Intelligent mail system and generating method based on multi-modal user value analysis

Abstract

The application discloses an intelligent mail system and a generation method based on multi-modal user value analysis, and belongs to the field of artificial intelligence, aiming at solving the problems of serious homogeneity of existing mail marketing content, artificial response delay, low utilization rate of historical communication data, and insufficient personalized touch capacity, etc. The system includes user value scoring, mail generation engine, automatic triggering and sending control subsystem; through the preset five-dimensional feature system, multi-source heterogeneous data processing, user value quantitative scoring and deep intention recognition are completed, driving the large language model to generate personalized mail and complete multi-dimensional compliance audit and content quality closed-loop control, combined with full-scene event triggering and dynamic priority scheduling mechanism to realize the automatic and accurate sending of mail. The application improves the matching degree of mail content and user demand, customer behavior response timeliness, mail reply rate and business conversion rate, and realizes intelligent and automatic closed-loop management of the whole process of mail marketing.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and customer relationship management (CRM), specifically involving an intelligent email system and generation method based on multimodal user value analysis. Background Technology

[0002] The existing technology has the following drawbacks:

[0003] Emails suffer from severe homogenization: Traditional email templates cannot dynamically adjust content based on user behavior, resulting in poorly targeted responses. Traditional template emails are static and universal, unable to dynamically adjust content based on real-time user behavior (such as browsing specific product pages). This leads to a severe disconnect between the pushed information and the user's current interests and role, lacking relevance and targeting. Customers do not feel understood and are easily misled into viewing them as spam, resulting in low response and conversion rates.

[0004] Response Delay: Manually composing emails takes an average of 15-30 minutes, making it impossible to respond to key user behaviors in real time. The core pain point is missing the marketing "golden window." Research shows that contacting potential customers within the first few minutes after they become interested yields the highest conversion rates. However, manually composing emails is cumbersome, taking an average of 15-30 minutes, by which time the user's attention has already shifted or been captured by competitors. This delay not only wastes time but also directly leads to lost business opportunities, preventing the initial investment in acquiring potential customers from realizing its value.

[0005] Low information utilization: A large amount of valuable information in historical emails has not been systematically extracted and reused. Enterprise historical emails contain a wealth of high-value "dark data" (such as customer needs, pain points, and project stages), but this data is scattered in personal inboxes in unstructured text form, creating "data silos." The lack of automated extraction tools and reliance on manual review are extremely inefficient, resulting in each communication lacking historical context, severely damaging customer experience and sales efficiency.

[0006] Insufficient personalization: There is a lack of quantitative models to assess user value, and email tone and content structure are not sufficiently differentiated. Email strategies rely heavily on the salesperson's subjective experience, making it impossible to quantify and differentiate customer value. Communication resources for high-potential customers are similar to those for ordinary leads, resulting in a failure to provide a good experience for the most important customers.

[0007] Therefore, the industry urgently needs an intelligent email automation solution that can build a comprehensive quantitative evaluation system based on multimodal user behavior data, achieve accurate user value segmentation and deep intent recognition, fundamentally solve the industry pain points of traditional email marketing that rely on human experience, suffer from severe homogenization, and lack of personalized matching, and can also rely on AI generation capabilities to achieve automated generation of email content that closely matches user needs and closed-loop control of compliance and quality, fully explore the core value of reusing historical communication data, completely break the industry dilemma of customer data silos and low information utilization, and achieve intelligent triggering and dynamic priority scheduling based on real-time user behavior and lifecycle status, accurately grasp the golden window of marketing, and solve the core problems of delayed human response and lost business opportunities. Summary of the Invention

[0008] This invention addresses the problems in existing enterprise sales development and customer relationship management scenarios, such as severe content homogenization in email marketing, delayed manual responses leading to missed marketing opportunities, siloed historical communication data resulting in low utilization of high-value information, and insufficient personalized outreach capabilities due to the lack of a quantitative user value assessment system. It proposes an intelligent email system and generation method based on multimodal user value analysis. By constructing a multimodal user value scoring subsystem based on a five-dimensional feature system, an AI email generation engine with closed-loop content quality control capabilities, and a real-time event-driven automated triggering and sending control subsystem, a fully automated closed loop is formed, encompassing quantitative user value assessment, intelligent personalized email generation, and precise outreach scheduling. This solves the core technical problems of existing technologies, such as poor targeting of static template emails, low efficiency and delayed response of manual writing, insufficient insight into user needs and purchase intentions, and unreasonable allocation of sales follow-up resources. The system achieves significant improvements in the matching degree between email content and user needs, greatly shortens the response time for high-value customer behavior, fully explores and reuses the value of historical communication data, optimizes the efficiency of sales resource allocation, and ultimately achieves significant improvements in email response rate and sales opportunity conversion rate.

[0009] In a first aspect, the present invention provides an intelligent email system based on multimodal user value analysis, and the technical solution adopted to solve the above-mentioned technical problems is as follows:

[0010] An intelligent email system based on multimodal user value analysis, comprising:

[0011] The user value scoring subsystem processes multi-source heterogeneous user data, performs standardized preprocessing, and constructs a scoring context. Based on a preset five-dimensional feature system, weight configuration rules, and scoring model configuration, it generates a structured scoring instruction Prompt_M. After inference by a large language model, it outputs a quantitative user value score with multi-dimensional analysis and simultaneously generates and outputs structured user intent recognition results. The five-dimensional feature system includes basic attribute dimensions, user interaction dimensions, behavioral dynamic dimensions, demand urgency dimensions, and historical email dimensions.

[0012] The email generation engine receives user value scores and user intent recognition results from the user value scoring subsystem, combines them with historical communication records and preset business rules to generate email composition prompts; it then submits these prompts to a large language model for inference, generating personalized email content, and performs automated compliance review and quality verification on the personalized content, completing a closed-loop control of content quality; and...

[0013] The automated triggering and sending control subsystem communicates with the user value scoring subsystem and the email generation engine, respectively. The automated triggering and sending control subsystem is used for:

[0014] Trigger control: Based on preset trigger rules, it receives user behavior events and changes in user status. When the trigger conditions are met, it triggers the user value scoring subsystem to perform user value assessment and user intent recognition, and starts the email generation engine to execute the entire personalized email generation process.

[0015] Sending Schedule: Receives qualified email content from the email generation engine that has completed closed-loop content quality control, and manages the emails to be sent in real time or on a schedule based on preset scheduling rules and preset priority queues to complete the automated sending of emails.

[0016] Optionally, the five-dimensional feature system includes the basic attribute dimension, user interaction dimension, behavioral dynamic dimension, demand urgency dimension, and historical email dimension. Specifically: the basic attribute dimension stores the user's static attribute information, including company size, industry, job role, and region. This static attribute information is used to match a preset ideal customer profile; the user interaction dimension stores user interaction behavior data, including the number of occurrences, frequency, and session duration of multi-channel interaction behaviors such as page visits, page downloads, telephone communication, online consultation, and social media comments. This interaction behavior data is used to quantify the user's interaction activity and interaction depth; the behavioral dynamic dimension stores user browsing, The downloaded content is tagged with content tags, which are used to identify the user's interests, role attributes, and behavioral motivations. In terms of urgency, the system stores high-value user interaction data, which corresponds to preset strong purchase intention behaviors. These behaviors include repeatedly visiting product pages, repeatedly visiting pricing pages, downloading technical documents, adding items to the cart, and visiting contact information pages. This high-value interaction data is used to identify the urgency of the user's purchase intention. Finally, the system stores user feedback data on historical emails, including email opening behavior, reply behavior, reply keywords, and sentiment. Sentiment is determined as positive, neutral, or negative based on preset keyword rules.

[0017] Optionally, the user value scoring subsystem is configured to perform the following steps: Multi-source data aggregation and preprocessing: Acquire multi-source heterogeneous user data, including static user information, interaction behavior sequences, historical communication records, product and target profiles, and historical scoring data; complete data standardization and integration, deduplication, and invalid data filtering to construct a complete scoring context including the above data; Perform feature encoding and vectorization transformation on the structured data, behavior sequence data, and unstructured text data in the scoring context to generate input features that can be processed by the large language model; Dynamic scoring instruction construction: Based on the preset five-dimensional feature system, weight configuration rules, and scoring model configuration, automatically generate a structured scoring instruction Prompt_M in combination with the input features; The scoring instruction Prompt_M is a dedicated instruction used to constrain the large language model to perform user value scoring inference, which explicitly defines the scoring dimensions, weight rules, scoring processing logic, and output format constraints; Among them, the five-dimensional feature system is the basic framework of the scoring dimensions, and the weight configuration rules are for each dimension. The scoring model configuration includes scoring logic decomposition rules, large model inference constraint rules, and output format standardization rules. Weight configuration rules include preset basic weight allocation rules between dimensions, industry-adaptive dimension weight adjustment rules, event weight coefficient rules for segmented behaviors, and weighted calculation rules for deep interactive behaviors in the user interaction dimension, as well as sentiment tendency score adjustment rules for the historical email dimension. Large language model inference and scoring output: The scoring instruction Prompt_M is submitted to the large language model for inference, obtaining a preset structured user value score that conforms to the scoring model configuration constraints. The user value score includes a quantified total user value score, as well as the original score, weighted score, and objective fact-based score reasoning analysis for each dimension in the five-dimensional feature system. User intent recognition: Based on the scoring context, key user behaviors are identified and analyzed to generate structured user intent recognition results. Key user behaviors include searching keywords and accessing or downloading key content.

[0018] Optionally, the email generation engine is configured to perform the following steps: Aggregating multi-source input information: receiving user value scores, user intent recognition results, and historical communication records of the target user from the user value scoring subsystem; wherein, the user value score includes the total user value score, user tier level, and detailed scores for each dimension of the five-dimensional feature system; the user intent recognition results include the user's core intent type and key topics; Generating email communication strategies: based on the aggregated multi-source input information and combined with a preset strategy generation template, generating structured email communication strategy prompts; submitting the email communication strategy prompts to a large language model for inference to obtain a structured email communication strategy including communication actions, intent type matching, and core communication strategies; Generating final email composition prompts: integrating the structured email communication strategy, the aggregated multi-source input information, and preset email composition rules to generate final email composition prompts; the preset email composition rules are part of the preset business rules.

