Profile construction module: used to construct a customer profile based on historical conversation content and analyze the customer's intention level, match and recommend a solution from the marketing plan database according to the customer profile and customer intention level, and obtain recommended conversation scripts by using the chatbot. Manual follow-up module: used to push the recommended conversation scripts and recommended solutions to the staff for manual follow-up.

2. The AI Agent marketing system based on a large model according to claim 1, characterized in that, The automatic outbound call made by using the outbound call robot according to the list of potential customers includes: A1. Collect historical policy data in the insurance database, screen out users who have not renewed their policies within a preset time period, have a click-through rate greater than a preset threshold, and have no complaint records from the historical policy data to obtain a list of potential customers. A2. Obtain the difference T between the current time and the policy expiration time based on the policy data in the list of potential customers 到期 , when T 到期 is less than the preset expiration threshold, initiate an outbound call based on the customer's virtual contact information; A3, record the customer ID and the outbound call time t 外呼 , the connection status, and the call transcription text to obtain an outbound call record form; A4. Judge whether the call transcription text contains the real contact information of the customer; if yes, save the outbound call record form; if not, judge whether to make a second outbound call according to the call transcription text.

3. The AI Agent marketing system based on a large model according to claim 2, characterized in that, The judgment of whether to make a second outbound call according to the call transcription text includes: A41. Use natural language processing technology to perform word segmentation on the call transcription text to obtain a number of call words. A42. Calculate the cosine similarity between a number of call words and the words in the insurance vocabulary list, mark the call words with a cosine similarity greater than a preset similarity threshold as keywords to obtain a number of keywords; wherein, the insurance vocabulary list is a custom common vocabulary list related to insurance business. A43, calculate the retry call score Retry_Score according to the formula ; where E dialog represents the information entropy of several call vocabulary, and is obtained according to the formula , P(c i ) represents the word frequency of the call vocabulary c i , N key represents the number of several keywords, and T re represents the time difference from when the outbound robot asks a question to when the customer responds for the first time; A44, save the customers whose retry score Retry_Score is greater than the preset score threshold to the retry call queue, and make a retry call after t 外呼 +T, and then jump to A3; where T represents the preset call cycle.

4. An AI Agent marketing system based on a large model according to claim 2, characterized in that, The use of the chatbot to guide customers to activate voice services includes: B1. Use the chatbot to parse the call transcription text through the first instruction statement to obtain the initial interest of the customer. B2. Use the chatbot to have an interactive conversation according to the initial interest of the customer, and automatically guide the customer to activate voice services according to the conversation content. B3. Judge whether the voice service is activated; if yes, save the customer to the manual service queue; if not, analyze and obtain a dynamic access period according to the historical conversation content.

5. The AI Agent marketing system based on a large model according to claim 4, characterized in that, The analysis and obtaining of the dynamic access period according to the historical conversation content includes: B31, using a predefined sentiment dictionary to match the number N of positive words that appear in the historical conversation content 正 and the number N of negative words 负 , and calculating the ratio of N 正 to N 负 to obtain the customer trust score S trust ; B32, collect the cumulative number N of times that customers refuse to activate the voice service reject , and calculate the dynamic access period T according to the formula next = T base ×(1 + γ×N reject / (1 + S trust )); where, T base represents the preset basic access period, and γ represents the period adjustment coefficient.

6. An AI Agent marketing system based on a large model according to claim 1, characterized in that, The construction of the customer profile according to the historical conversation content includes: C1-1. Collect the historical conversation corpus of a number of customers, and use the BIO tagging system to label the corpus to obtain a labeled dataset, and use the labeled dataset to train a BiLSTM-CRF model to obtain an interest extraction model. Wherein, the labeled categories include price sensitivity, safety needs, service efficiency, and additional benefits, and the input of the keyword extraction model is text data, and the output is a number of words and the labeled category of each word. C1-2. Input the historical conversation content into the keyword extraction model to obtain a number of keywords and interest types. C1-3. Sort the sentences containing keywords in the historical conversation content in chronological order to obtain the sentence position rankings. C1-4, according to formula I k = ∑(TF-IDF(w i , k) × e^(-β × rank(wi))) / N u Calculate the interest intensity I of the customer in the k-th type k ; where, TF-IDF(w i , k) represents the TF-IDF value of the keyword w i in the interest type k, rank(w i) represents the ranking of the position of the statement where the keyword w i first appears in the historical conversation content, and N u represents the total number of dialogue turns; C1-5, according to the interest intensity I k Sort in the order of price sensitivity, safety requirements, service efficiency, and additional benefits to obtain the customer portrait P client = [I1, I2, I3, I4]; where I1 represents the interest intensity of price sensitivity, I2 represents the interest intensity of safety requirements, I3 represents the interest intensity of service efficiency, and I4 represents the interest intensity of additional benefits.

7. The AI Agent marketing system based on a large model according to claim 6, characterized in that, The method for obtaining the customer intention level includes: C2-1, count the time from the first conversation to the consent to activate the voice service, and obtain the activation time T 激活 ; C2-2, calculate the customer intention level L according to the formula ; where intent ; among them respectively represent the reciprocal of the activation time T 激活 and the total number of dialogue turns N u , k represents the regression coefficient, and λ represents the interaction weight coefficient.

8. An AI Agent marketing system based on a large model according to claim 6, characterized in that, The matching and recommending of a solution from the marketing solution database based on the customer profile and the customer intention level includes: C3-1, extracting the marketing portraits P of several marketing plans from the marketing plan database plan and the lowest intention level L plan ; C3-2, according to the hybrid similarity formula Calculate the matching degree between the customer and the j-th marketing plan where cos() represents the cosine similarity calculation formula; C3-3, according to Filter out the marketing plan with the highest matching degree to obtain the recommended plan.

9. An AI Agent marketing system based on a large model according to claim 8, characterized in that, The obtaining of recommended conversation scripts using a chatbot includes: D1. Extract the personal information of the customer from the insurance database according to the customer ID; wherein, the personal information includes the customer's basic information, insurance purchase history, and claim settlement records. D2. Input the customer's personal information, customer profile, intention level, and recommended solution into the chatbot, and use the second instruction statement to guide the chatbot to generate recommended conversation scripts for the recommended solution.

10. A marketing method for an AI Agent based on a large model, which is applied to a marketing system for an AI Agent based on a large model according to any one of claims 1-9, characterized in that, It includes: S1. Use an outbound robot to automatically make outbound calls and have conversations according to the list of potential customers, and obtain the customer's true contact information based on the conversation content. S2. Start the chatbot according to the customer's true contact information, and use the chatbot to guide the customer to activate the voice service. S3. Construct a customer profile and analyze the customer intention level based on the historical conversation content, match a recommended solution from the marketing solution database according to the customer profile and the customer intention level, and obtain recommended conversation scripts using the chatbot. S4. Push the recommended conversation scripts and recommended solutions to the staff for manual follow-up.

Images