[0019] Optionally, the email generation engine is configured to perform the following steps: Full context parsing: Parse the generation instructions, user background information, email communication strategy, and historical communication records in the email composition prompts using a large language model to clarify the core objectives and constraints of email generation; Constrained content generation: Based on the parsing results, generate personalized email content that matches the user's core needs and aligns with the email communication strategy by adhering to preset content compliance requirements, format specifications, and communication style constraints using a large language model; Standardized structured output: Output structured email generation results according to the preset format requirements agreed upon in the email composition prompts using a large language model. The generated email results include at least the email title, email body, and content summary.

[0020] Optionally, the email generation engine is configured to perform the following automated compliance review and content quality closed-loop control steps: Multi-dimensional automated review: Based on preset content review rules, multi-dimensional verification is performed on the personalized email content generated by the large language model. Multi-dimensional verification includes at least key information matching verification, brand specification compliance verification, content compliance verification, and format specification verification; Review passed processing: If the personalized email content passes all verifications, it is determined to be qualified content and the email is pushed to the sending queue; At the same time, manual quality labeling of email content is supported, and the labeled data is used for optimization and iteration of the generation effect of the large language model; Closed-loop control of review failed: If the personalized email content fails the verification, the reasons for the failure are analyzed and located, and the following branch processing is executed: Automatic regeneration branch: Based on the reasons for the failure, optimization constraint instructions are generated, and personalized email content is regenerated in combination with the original email composition special prompt words. Automated review is performed again until the content passes the verification or the preset maximum number of retries is reached; Manual intervention branch: If the verification still fails after reaching the maximum number of retries, or if the content involves preset complex processing scenarios, the content is routed to the manual processing stage to receive qualified email content corrected by the human or the human-supplemented generation instructions to complete the quality control of the email content.

[0021] Optionally, the automated triggering and sending control subsystem is configured to execute the following triggering control steps: Rule reception: Based on preset multi-dimensional triggering rules, continuously receive user full-link behavior events and user status changes; multi-dimensional triggering rules include at least user behavior event triggering rules, scheduled task triggering rules, user status change triggering rules, and historical communication response triggering rules; Trigger determination: When the received event or status change matches the triggering condition of any triggering rule, complete the trigger validity verification and lock the target user to be followed up; Process initiation: Send a trigger command to the user value scoring subsystem, triggering the user value scoring subsystem to pull multi-source heterogeneous data of the target user, start the user value assessment and user intent recognition process, and simultaneously send a start command to the email generation engine to trigger the personalized email generation process.

[0022] Optionally, the automated triggering and sending control subsystem is configured to execute the following sending scheduling steps: Email reception and queuing: Receives qualified email content from the email generation engine that has undergone closed-loop content quality control, and stores the emails to be sent in a preset priority queue; Dynamic priority determination: Assigns a sending priority to each email to be sent based on preset priority rules; The determination dimensions of the priority rules include at least the user tier level corresponding to the user value score, the urgency of the triggering event, the timeliness requirements for email follow-up, and the response status of historical communication; Scheduling and sending control: Performs scheduling management on emails to be sent according to the assigned priority, supporting three sending modes: real-time sending, preset timed sending, and best-time delivery sending; For emails that fail to send, performs resending operations according to preset retry rules until successful sending or the maximum number of retries is reached; Closed-loop sending management: Records the sending status and result of each email, completing closed-loop control and data archiving of the entire email sending process.

[0023] Secondly, this invention provides an intelligent email generation method based on multimodal user value analysis, and the technical solution adopted to solve the above-mentioned technical problems is as follows:

[0024] A smart email generation method based on multimodal user value analysis, applied to the aforementioned email system, includes the following steps:

[0025] Trigger control: Through the automated trigger and sending control subsystem, user behavior events and user status changes are received based on preset trigger rules. When the trigger conditions are met, the entire process of user value assessment and email generation is triggered.

[0026] User value assessment and intent recognition: The user value scoring subsystem processes multi-source heterogeneous user data, completes standardized preprocessing and constructs a scoring context. Based on the preset five-dimensional feature system, weight configuration rules and scoring model configuration, it generates a structured scoring instruction Prompt_M. The scoring instruction Prompt_M is submitted to the large language model for inference, outputting a quantitative user value score with multi-dimensional analysis, and simultaneously generating and outputting structured user intent recognition results.

[0027] Email prompt word generation: Through the email generation engine, user value scores and user intent recognition results are received, and combined with historical communication records and preset business rules, special prompt words for email composition are generated.

[0028] Personalized email content generation: Submit email composition prompts to a large language model for inference to generate personalized email content;

[0029] Content quality closed-loop control: Automated compliance review and quality verification are performed on the generated personalized email content to complete the content quality closed-loop control and output qualified email content that has passed the review;

[0030] Email scheduling and automated sending: Through the automated triggering and sending control subsystem, qualified email content is received, and based on preset scheduling rules, emails to be sent are scheduled and managed in real time or on a timed basis through preset priority queues to complete the automated sending of emails.

[0031] Optionally, the five-dimensional feature system includes the basic attribute dimension, user interaction dimension, behavioral dynamic dimension, demand urgency dimension, and historical email dimension. Specifically: the basic attribute dimension corresponds to the user's static attribute information, including company size, industry, job role, and region, used to match a pre-defined ideal customer profile; the user interaction dimension corresponds to user interaction behavior data, including the number of occurrences, frequency, and session duration of multi-channel interaction behaviors such as page visits, page downloads, telephone communication, online consultation, and social media comments, used to quantify the user's interaction activity and interaction depth; the behavioral dynamic dimension corresponds to the user's browsing... The download content is tagged with relevant tags to identify user interests, role attributes, and behavioral motivations. The urgency dimension corresponds to high-value user interaction data, which in turn corresponds to pre-defined strong purchase intention behaviors. These behaviors include repeatedly visiting product pages, repeatedly visiting pricing pages, downloading technical documents, adding items to the cart, and visiting contact information pages, used to identify the urgency of a user's purchase intention. The historical email dimension corresponds to user feedback data on historical emails, including email opening behavior, reply behavior, reply keywords, and sentiment. Sentiment is determined as positive, neutral, or negative based on pre-defined keyword rules.

[0032] This invention discloses an intelligent email system and generation method based on multimodal user value analysis, which offers several advantages over existing technologies: First, it constructs a multimodal user value scoring subsystem based on a five-dimensional feature system to accurately quantify user value and identify deep-seated intentions, providing an objective basis for personalized outreach. Second, it designs an AI email generation engine with a quality closed loop to address the pain point of content homogenization, ensuring content compliance and relevance. Third, it establishes a real-time triggering and dynamic priority scheduling mechanism to seize the golden marketing window and optimize resource allocation. This invention achieves a fully automated closed loop for email marketing, significantly reducing costs and improving efficiency, thereby increasing lead conversion rates. Attached Figure Description

[0033] Appendix Figure 1 This is an architecture diagram of the intelligent email system based on multimodal user value analysis according to the present invention;

[0034] Appendix Figure 2 This is a flowchart illustrating the workflow for generating user value scores within the user value scoring subsystem.

[0035] Appendix Figure 3 This is a flowchart of the email generation engine's workflow;

[0036] Appendix Figure 4 This is a flowchart of the intelligent email generation method based on multimodal user value analysis of the present invention. Detailed Implementation

[0037] To make the technical solution, the technical problem solved, and the technical effect of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with specific embodiments.

[0038] Example 1

[0039] Combined with appendix Figure 1 This embodiment proposes an intelligent email system based on multimodal user value analysis, including a user value scoring subsystem, an email generation engine, and an automated triggering and sending control subsystem. This embodiment uses a sales development representative (SDR) customer follow-up scenario for a B2B enterprise SaaS product as an example to provide a detailed description of the intelligent email automation system of this invention. In this embodiment, the large language model that the system interfaces with is either a commercially available large model API that the enterprise can call or an internally deployed open-source large model. This embodiment does not limit the specific model type, as long as it can complete structured instruction reasoning and text generation tasks.

[0040] The intelligent email system based on multimodal user value analysis described in this embodiment includes a user value scoring subsystem, an email generation engine, and an automated triggering and sending control subsystem. The specific implementation methods of each module are described in detail below.

[0041] (I) User Value Scoring Subsystem

[0042] In this embodiment, the user value scoring subsystem is the core decision-making module of the entire intelligent email automation system. It interfaces with the enterprise CRM system, the official website user behavior tracking system, the enterprise email system, the online customer service system, and the user tag management platform, while also communicating bidirectionally with the automated triggering and sending control subsystem and the email generation engine. Its core functions are to integrate and process multi-source heterogeneous user data, quantify user value scoring, and accurately identify users' deep intentions, providing comprehensive and objective decision-making basis for subsequent email communication strategy generation and personalized content creation.

[0043] In this embodiment, all preset rules, weights, templates, and thresholds of the subsystem can be visually adjusted by enterprise managers through the system configuration backend. This allows for the adaptation of scoring standards to different industries, business lines, and customer groups, ensuring that the scoring logic is highly aligned with the enterprise's business objectives and ideal customer profile.

[0044] 1. Core Preset Configuration Description

[0045] In this embodiment, the specific implementation methods of the various preset rules, systems, and configurations mentioned are as follows. All configurations are stored in the system's scoring rule library and can be called, updated, and iterated as needed:

[0046] (1) Specific implementation rules of the five-dimensional feature system

[0047] The five-dimensional feature system is the core framework for user value scoring, including basic attribute dimensions, user interaction dimensions, behavioral dynamics dimensions, urgency of needs dimensions, and historical email dimensions. The specific implementation rules, data sources, and scoring logic for each dimension are as follows:

[0048] Basic attribute dimension: This dimension stores the user's static attribute information. The data comes from the enterprise CRM system, business registration information platform, and user registration information, specifically including the size of the user's company, industry, job role, and region. It is primarily used to match the enterprise's preset ideal customer profile (hereinafter referred to as ICP). The higher the match with the ICP, the higher the basic score for this dimension. In this embodiment, the enterprise's preset ICP is: enterprise service / SaaS industry, company size of 50 or more employees, director or higher decision-making position, and first- or second-tier city in China.

[0049] User interaction dimension: This dimension stores user interaction behavior data across all channels, sourced from the official website's tracking system, online customer service system, call center system, and social media management platform. Specifically, it includes the number of occurrences, frequency, and duration of each session for behaviors such as page visits, document downloads, phone calls, online inquiries, and social media comments. Its core purpose is to quantify user interaction activity and depth. In this embodiment, weighted calculation rules are set for high-frequency visits, long-duration sessions, and multiple logins—deep interaction behaviors. Within the same statistical period, the higher the frequency of user interaction and the longer the duration of each single session, the higher the score for this dimension.

[0050] Behavioral Dynamics Dimension: This dimension stores preset content tags corresponding to user browsing and downloading content. Data is sourced from the official website's content management system and user behavior tagging platform, specifically including topic tags for pages visited, category tags for downloaded files, and keyword tags for search content. Its core purpose is to identify user interests, roles, and core behavioral motivations. In this embodiment, preset content tags include: product features, pricing schemes, technical documentation, customer case studies, API integration, compliance explanations, etc. Tag clustering determines the user's core focus; the higher the match between tags and the company's product canvas, the higher the score for this dimension.

[0051] Urgency of Need Dimension: This dimension stores high-value user interaction data, sourced from the official website's tracking system and user behavior event platform. These high-value interactions correspond to pre-defined strong purchase intention behaviors, specifically including repeatedly visiting product detail pages, repeatedly visiting pricing pages, downloading technical documents, applying for trials, adding solutions to a comparison list, and repeatedly viewing the contact us page. It is primarily used to identify the urgency of a user's purchase intention and is a core weighted dimension for user value scoring. In this embodiment, differentiated weight coefficients are preset for different high-value behaviors; for example, downloading product technical documents has a weight coefficient of 2.0, and browsing product pages has a weight coefficient of 1.0. The higher the frequency and weight of high-value behaviors, the higher the score for this dimension.

[0052] Historical Email Dimension: This dimension stores complete user feedback data for historical emails. The data originates from the enterprise email system and email marketing platform, specifically including whether the email was opened, whether a link was clicked, whether a reply was made, the keywords in the reply content, and sentiment. Sentiment is categorized as positive, neutral, or negative based on the enterprise's preset keyword rules. In this embodiment, preset positive keywords include: "interested," "please quote," "arrange a demonstration," and "send detailed information." Negative keywords include: "not needed," "not considering for now," "don't send anymore," and "complain." Simultaneously, preset scoring adjustment rules are applied: positive user feedback corresponds to +5 points to the total score, no feedback results in no score adjustment, and negative feedback corresponds to -10 points to the total score.

[0053] (2) Specific implementation of weight allocation rules

[0054] The weighting configuration rules are the core constraints for scoring calculation. They are preset as an editable configuration file, which enterprises can flexibly adjust according to their own business needs. The basic configuration and adaptation rules in this embodiment are as follows:

[0055] Basic weighting rules between dimensions (default configuration): Basic attributes 20%, user interaction 20%, behavioral dynamics 15%, urgency of needs 35%, historical emails 10%;

[0056] Industry-specific dimensional weighting adjustment rules: In the SaaS industry scenario, the weight of urgency of demand is increased to 40%, correspondingly reducing the weight of behavioral dynamics to 10%; in the manufacturing industry scenario, the weight of basic attributes is increased to 30%, correspondingly reducing the weight of historical emails to 5%.

[0057] Event weighting coefficient rules for segmented behaviors: For deep interactive behaviors in the user interaction dimension and high-value behaviors in the demand urgency dimension, corresponding weighting coefficients are preset for quantitative scoring calculation of single behaviors.

[0058] Weighted calculation rules for deep interactive behaviors: For user interaction, deep interactive behaviors such as staying on a single page for more than 120 seconds, visiting more than 3 times a week, and actively initiating online consultations will be given an additional 1.5 times weight on the basic frequency statistics.

[0059] Sentiment Score Adjustment Rules: Based on user feedback from historical emails, a fixed score is set to positively add points or negatively deduct points from the user's total value score, thus achieving a closed-loop feedback of historical communication effectiveness.

[0060] (3) Specific definition of scoring model configuration

[0061] The scoring model is configured as a preset JSON / YAML format configuration file, stored in the system's model configuration library, and serves as the core basis for generating scoring instructions and constraining the inference logic of large models. In this embodiment, the scoring model configuration fixedly includes the following core content:

[0062] Basic model information includes the model's unique ID, model name, and model scoring objective (e.g., "Identify high-intent potential customers and prioritize sales follow-up resources").

[0063] Scoring Dimension Definition: Includes the unique ID, dimension name, weight percentage, scoring guidelines, and corresponding data source for each dimension in the five-dimensional feature system;

[0064] Inference constraint rules: Clarify the core requirements for scoring large models, such as scoring must be based on objective facts in context, subjective assumptions are not allowed, and reasons for scoring must be provided for each dimension;

[0065] Standardized output format rules: Mandate the JSON structure, required fields, and field formats for large model outputs to ensure that the output results can be directly parsed and used by the system.

[0066] (4) Data preprocessing rules

[0067] In this embodiment, the preset data preprocessing rules are as follows, which are used to ensure the validity and standardization of the input data:

[0068] Data deduplication rules: For the same type of event triggered by the same user within 5 minutes, only the earliest valid record is retained to avoid scoring bias caused by duplicate behavior;

[0069] Invalid data filtering rules: In page browsing events, if the user stays on the page for less than 3 seconds, it is judged as an invalid browsing event and filtered; empty sessions without clear content, behavioral data of test accounts, and data of blacklisted users are all directly filtered and not included in the scoring context;

[0070] Standardized format rules: User interaction behavior sequences are uniformly standardized into JSON format. Each event must contain three core fields: timestamp, event_type, and event_params. Historical communication records first undergo NER entity recognition and sentiment analysis, and then are converted into structured records and incorporated into the scoring context.

[0071] (5) User intent recognition preset rules

[0072] In this embodiment, the core classification, judgment criteria, and output rules for user intent recognition are pre-defined and fully compatible with the subsequent strategy generation stage of the email generation engine, as detailed below:

[0073] There are four pre-defined categories of core intents: procurement intent, technical intent, competitor intent, and general consulting intent. Each type of intent corresponds to a clear behavioral judgment criterion.

[0074] Purchasing intent: The determination is based on user behavior such as viewing the pricing page, applying for a trial, requesting a quote, and viewing the cooperation process;

[0075] Technical intent: The determination is based on user behaviors such as downloading technical documentation, viewing API documentation, consulting technical parameters, and reviewing integration solutions;

[0076] Competitor intent: The determination is based on user behavior such as searching for competitor names, visiting competitor comparison pages, and downloading competitor comparison reports;

[0077] General consultation intent: The determination is based on user behavior such as submitting online consultations, leaving messages, and inquiring about after-sales / compliance issues;

[0078] Output rules: Force output of standardized JSON format results, including four required fields: core intent, secondary intent, key terms, and intent confidence.

[0079] 2. Specific Implementation Steps of the User Value Scoring Subsystem

[0080] The user value scoring subsystem is executed strictly according to the following steps, the specific implementation methods of each step are as follows, and see the appendix for details. Figure 2This is a flowchart illustrating the workflow for generating user value scores within the user value scoring subsystem:

[0081] Step 1: Multi-source data aggregation and preprocessing

[0082] The system first acquires multi-source heterogeneous user data of the target user, including user static information, interaction behavior sequences, historical communication records, product and target profile information, and user historical rating data (if any). It then completes the standardized integration of all data, deduplication, and filtering of invalid data to construct the complete rating context C for this rating.

[0083] In this embodiment, the scoring context C is defined as: C={U,I,H,P,S_prev}, where:

[0084] U: User static information, that is, the static profile data of users and companies corresponding to the basic attribute dimensions in the five-dimensional feature system;

[0085] I: A sequence of user interaction records, a complete list of behavioral events such as page views, downloads, and inquiries, sorted by timestamp;

[0086] H: Historical communication records, i.e., the history of text conversations with the user, such as emails, chat logs, and phone minutes;

[0087] P: Product and target profile information, namely the product canvas keywords and the characteristics of the ideal customer profile ICP preset by the enterprise;

[0088] S_prev: User's historical rating data, i.e., the user's total score and scores in each dimension in the previous round. If it is the first rating, it will be empty.

[0089] After data aggregation is completed, the system performs feature encoding and vectorization transformation on different types of data in the scoring context to generate input features that can be processed by large language models. The specific transformation rules are as follows:

[0090] Structured data (user static information): directly perform tag encoding and numerical conversion to form standardized numerical features;

[0091] Behavioral sequence data (user interaction records): The user's behavioral history is organized in time series and converted into vector embeddings that represent user behavior patterns and interests through a pre-trained model.

[0092] Unstructured text data (historical communication records): Through keyword extraction, sentiment analysis, and entity recognition, core pain points, needs, and feedback information are extracted, converted into corresponding word vectors and numerical sentiment features, and incorporated into the input features.

[0093] Step 2: Constructing Dynamic Scoring Instructions

[0094] Based on a pre-defined five-dimensional feature system, weight configuration rules, and scoring model configuration, the system automatically generates structured scoring instructions (hereinafter referred to as Prompt_M) by combining pre-processed input features.

[0095] Prompt_M is a dedicated instruction used to constrain large language models in performing user value rating inference. The system automatically populates the core content of the rating model configuration and the aggregated rating context C into the predefined meta-prompt word template to generate a standardized Prompt_M, ensuring consistent standards and logic for each rating. Prompt_M explicitly defines the rating dimensions, weight rules, rating processing logic, and output format constraints, forcing large models to output parsable, structured results.

[0096] In this embodiment, the core structure of the predefined meta-prompt word template is as follows:

[0097] Instruction: Please act as a professional sales analysis AI, based on all the provided context information, and strictly follow the given scoring model dimensions to assess user value.

[0098] enter:

[0099] Scoring Model: [Model Name], Scoring Objective: [Model Scoring Objective]

[0100] Scoring Dimensions and Weights: [Iterate through the scoring model configuration to generate a complete list of "Dimension Name: Weight Percentage"]

[0101] Contextual Information: [Fill in the full set of core information for the standardized rating context C, including user static information, interaction behavior summary, historical communication record summary, product and target profile, and historical rating data]

[0102] Processing requirements:

[0103] Each rating dimension was analyzed one by one, and for each dimension, objective evidence supporting the rating was extracted from the context, strictly following the rating guidelines for the corresponding dimension.

[0104] Give an original score of 1-100 for each dimension, along with a brief reason for the score. The reason for the score must be based on objective facts in the context and must not be subjective or arbitrary.

[0105] Calculate the weighted score for each dimension: Weighted score = Original score × Weight percentage of the corresponding dimension;

[0106] If historical user ratings exist, trends can be referenced in the analysis, but the final rating must be determined independently based on the current context. Output requirements:

[0107] The output must be a strictly formatted JSON object, and no additional explanatory text may be output.

[0108] The JSON object must contain the total score (total_score) and details of the scores for each dimension. Each dimension detail must include the dimension name, the original score, the weighted score, and the reason for the score.

[0109] Step 3: Large Language Model Inference and Scoring Output

[0110] The system submits the constructed scoring instruction Prompt_M to the large language model for inference, and obtains user value scoring results in a preset JSON format that conforms to the scoring model configuration constraints.

[0111] In this embodiment, the user value score output by the large language model always includes the following core content:

[0112] Associated model information: the ID and name of the corresponding scoring model to ensure the traceability of scores;

[0113] Quantified total user value score: the sum of weighted scores across all dimensions, with a score range of 0-100.

[0114] Dimensional Score Details: The original score, weighted score, and analysis of the scoring reasons based on objective facts for each dimension in the five-dimensional feature system, ensuring that the scoring results are interpretable and traceable.

[0115] After obtaining the scoring results, the system will perform format and integrity checks on the output content. If it does not meet the preset format requirements, it will regenerate Prompt_M and submit it for inference until a standardized scoring result that meets the requirements is obtained.

[0116] Step 4: User Intent Recognition

[0117] Based on the constructed rating context, the system combines key user behaviors (such as those over a period of time, especially the recent 2-4 weeks), interaction content, and historical communication information to identify and analyze the user's core needs and generate structured user intent recognition results.

[0118] Key user behaviors include search keywords, accessed or downloaded content, tags for interactive events, and core questions asked. Based on pre-defined intent classification rules, the system uses behavioral tag clustering and semantic reasoning via a large language model to accurately identify user intent. It then outputs standardized JSON-formatted intent recognition results, which are simultaneously pushed to the email generation engine as the core basis for formulating email communication strategies.

[0119] 3. Complete scoring execution example of this embodiment

[0120] This example uses the virtual user "Zhang San" to fully demonstrate the entire execution process of the user value scoring subsystem, forming a complete closed loop with the subsequent implementation of the email generation engine and automated triggering and sending control subsystem:

[0121] Multi-source data aggregation to construct scoring context C

[0122] User static information U: Zhang San, technical director of a company with 50-200 employees, in the enterprise services / SaaS industry, located in Beijing, perfectly matching the company's preset ICP;

[0123] User interaction record sequence I: A total of 5 visits to the official website within 2 days, including 2 visits to the pricing page (with a maximum stay of 300 seconds per visit), 1 visit to the AI ​​CRM product page, 1 visit to the product feature details page, and 1 technical document download. The stay time on all pages exceeded 100 seconds.

[0124] Historical Communication Record H: Zhang San replied to the first email from the sales team, clearly mentioning that the current team uses Excel and simple project management tools, which has pain points such as scattered data and difficulty in following up with customers. He focused on the value of AI in sales forecasting and customer insights, asked about relevant customer cases, and his overall sentiment was positive.

[0125] Product and Target Profile Information P: The core keywords of the product canvas are AI CRM, sales forecasting, customer insights, and automated processes; the core characteristics of the ICP are technology decision-makers, enterprise service industry, companies with more than 50 employees, and the existence of sales management pain points;

[0126] Historical rating data S_prev: None, indicating the first rating.

[0127] The data preprocessing and feature transformation system completes data deduplication and invalid data filtering, confirming that all behaviors are valid events; it encodes structured user static information, converts behavior sequences into user interest vectors, extracts pain point keywords and sentiment features from historical email text, and completes the standardization processing of all features.

[0128] The system dynamically generates Prompt_M and submits it to the inference system. Based on the SaaS industry standard scoring model configuration and combined with Zhang San's scoring context, it automatically fills in the meta-prompt word template, generates a standardized Prompt_M, and submits it to the large language model for inference.

[0129] The large language model outputs structured scoring results that conform to the required format, specifically: total user value score of 83.5 points; including: basic attribute dimension raw score of 85 points and weighted score of 17 points; user interaction dimension raw score of 80 points and weighted score of 16 points; behavioral dynamic dimension raw score of 82 points and weighted score of 12.3 points; demand urgency dimension raw score of 90 points and weighted score of 31.5 points; and historical email dimension raw score of 75 points and weighted score of 6.7 points; and also outputs the corresponding scoring analysis for each dimension.

[0130] The user intent recognition and output system identifies intent based on Zhang San's behavior and communication context. The output results are: core intent "seeking technical solutions and customer cases", secondary intent "understanding product functions and procurement solutions", key keywords "AI sales forecasting, customer insights, data fragmentation pain points, customer cases", and intent confidence level of 95%.

[0131] (II) Email Generation Engine

[0132] In this embodiment, see Figure 3 The email generation engine communicates and interfaces with the user value scoring subsystem and the automated triggering and sending control subsystem. Its core function is to receive the user value score and user intent recognition results output by the user value scoring subsystem, combine them with the target user's historical communication records and the enterprise's preset business rules, complete the entire process of email communication strategy decision-making, automated generation of personalized email content, and content compliance and quality verification, and finally output qualified email content to the automated triggering and sending control subsystem.

[0133] In this embodiment, all preset rules, templates, and parameters of the email generation engine can be visually adjusted by enterprise managers through the system configuration backend to adapt to the communication needs of different business lines, industries, and customer groups.

[0134] 1. Basic Preset Configuration Instructions

[0135] In this embodiment, the specific implementation of the core configurations such as preset business rules, templates, and thresholds mentioned is as follows: all configurations are stored in the system's business rule base and can be called and updated as needed:

[0136] User tiering rules: The system pre-sets the mapping rules from the user's total value score to the customer level, which cannot be adjusted across levels. Specifically: 0-40 points correspond to level C (low-value customer), 41-70 points correspond to level B (medium-value customer), and 71-100 points correspond to level A (high-value / S-level customer). Different levels correspond to different communication strategies, tone styles and resource allocations.

[0137] Pre-defined business rules include: brand communication guidelines, product information release rules, advertising law compliance requirements, call-to-action (CTA) guidelines, industry-specific communication script library, benchmark customer case library, email signature and compliance information templates, and a list of prohibited words;

[0138] Strategy generation template: The system's predefined structured prompt word templates include six core modules: role setting, background information input box, scoring target, optional intent type library, optional communication action library, and mandatory output format constraints, ensuring that the generated communication strategies are standardized and parsable.

[0139] Email composition rules: This is a sub-module of business rules, which clarifies the core elements that emails must include, tone and style adaptation rules, format specifications, output format constraints, HTML code writing requirements, and rules for disabling placeholders, etc.

[0140] Content review rules: Clearly define the specific standards, judgment logic, and handling rules for non-compliant items in the multi-dimensional verification process;

[0141] Retry and fallback rules: The maximum number of retries for regenerating email content is preset to 3 (configurable), and the triggering conditions for manual intervention and the definition standards for complex processing scenarios are clearly defined.

[0142] 2. Specific execution steps of the email generation engine

[0143] The email generation engine strictly follows these steps in sequence, and the specific implementation methods for each step are as follows:

[0144] Step 1: Aggregation of multi-source input information

[0145] The system automatically retrieves and aggregates all the input information required for email generation, specifically including:

[0146] User value scoring data: Received from the user value scoring subsystem, including total user value score, mapped user tier level, original scores for each dimension of the five-dimensional feature system, weighted scores, and analysis of scoring reasons;

[0147] User intent recognition results: received from the user value scoring subsystem, including the user's core intent type, secondary intent type, key keywords, and intent confidence level;

[0148] Historical communication records of the target user: Full historical interaction data synchronized from the enterprise CRM system, enterprise email system, and online customer service system, specifically including full text of past emails with the user, online chat logs, transcribed text of telephone communication minutes, and historical follow-up notes, to ensure that the generated content has complete context awareness capabilities;

[0149] Pre-defined business rules: Based on the user's industry, customer segmentation, and follow-up scenario, automatically match the corresponding business rules, script library, case library, and compliance requirements.

[0150] Step 2: Generate Email Communication Strategy

[0151] Based on aggregated multi-source input information and combined with preset strategy generation templates, the system automatically generates structured email communication strategy prompts. These prompts are then submitted to a large language model for inference to obtain standardized and structured email communication strategies.

[0152] In this embodiment, the predefined optional intent types in the strategy generation template correspond perfectly to the intent classifications of the user value scoring subsystem, including purchasing intent, technical intent, competitor intent, and general consultation intent. Figure 4 Major categories, with subcategories under each major category;

[0153] Predefined optional communication actions include: sending customer case studies, inviting product demonstrations, providing product quotations, sending technical documents, answering customer questions, advancing project coordination, reactivating customers, and conducting in-depth needs research.

[0154] The system mandates that the large language model output structured results in JSON format, which includes three core fields: communication_actions (an array of communication actions with predefined action labels), intention_types (an array of matched intent types), and strategy (the core communication strategy text, which clarifies the email's core entry point, content focus, tone requirements, and pitfalls to avoid), ensuring that the output results can be directly parsed and used by the system.

[0155] Step 3: Generating the final email composition prompts

[0156] The system integrates the structured email communication strategy generated in step 2, the full input information gathered in step 1, and the preset email composition rules, and fills them into the predefined email generation meta prompt template to generate the final, standardized email composition prompt.

[0157] In this embodiment, the email composition prompts are fixed to include the following core modules: role setting, generation goal, full context information (user profile, rating results, intent, historical communication records), overall email communication strategy, mandatory requirements for email composition, and mandatory constraints on output format.

[0158] Among them, the strict requirements for email composition are to strictly follow the preset email composition rules, which clearly require that: a personalized salutation must be included; the email must be related to the business pain points mentioned by the customer; there must be a clear call to action; overly sales-oriented language is prohibited; the tone and style must be appropriate for the customer's segmentation; placeholders such as XXX and {} are prohibited; HTML styles must be embedded; and the email must be compatible with mainstream email clients.

[0159] The output format constraint requires the large language model to output results in standard JSON format, which must include four core fields: subject (email title), message_body (complete HTML code of the email body), message_body_text (plain text content of the email body), and summary (email content summary), with no additional redundant content.

[0160] Step 4: Generate email content using a large language model

[0161] The system submits the generated email composition prompts to the large language model for inference, obtaining the email generation result. The execution process of the large language model is as follows:

[0162] Full context analysis: The large language model fully analyzes the generation instructions, user background information, communication strategies, and historical communication context in the prompt words, clarifying the core objectives, constraints, and core user needs of email generation;

[0163] Constrained content generation: Based on the analysis results, it strictly follows the preset content compliance requirements, format specifications, and communication style constraints to generate personalized email content that matches the user's core needs and fits the established communication strategy;

[0164] Standardized and structured output: Output emails in JSON format that conforms to the format requirements of the prompt words, ensuring that the system can directly parse and call the content of each field.

[0165] Step 5: Automated compliance review and closed-loop management of content quality

[0166] The system performs multi-dimensional automated verification on the email content generated by the large language model based on preset content review rules, completing closed-loop control of content quality. The specific implementation method is as follows:

[0167] (1) Multi-dimensional automated review rules

[0168] The system performs validation across four core dimensions. If any dimension fails validation, the content is deemed unqualified. The specific validation rules are as follows:

[0169] Key information matching verification: Verify whether the email contains a personalized salutation, whether it matches the user's core pain points and intentions, whether it contains a clear CTA call to action, and whether it omits product information and compliance information that are required by business rules. If no core information is missing, the verification passes.

[0170] Brand compliance verification: Verify whether the email contains a standard corporate signature and compliance information, whether it uses absolute terms prohibited by the Advertising Law (such as "best", "lowest", "top", etc.), and whether it complies with the company's brand language guidelines. If there is no violation, the verification passes.

[0171] Content compliance verification: Verify whether the email contains sensitive words, prohibited language, or content that does not comply with industry regulatory requirements. If no prohibited content is found, the verification passes.

[0172] Format compliance check: Check the email's HTML code for syntax errors, whether the encoding is UTF-8, whether placeholders such as XXX and {} are present, and whether it meets email compatibility requirements. If there are no format issues, the check passes.

[0173] (2) Branch processing of audit results

[0174] Approval Processing: If the email content passes all dimensions of verification, the system determines it to be qualified and directly pushes the email to the pending-send queue of the automated triggering and sending control subsystem. At the same time, the system allows sales personnel to mark the quality of the final email content. The good and bad labels and modified content will be stored in the system's sample library for subsequent optimization strategy generation templates, email composition rules, prompt word engineering, and continuous iteration of the large language model's generation effect.

[0175] Closed-loop management of failed verification: If the email content fails verification, the system automatically locates and records the reason for the failure and the specific problem, and executes the following two branches of processing:

[0176] Automatic branch regeneration: The system generates supplementary optimization constraint instructions based on the reasons for the failure of the validation, merges them into the original email composition prompts, resubmits them to the large language model to generate email content, and performs a full automated review again after the generation is completed; this cycle continues until the content passes the review or the preset maximum number of retries is reached (3 times by default in this embodiment).

[0177] Manual intervention option: If the maximum number of retries is reached and the email still fails to pass review, or if the email content involves preset complex processing scenarios, the system will automatically route the task to the corresponding salesperson's manual processing queue. These preset complex processing scenarios include: email content involving large quotations exceeding 500,000 RMB, customized technical solution descriptions, compliance-sensitive statements, and content requiring cross-departmental collaborative confirmation. The manual processing stage allows salespeople to directly modify the email content to generate a qualified version, or to add more specific generation instructions to re-trigger AI generation, ultimately completing the quality control of the email content.

[0178] 3. Complete email generation and execution example in this embodiment

[0179] This example follows the virtual user "Zhang San" case from the aforementioned user value scoring subsystem, fully demonstrating the entire execution process of the email generation engine:

[0180] Multi-source information aggregation: The system receives Zhang San's user value score (total score 83.5 points, A-level high-value customer), five-dimensional score details, and user intent recognition results (core intent: seeking technical solutions and customer cases, key themes: AI sales prediction, customer insights, data fragmentation pain points). Simultaneously, it pulls the full text of Zhang San's past email exchanges with sales, the company's preset business rules for communicating with high-value customers in the SaaS industry, and email composition rules.

[0181] Email communication strategy generation: Based on the aggregated information, the system fills in the strategy generation template, generates email communication strategy prompts and submits them to the large language model, and finally obtains the structured strategy results: the communication action is "sending case studies and inviting product demonstrations", the intent type is "seeking technical solutions and seeking customer cases", and the core strategy is "taking the data-driven pain points mentioned by users as the starting point, focusing on matching the value of AI sales prediction function, building trust with real-world cases from companies of similar size, naturally leading to product demonstration invitations, using a professional and pragmatic tone, avoiding over-selling, and not proactively mentioning pricing;

[0182] Generate email composition prompts: The system integrates the above communication strategies, full user information, historical email content, and email composition rules to generate the final email generation prompts, explicitly requiring the generation of HTML format emails that conform to the specifications, and forcibly outputting the agreed JSON structure;

[0183] Large model generates email content: The system submits prompt words to the large language model and obtains email generation results that meet the format requirements, including email title, HTML body, plain text content, and content summary;

[0184] Automated review and closed-loop processing: The system performs multi-dimensional review on the generated emails. After verification, if the email contains a personalized salutation, matches the user's pain points and intent, has a clear demo invitation CTA, complies with brand guidelines, has no illegal content, has no HTML format errors, and has no placeholders, all verification items pass and the email is judged as qualified content and pushed to the queue to be sent.

[0185] (III) Automated Triggering and Sending Control Subsystem

[0186] In this embodiment, the automated triggering and sending control subsystem is bidirectionally connected to the user value scoring subsystem and the email generation engine, and also interfaces with the enterprise website's data tracking system, CRM system, user behavior data platform, and enterprise email sending gateway. Its core functions are: first, to receive user behavior and status in real time based on preset rules, accurately triggering the entire process of user value assessment and automated email generation; second, to dynamically prioritize qualified emails based on customer value and event urgency, achieving accurate, timely, and compliant email sending, while simultaneously completing closed-loop control and data archiving of the entire sending process.

[0187] In this embodiment, the event reception of this subsystem is implemented based on the RocketMQ real-time message processing platform. User behavior events are enqueued as soon as they are generated and consumed immediately after being enqueued. It supports horizontal scaling of the consumer end and can achieve millisecond-level event response in high-concurrency scenarios. All preset rules, thresholds and parameters can be adjusted by enterprise administrators through the system configuration backend to adapt to the triggering and sending requirements of different business scenarios.

[0188] 1. Core Preset Configuration Description

[0189] In this embodiment, the specific implementation methods of the various preset rules, thresholds, and parameters mentioned are as follows. All configurations are stored in the system's rule configuration library and can be updated and called as needed:

[0190] Multi-dimensional trigger rule library: The system has four preset core trigger rules, covering customer follow-up needs across all scenarios. Each rule category supports enterprise-defined addition, editing, and enabling / disabling. The specific definitions and implementation rules are as follows:

[0191] User behavior event triggering rules: Real-time event response rules are used to receive users' instantaneous high-value behaviors. The preset triggering condition is "the user completes a specified behavior". It supports two modes: single event triggering and multiple event combination triggering. The preset high-value triggering events include: users downloading product technical documents, repeatedly visiting product details page / pricing page, submitting trial application / form inquiry, consulting product-related issues online, watching product demonstration videos, etc.

[0192] Scheduled task triggering rules: Periodic proactive triggering rules are used for routine follow-up and nurturing of batch customers. Preset configuration items include: execution cycle (daily / weekly / monthly), target customer group screening conditions (such as "all customers who have been silent for more than 60 days", "customers who have not interacted in the past 7 days and are being nurtured"), and planned sending time window (such as 9:00-11:00 on weekdays).

[0193] User status change triggering rules: Lifecycle transition triggering rules, based on user value score and interaction behavior to determine user lifecycle status, and trigger follow-up process when user status changes; the system presets quantitative judgment thresholds for user lifecycle status: new lead (within 30 days after registration, total score <50), nurturing (total score 50-70, recent interaction but no conversion), high intent (total score ≥70, or high urgency behavior), dormant customer (no interaction behavior for 60 consecutive days).

[0194] Historical Communication Response Trigger Rules: Two-way communication follow-up trigger rules are used for automated follow-up after customer email replies and online inquiries. Preset trigger conditions include: no follow-up within 24 hours after the customer sends a positive reply, automatic sending of supplementary information after the customer inquires about a problem, and scheduled follow-up for customers who have not replied to historical emails.

[0195] Trigger validity verification rules: Preset anti-duplicate triggering and invalid triggering filtering rules, specifically: the same user and the same triggering rule will only trigger once within 24 hours to avoid repeatedly sending emails to harass customers; filter triggering events of test accounts, internal employee accounts, and blacklisted users, and do not start subsequent processes.

[0196] Priority rules: The preset priority calculation formula, weight of each dimension, and corresponding rules for sending timeliness are explained in detail in the sending scheduling steps.

[0197] Sending mode configuration rules: Clearly define the triggering conditions and execution logic for three modes: real-time sending, preset timed sending, and optimal arrival time sending.

[0198] Retry rules: The maximum number of retries for emails that fail to send is preset to 3, with retry intervals of 1 hour, 6 hours, and 24 hours respectively. If the email still fails after reaching the maximum number of retries, it will be marked as a sending error and pushed to the corresponding sales personnel for handling, and will no longer be automatically retried.

[0199] Manual approval rules: Preset manual approval trigger conditions before emails are sent to form a human-machine collaborative management system. Specifically, emails containing quotations with an estimated amount of ≥500,000, emails sent to ≥1,000 people in batches, emails with compliance statements / legal clauses in the subject line, or emails involving special industry regulatory requirements must be routed to the corresponding approver for review. Only after the review is approved can the emails enter the sending queue.

[0200] 2. Specific execution steps for triggering control

[0201] When the automated triggering and sending control subsystem performs triggering control, it strictly follows the steps in sequence, as detailed below:

[0202] Step 1: Rule Reception

[0203] Based on preset multi-dimensional triggering rules, the system continuously receives user events across the entire lifecycle through a real-time message queue and scans user lifecycle status and historical communication data through daily scheduled tasks to achieve uninterrupted reception of triggering conditions.

[0204] Regarding user behavior event triggering rules: By connecting to the enterprise's official website tracking system and user behavior data platform, we receive user behavior event streams in real time, and each event is matched with preset triggering conditions in real time;

[0205] For scheduled task triggering rules, user status change triggering rules, and historical communication response triggering rules: the system performs a full scan of user data every day at midnight, updates user lifecycle status, matches the conditions of the corresponding triggering rules, and marks users who meet the conditions and need to be followed up.

[0206] Step 2: Trigger Decision

[0207] When the received event or the scanned user status matches the trigger condition of any enabled trigger rule, the system immediately performs a trigger validity check:

[0208] Verify whether the user is on the blacklist or is an invalid test account. If the user is invalid, terminate the process directly and do not trigger subsequent operations.

[0209] Verify whether the user and the rule are within the cooldown period for preventing repeated triggering. If the rule has already been triggered and is within the cooldown period, terminate the process to avoid repeated harassment.

[0210] Once all verifications pass, the event is deemed a valid trigger. The target user to be followed up is identified, and the trigger rule ID, trigger event, and trigger time are recorded to form a trigger task.

[0211] Step 3: Full Process Startup

[0212] The system initiates the entire automated process by performing the following actions in parallel based on the triggered task:

[0213] Send a trigger command to the user value scoring subsystem. The command carries the unique identifier of the target user and the trigger event information, which triggers the user value scoring subsystem to pull the user’s multi-source heterogeneous data and immediately start the user assessment and intent recognition process.

[0214] Receive the output of the user value scoring subsystem. After the scoring and intent recognition are completed, send a start command to the email generation engine at the same time, carrying the user score, intent recognition result and trigger scenario information, to trigger the email generation engine to execute the entire personalized email generation process.

[0215] The entire process is tracked, and if any step fails, the error information is recorded and sent to the relevant person in charge for handling.

[0216] 3. Specific execution steps of the sending schedule

[0217] When the automated triggering and transmission control subsystem executes the transmission schedule, it strictly follows the steps in sequence, as detailed below:

[0218] Step 1: Receiving Emails and Joining the Queue

[0219] The system receives qualified email content from the email generation engine that has completed closed-loop content quality control. At the same time, it retrieves the target user information, trigger scenario information, and user value score data corresponding to the email. First, it checks whether the email requires manual approval. If it triggers the preset manual approval rules, it is first routed to the approval queue. After approval, it is stored in the priority queue to be sent. If no approval is required, the email to be sent is directly stored in the preset priority queue, recording the basic information of the email, target user, generation time, and required sending time.

[0220] Step 2: Dynamic Priority Determination

[0221] Based on preset priority rules, the system calculates a priority score for each email to be sent in the queue, assigns a sending priority, and clarifies the sending timeliness level.

[0222] In this embodiment, the priority score is calculated using the following formula: Priority Score = (Customer Value / 100) × Event Weight × Time Decay Factor. The specific definitions and implementation rules for each parameter are as follows:

[0223] Customer value: Calculated by multiplying the total user value score by the urgency coefficient. The urgency coefficient is 1.0 by default and can be configured up to 2.0 to amplify the priority of users with high urgency.

[0224] Event weight: The weight is preset from 1 to 5 points, reflecting the importance of the triggering event. The preset values ​​are: requesting a product demonstration 5 points, visiting the pricing page 4 points, downloading technical documents 3 points, form submission / online consultation 2 points, ordinary page browsing 1 point, and scheduled nurturing tasks 1 point.

[0225] Time decay factor: The calculation formula is 1 / (1 + number of hours after the event / 24). The more recent the event, the larger the decay factor and the higher the priority, ensuring that the latest user behavior is responded to first.

[0226] Based on priority scores, the system presets sending timeliness levels and response rules:

[0227] Priority score ≥ 4.0: highest priority, must be sent within 5 minutes;

[0228] Priority score 2.0-4.0: High priority, requires transmission to be completed within 1 hour;

[0229] Priority score < 2.0: Normal priority, execute batch planned sending, complete the sending within the preset off-peak period or specified time window;

[0230] Historical communication response status rules: This corresponds to the target user's historical email feedback types, response timeliness, and interaction status for historical inquiries. The preset adaptation coefficient rules are as follows: 1.5 for users who have responded positively to emails / engaged in inquiries within the past 30 days; 1.0 for users who have not responded to any emails / engaged inquiries within the past 30 days; and 0.5 for users who have responded negatively / explicitly refused within the past 30 days. This adaptation coefficient is incorporated into the priority score calculation formula, and the final calculation formula is updated to: Priority Score = (Customer Value / 100) × Event Weight × Time Decay Factor × Historical Communication Response Adaptation Coefficient.

[0231] Step 3: Scheduling and Sending Control

[0232] The system schedules and manages emails to be sent according to their assigned priority and delivery timeliness. High-priority tasks can be prioritized for delivery, ensuring that high-value and high-urgency emails reach customers first. The system supports three sending modes, the specific implementation methods of which are as follows:

[0233] Real-time sending: For the highest priority and high priority emails, the system immediately calls the enterprise email sending gateway to execute the email sending operation, ensuring that the emails are delivered within the time limit.

[0234] Preset scheduled sending: For emails triggered by scheduled tasks or emails with user-specified sending times, the system adds the emails to the corresponding time sending plan according to the preset sending time, and automatically executes sending after the specified time.

[0235] Optimal delivery time: For nurturing emails with normal priority, the system intelligently matches the time period with the highest open rate based on the user's and similar customer groups' historical email open and reply rates, and automatically schedules the emails to be sent during that time period to improve email delivery effectiveness.

[0236] Meanwhile, the system implements retry control for sending failures: if the email sending gateway returns a sending failure, the system automatically records the reason for the failure, and re-executes the sending operation after a specified interval according to the preset sending retry rules; until the sending is successful, or the maximum number of retries is reached, the retries are terminated, the exception is marked, and the corresponding person in charge is notified.

[0237] Step 4: Send closed-loop management

[0238] The system records the sending status of each email in real time (pending, sending, sent successfully, sent failed, under approval, rejected), sending time, reason for failure, message ID returned by the email gateway, and other end-to-end data. After the email is sent, the sending result and the user's subsequent email opening, clicking, and reply behavior data are synchronously archived to the enterprise CRM system and user behavior database, and simultaneously fed back to the user value scoring subsystem for subsequent iterative optimization of user scoring, forming a closed-loop data process.

[0239] 4. Complete triggering and sending execution example of this embodiment

[0240] This example follows the previous case of the virtual user "Zhang San" and fully demonstrates the entire execution process of the automated triggering and sending control subsystem:

[0241] Rule reception and trigger determination: Zhang San completes the download of technical documents on the official website. This action event is synchronized to the system in real time and matches the enabled "user downloads technical documents" action event trigger rule. The system performs a trigger validity check. Zhang San is a valid customer and the rule has not been triggered on him in the past 24 hours. The check passes and is determined to be a valid trigger. The target user Zhang San is locked and a trigger task is generated.

[0242] Full-process startup: The system sends a trigger command to the user value scoring subsystem to trigger the user value assessment and intent recognition of Zhang San, obtaining Zhang San's total user value score of 83.5 points, A-level high-value customer, and core intent of seeking technical solutions and customer cases; simultaneously, a startup command is sent to the email generation engine to trigger the entire email generation process, and finally obtains the qualified email content that has passed the review.

[0243] Email queuing and priority determination: This email does not require manual approval and will directly enter the priority queue; the system calculates the priority score: Zhang San's customer value is 83.5 × 1.6 (default urgency coefficient) = 133.6, the event weight is 3 (downloading technical documents), the event occurred 10 minutes ago, the time decay factor is close to 1.0, the final priority score is ≈4.0, which belongs to the highest priority and requires to be sent within 5 minutes.

[0244] Scheduling and Closed-Loop Management: The system immediately executes real-time sending, calls the email gateway to complete the email sending, and returns a successful sending status; the system records the entire sending data of the email, synchronously archives it to the CRM system, and continuously tracks the opening, clicking, and reply data of the email, feeding it back to the user value scoring subsystem for subsequent scoring optimization.

[0245] The intelligent email system based on multimodal user value analysis demonstrated in this embodiment achieves intelligent marketing across the entire sales email marketing chain through the organic collaboration and process-oriented connection of three core modules: user value scoring subsystem, email generation engine, and automated triggering and sending control subsystem. This chain enables the system to perceive multi-source heterogeneous user data, quantitatively evaluate multi-dimensional user value, accurately identify deep user intent, make personalized communication strategy decisions, intelligently generate compliant email content, and trigger scheduling and automated precise delivery across all scenarios.

[0246] The core innovation of this system lies in constructing a three-in-one collaborative architecture of "quantified decision boundary + intelligent generation closed loop + real-time scheduling and execution," which achieves a deep integration of rule determinism and AI flexibility.

[0247] With a user value scoring subsystem based on a five-dimensional feature system, a quantitative decision boundary for customer segmentation and communication strategies is constructed to ensure the accuracy of email marketing objectives and the interpretability of strategies. This defines a clear range of needs for AI content generation and avoids the subjective bias of traditional experience-driven approaches and the risk of deviation from content generated purely by AI.

[0248] Driven by a large language model, the email generation engine completes communication strategy reasoning, personalized content generation, and multi-dimensional compliance quality verification within the boundaries of quantitative decision-making, forming a closed loop of "strategy-generation-review-optimization". This completely breaks through the homogeneity limitations of traditional static templates and achieves highly customized content that matches user needs.

[0249] With a real-time event-driven automated triggering and sending control subsystem, a full-scenario triggering rule system and dynamic priority scheduling mechanism are built. This not only ensures minute-level instant response to high-value business opportunities and firmly grasps the golden window of marketing, but also realizes full-link controllability and traceability of the sending process, seamlessly adapting to the enterprise's existing sales management process.

[0250] Through the above architecture, the system not only overcomes the core defects of traditional email marketing, such as serious content homogenization and lack of personalization due to reliance on static templates, high response delay and lost business opportunities due to reliance on manual writing, and low utilization of historical communication data and lack of contextual support for customer follow-up, but also solves the industry pain points of pure AI-generated content being insufficiently targeted, having uncontrollable compliance, and being difficult to integrate into the enterprise's standardized sales process and customer management system.

[0251] Ultimately, the system generates personalized email content that fits the user profile, matches the core pain points, and complies with brand guidelines and compliance requirements. Through the automated triggering and sending control subsystem, it seamlessly integrates with the enterprise's existing CRM system, email sending gateway, and sales follow-up system. While retaining the ability for manual approval, content intervention, and quality labeling, it achieves a closed-loop, traceable management of the entire process from user behavior insight, value assessment, content generation to precise targeting and iterative feedback.

[0252] Therefore, this system not only significantly improves the automation and intelligence of email marketing in B2B sales scenarios, but also provides comprehensive value in reducing sales labor costs, ensuring marketing compliance, improving email response rates and lead conversion rates, optimizing marketing resource allocation efficiency, and standardizing sales follow-up processes. It provides a feasible technical path and practical paradigm for enterprises to build an intelligent email marketing system.

[0253] Example 2

[0254] Combined with appendix Figure 4 This embodiment proposes an intelligent email generation method based on multimodal user value analysis, applied to the email system described in Embodiment 1. This embodiment uses a sales development representative (SDR) customer follow-up scenario for a B2B enterprise SaaS product as an example to provide a detailed explanation of each execution step of the method. The large language model that this method interfaces with can be a commercially available large model API that the enterprise can call or an internally deployed open-source large model. This embodiment does not limit the specific model type, as long as it can complete structured instruction reasoning and text generation tasks.

[0255] In this embodiment, all preset rules, weights, thresholds, and templates involved in the method can be visually adjusted by enterprise managers through the system configuration backend, which can adapt to the business needs of different industries, different business lines, and different customer groups. All preset configurations are stored in the corresponding rule library and configuration library of the system, and can be called, updated, and iterated as needed.

[0256] This method specifically includes the following steps:

[0257] Step 1: Trigger Control

[0258] The automated triggering and sending control subsystem receives user behavior events and changes in user status based on preset triggering rules. When the triggering conditions are met, it triggers the entire process of user value assessment and email generation. The specific implementation method for this step is as follows:

[0259] Preset Trigger Rule Explanation: This step uses multi-dimensional trigger rules, with four core rule categories pre-defined to cover customer follow-up needs across all scenarios. Enterprises can customize, add, edit, and enable / disable individual rules, specifically including:

[0260] User behavior event triggering rules: Real-time event response rules, which receive users' instantaneous high-value behaviors, with the preset trigger condition being "user completes a specified behavior", supporting both single event triggering and multi-event combination triggering modes; preset high-value triggering events include: users downloading product technical documents, repeatedly visiting product details pages / pricing pages, submitting trial applications / form inquiries, consulting product-related issues online, watching product demonstration videos, etc.

[0261] Scheduled task triggering rules: Periodic proactive triggering rules are used for routine follow-up and cultivation of batch customers. Preset configuration items include execution cycle (daily / weekly / monthly), target customer group screening conditions, and planned sending time window;

[0262] User status change triggering rules: Lifecycle transition triggering rules, which comprehensively determine the user's lifecycle status based on user value score and interaction behavior, and trigger a follow-up process when the user's status changes;

[0263] Historical communication response trigger rules: Two-way communication follow-up trigger rules are used for automated follow-up after customer email replies and online inquiries. Preset trigger conditions include no follow-up within 24 hours after a customer sends a positive reply, automatic sending of supplementary information after a customer inquires about a problem, and scheduled follow-up after a customer fails to reply to historical emails.

[0264] Rule reception and trigger determination: The system continuously receives user behavior events across the entire lifecycle through a real-time message queue and scans user lifecycle status and historical communication status through daily scheduled tasks; when the received event and the scanned user status match the triggering conditions of any enabled triggering rule, the system immediately performs a trigger validity check.

[0265] Trigger validity verification rules: Verification content includes: 1. Whether the user is on the blacklist, whether it is an invalid test account / internal employee account, invalid users will terminate the process directly; 2. Whether the user and the triggering rule are within the preset anti-duplicate triggering cooldown period (default 24 hours), if it has been triggered within the cooldown period, the process will terminate directly to avoid repeatedly sending emails to harass customers.

[0266] Process Initiation: After all verifications pass, the process is deemed a valid trigger. The target user to be followed up is identified, the trigger rule ID, trigger event, and trigger time are recorded, a trigger task is generated, and the entire process of subsequent user value assessment and email generation is initiated.

[0267] Step 2: User Value Assessment and Intent Identification

[0268] The user value scoring subsystem processes multi-source heterogeneous user data, performs standardized preprocessing, and constructs a scoring context. Based on a pre-defined five-dimensional feature system, weight configuration rules, and scoring model configuration, it generates a structured scoring instruction (hereinafter referred to as Prompt_M). The scoring instruction Prompt_M is then submitted to a large language model for inference, outputting a quantified user value score with multi-dimensional analysis, and simultaneously generating and outputting structured user intent recognition results. Details are as follows:

[0269] Core Preset System Description

[0270] The five-dimensional feature system is the core framework for scoring user value. It includes basic attribute dimensions, user interaction dimensions, behavioral dynamics dimensions, demand urgency dimensions, and historical email dimensions. The definitions, data sources, and scoring logic of each dimension are completely consistent with those in Example 1.

[0271] Weighting configuration rules: These include preset basic weighting rules between dimensions, industry-adapted dimension weighting adjustment rules, event weighting coefficient rules for segmented behaviors, weighted calculation rules for deep interactive behaviors, and sentiment tendency scoring adjustment rules. The specific configuration is completely consistent with Example 1.

[0272] Scoring model configuration: This is a preset JSON / YAML format configuration file, which contains basic model information, scoring dimension definitions, inference constraint rules, and output format standardization rules. It is the core basis for generating Prompt_M, and the specific definition is completely consistent with Example 1.

[0273] Multi-source data aggregation and preprocessing: The system acquires multi-source heterogeneous user data of the target user, including user static information, interaction behavior sequences, historical communication records, product and target profile information, and user historical rating data (if any); it completes data standardization integration, deduplication, and invalid data filtering to construct the complete rating context C for this rating; simultaneously, it performs feature encoding and vectorization transformation on the structured data, behavior sequence data, and unstructured text data in the rating context to generate input features that can be processed by large language models. The specific rules for data deduplication, invalid data filtering, standardization, and feature transformation are completely consistent with those in Example 1.

[0274] Dynamic scoring instruction construction: Based on a pre-defined five-dimensional feature system, weight configuration rules, and scoring model configuration, the system automatically generates structured scoring instructions (Prompt_M) using predefined meta-prompt word templates, combined with pre-processed input features. Prompt_M explicitly defines the scoring dimensions, weight rules, scoring processing logic, and output format constraints. It is used to constrain the large language model to perform standardized user value scoring reasoning, ensuring that the output results are parsable and traceable.

[0275] Large Language Model Inference and Scoring Output: The system submits the constructed Prompt_M to the large language model for inference, obtaining a user value score in a preset JSON format that conforms to the scoring model configuration constraints. The score includes a quantified total user value score (0-100 points), as well as the original score, weighted score, and objective fact-based reasoning analysis for each dimension in the five-dimensional feature system. The system will perform format and completeness checks on the output results; if they do not meet the requirements, the system will regenerate and resubmit the inference until a compliant result is obtained.

[0276] User Intent Recognition: Based on the constructed rating context, the system combines key user behaviors (e.g., within a certain period, especially the recent 2-4 weeks), interaction content, and historical communication information. Through behavioral tag clustering and semantic reasoning using a large language model, it identifies and analyzes the user's core needs, generating standardized JSON-formatted user intent recognition results, including core intent, secondary intent, key keywords, and intent confidence. The preset core user intent classification and judgment criteria are completely consistent with those in Example 1.

[0277] Step 3: Email Notification Message Generation

[0278] The email generation engine receives the user value score and user intent recognition results, combines them with historical communication records and preset business rules, and generates email composition prompts. The specific implementation of this step is as follows:

[0279] Multi-source input information aggregation: The system automatically pulls and aggregates all the input information required for email generation, including: 1. User value scoring data (total score, user tier level, and detailed five-dimensional scoring); 2. User intent recognition results; 3. Complete historical communication records of the target user (past emails, chat logs, follow-up minutes, etc.); 4. Enterprise-preset business rules, script libraries, case libraries, and email composition rules. The mapping rules for user tier levels are as follows: 0-40 points correspond to C-level (low-value customers), 41-70 points correspond to B-level (medium-value customers), and 71-100 points correspond to A-level (high-value / S-level customers). Different levels correspond to different communication strategies and tone requirements.

[0280] Email communication strategy generation: Based on aggregated multi-source input information and a pre-defined strategy generation template, the system automatically generates structured email communication strategy prompts. These prompts are then submitted to a large language model for inference, yielding a standardized JSON-formatted email communication strategy. The result consistently includes three core fields: communication action, matched intent type, and core communication strategy, ensuring that the output can be directly parsed and called by the system. The structure of the strategy generation template, the predefined optional intent types, and the optional communication actions are completely consistent with Example 1.

[0281] The final email composition prompt generation: The system integrates the generated structured email communication strategy, all input information, and preset email composition rules, and fills them into a predefined email generation prompt template to generate the final, standardized email composition prompt. The email composition rules clearly define the core elements that an email must include, tone and style requirements, format specifications, prohibited content, and output format constraints, with specific definitions completely consistent with Example 1. The final generated prompt includes six core modules: role setting, generation goal, full context information, communication strategy, writing requirements, and mandatory output format constraints.

[0282] Step 4: Generate personalized email content

[0283] The email composition prompts are submitted to a large language model for inference to generate personalized email content.

[0284] The system submits the generated email composition prompts to the large language model, which then performs reasoning and generation according to the following logic:

[0285] Full context analysis: Completely analyzes the generation instructions, user background information, communication strategies, and historical communication context in the prompt words to clarify the core objectives, constraints, and core user needs of the email generation;

[0286] Constrained content generation: Based on the analysis results, it strictly follows the preset content compliance requirements, format specifications, and communication style constraints to generate personalized email content that matches the user's core needs and fits the established communication strategy;

[0287] Standardized structured output: Following the format requirements agreed upon by the prompt words, output emails in standard JSON format. The output must include four required fields: email title, complete HTML code for the email body, plain text content of the email body, and email content summary. There is no extra redundant content.

[0288] After the system obtains the output of the large language model, it first performs basic format validation to confirm that the output content conforms to the preset JSON structure and that the required fields are complete before proceeding to the subsequent content quality control stage.

[0289] Step 5: Closed-loop management of content quality

[0290] The generated personalized email content undergoes automated compliance review and quality verification to complete closed-loop content quality control, outputting approved and qualified email content. The specific implementation method for this step is as follows:

[0291] Multi-dimensional automated review: Based on preset content review rules, the system performs checks on the generated email content in four core dimensions. If any dimension fails the check, the content is deemed unqualified. The specific review rules are exactly the same as in Example 1.

[0292] Key information matching verification: Verify whether it contains personalized names, whether it matches the user's core pain points and intentions, whether it contains clear CTA calls to action, and whether any required product and compliance information is missing;

[0293] Brand compliance verification: Verify whether it contains standard corporate signature and compliance information, whether it uses absolute terms prohibited by advertising law, and whether it complies with corporate brand language standards;

[0294] Content compliance verification: Verify whether the content contains sensitive words, prohibited language, or content that does not comply with industry regulatory requirements;

[0295] Formatting compliance check: Check the HTML code for syntax errors, whether the encoding is UTF-8, whether placeholders such as XXX and {} are present, and whether it meets the compatibility requirements of mainstream email clients.

[0296] Approval Processing: If the email content passes all dimensions of verification, the system determines it to be qualified and directly pushes it to the pending-send queue of the automated triggering and sending control subsystem; at the same time, it supports sales personnel to mark the quality of the final email content, and the marked data is stored in the system sample library for subsequent optimization of prompt word templates and iteration of large model generation effects.

[0297] Closed-loop management of failed verification: If the email content fails verification, the system automatically locates and records the reason for the failure and the specific problem, and executes the following two branches of processing:

[0298] Automatic branch regeneration: The system generates supplementary optimization constraint instructions based on the reasons for the failure of the validation, merges them into the original email composition prompts, resubmits them to the large language model to generate email content, and performs a full automated review again after the generation is completed; this cycle continues until the content passes the review or the preset maximum number of retries is reached (3 times by default in this embodiment).

[0299] Manual intervention branch: If the maximum number of retries is reached and the email still fails to pass the review, or if the email content involves preset complex processing scenarios, the system will automatically route the task to the corresponding salesperson's manual processing queue. The manual step supports directly modifying the email content to generate a qualified version, or adding clear generation instructions to re-trigger AI generation, ultimately completing the quality control of the email content. Among them, the preset complex processing scenarios include: email content involving large quotations of over 500,000, customized technical solution descriptions, compliance-sensitive statements, and content requiring cross-departmental collaborative confirmation, the specific definitions of which are completely consistent with Example 1.

[0300] Step 6: Email Scheduling and Automated Sending

[0301] The automated triggering and sending control subsystem receives qualified email content and, based on preset scheduling rules, manages the emails to be sent in real-time or on a timed basis through a preset priority queue, thus completing the automated sending of emails. The specific implementation method of this step is as follows:

[0302] Email Reception and Queuing: The system receives approved emails and simultaneously retrieves corresponding target user information, trigger scenario information, and user value score data. First, it verifies whether the email triggers preset manual approval rules. If approval is required, it is routed to the approval queue and then stored in the pending-send priority queue after approval. If no approval is required, the email is directly stored in the preset priority queue, recording basic email information, target user, generation time, and required delivery time. The preset manual approval rules are completely consistent with those in Example 1.

[0303] Dynamic Priority Determination: Based on preset priority rules, the system calculates a priority score for each email to be sent in the queue, assigns a sending priority, and assigns a corresponding sending timeliness level. In this embodiment, the priority score calculation formula is: Priority Score = (Customer Value / 100) × Event Weight × Time Decay Factor. The definitions, calculation rules, preset event weights, and sending timeliness level classifications of each parameter are completely consistent with Embodiment 1. In addition, for the historical communication response dimension rules: corresponding to the target user's historical email feedback type, historical consultation response timeliness, and interaction, the preset adaptation coefficient rules are: if the user has positive email replies / consultation interactions in the past 30 days, the adaptation coefficient is 1.5; if the user has no email / interaction responses in the past 30 days, the adaptation coefficient is 1.0; if the user has negative replies / explicit rejections in the past 30 days, the adaptation coefficient is 0.5. This adaptation coefficient is incorporated into the priority score calculation formula, and the final calculation formula is updated to: Priority Score = (Customer Value / 100) × Event Weight × Time Decay Factor × Historical Communication Response Adaptation Coefficient.

[0304] Sending Schedule Control: The system manages the scheduling of emails to be sent according to their assigned priority and delivery timeliness level. High-priority tasks can be prioritized for delivery, ensuring that high-value and high-urgency emails reach customers first. The system supports three sending modes, with specific implementation rules as follows:

[0305] Real-time sending: For the highest priority and high priority emails, the enterprise email sending gateway is immediately invoked to execute the sending operation, ensuring that the emails are delivered within the time limit.

[0306] Preset scheduled sending: For emails triggered by scheduled tasks or emails with user-specified sending times, add them to the corresponding sending plan according to the preset time, and automatically execute sending after the specified time;

[0307] Optimal delivery time for sending: For nurturing emails with normal priority, the system intelligently matches and schedules sending emails during the time period with the highest open rate based on historical email open and reply rates for the user and similar customer groups, improving email delivery effectiveness. Simultaneously, the system implements retry control for sending failures: If the email sending gateway returns a sending failure, the system automatically records the reason for the failure and resends according to preset retry rules (default maximum of 3 retries, with retry intervals of 1 hour, 6 hours, and 24 hours respectively); retrying continues until successful delivery, or until the maximum number of retries is reached, at which point the system terminates the retry process, marks the error, and notifies the relevant responsible person.

[0308] Sending closed-loop management: The system records the sending status, sending time, failure reason, gateway return message ID, and other full-link data for each email in real time; after the email is sent, the sending result and the user's subsequent email opening, clicking, and reply behavior data are synchronously archived to the enterprise CRM system and user behavior database, and simultaneously fed back to the user value scoring subsystem for subsequent iterative optimization of user scoring, forming a closed-loop data process.

[0309] Complete method execution example of this embodiment

[0310] This example uses the virtual user "Zhang San" to fully demonstrate the entire execution process of this method, forming a closed loop with the system example in Implementation Example 1:

[0311] Trigger Control: When Zhang San downloads technical documents from the official website, this action is synchronized to the system in real time and matches the "user downloads technical documents" behavior event trigger rule. The system performs a trigger validity check, confirms that Zhang San is a valid customer and has not triggered the same rule within 24 hours. If the check passes, a trigger task is generated, and the entire process is started.

[0312] User value assessment and intent recognition: The system pulls multi-source heterogeneous data of Zhang San, constructs a scoring context, and completes data preprocessing and feature transformation; based on the SaaS industry scoring model configuration, it generates Prompt_M, submits it to the large language model for inference, and finally outputs a total user value score of 83.5 points, as well as five-dimensional scoring details and reason analysis; at the same time, it completes user intent recognition, outputting the core intent as "seeking technical solutions and customer cases", and the key keywords as "AI sales forecasting, customer insights, data fragmentation pain points, and customer cases".

[0313] Email prompt generation: The email generation engine receives scoring and intent recognition results, and aggregates Zhang San's historical communication records, the company's preset business rules and writing requirements; it first generates email communication strategy prompts, submits them to the big model to obtain structured communication strategies (communication actions: sending case studies, inviting product demonstrations; core strategy: taking the data-scattered pain points mentioned by the user as the starting point, focusing on matching the value of AI sales prediction function, building trust with case studies of companies of similar size, leading to product demonstration invitations, and using a professional and pragmatic tone); then it integrates all the information to generate the final email writing prompts.

[0314] Personalized email content generation: The system submits email composition prompts to a large language model to obtain emails that meet the format requirements, including email title, HTML body, plain text content, and content summary.

[0315] Content quality closed-loop control: The system performs multi-dimensional automated review of generated emails to confirm that the emails contain personalized salutations, match user pain points and intentions, have clear CTAs, comply with brand guidelines, contain no illegal content, have correct HTML formatting, and contain no placeholders. Once all checks pass, the emails are deemed qualified and pushed to the queue to be sent.

[0316] Email scheduling and automated sending: The system calculates the email's priority score as 4.0, which is the highest priority, requiring it to be sent within 5 minutes; it immediately calls the email gateway to execute real-time sending and finally returns a successful sending status; the system records the entire sending process data, synchronously archives it to the CRM system, and continuously tracks email open, click, and reply data, feeding it back to the user value scoring subsystem for iterative optimization.

[0317] Through the sequential execution of the above steps, this method can complete the entire process—from user behavior perception, multi-dimensional value assessment, deep intent identification, communication strategy decision-making, personalized compliance email generation, to precise outreach—within 5 minutes when dealing with high-value, real-time behavior-triggered scenarios such as "customers downloading technical documents" or "repeatedly visiting pricing pages." For scheduled, batch scenarios such as welcoming new leads, activating dormant customers, and nurturing regular customers, it achieves standardized, fully automated execution of the entire process through preset rules, enabling large-scale, precise customer outreach. This method integrates the objective decision-making boundaries of a five-dimensional quantitative scoring system, the personalized content creation capabilities of a large language model, and the response efficiency of a real-time trigger scheduling mechanism. While fully integrating into the enterprise's existing customer management system, sales follow-up processes, brand compliance, and manual approval requirements, it achieves automation, intelligence, closed-loop management, and traceability of the entire intelligent email marketing process. This reduces the manual email writing costs for sales representatives, improves the matching degree between email content and customer needs, email response rates, and lead conversion rates, optimizes the allocation efficiency of marketing and sales follow-up resources, and maximizes the return on marketing investment.

[0318] In summary, the core innovation of the intelligent email system and generation method based on multimodal user value analysis of this invention lies in constructing a three-in-one collaborative architecture of "quantified decision-making boundary + intelligent generation closed loop + real-time scheduling and execution." This achieves deep integration and boundary control of the objective decision-making capability of the multi-dimensional quantitative scoring system, the personalized content creation capability of the large language model, and the real-time response and process control capability of the automated rule engine. This solution establishes the core division of labor principle of "quantified data determines the communication direction, AI model determines the content expression, and rule engine determines the timing of outreach."

[0319] The system efficiently integrates multi-dimensional customer data, including static user attributes, behavioral time-series data, historical communication texts, and product profile information, through the aggregation and standardized preprocessing of multi-source heterogeneous data. Leveraging a pre-defined five-dimensional feature system and configurable weight rules, it achieves interpretable and quantifiable assessment of user value and accurate identification of deep-seated intentions. High-value real-time behavioral scenarios trigger minute-level end-to-end responses, while batch customer nurturing scenarios achieve large-scale automated follow-up through standardized rules. Based on this, the email generation engine first formulates precise communication strategies based on the quantitative decision results, then drives a large language model to generate personalized email content highly tailored to user pain points and needs within pre-defined business rules and compliance boundaries. Simultaneously, through multi-dimensional automated review, multiple rounds of iterative optimization, and a manual oversight mechanism, a complete closed-loop control of content quality is formed.

[0320] Based on the generated compliant and qualified email content, the automated triggering and sending control subsystem uses a dynamic priority scheduling mechanism to match the optimal sending time and reach strategy for emails of different value levels and urgency. It seamlessly integrates with the enterprise's existing CRM system, email gateway, and sales follow-up processes. While retaining the capabilities of manual approval, content intervention, and quality labeling, it achieves closed-loop management of the entire process from user behavior insight and content generation to precise reach and iterative feedback. This solution effectively addresses core industry pain points such as the severe homogenization of traditional email marketing templates, delayed manual responses leading to missed business opportunities, low utilization of historical communication data, and insufficient customer value segmentation and personalized reach capabilities. It also compensates for the shortcomings of purely AI-generated content, including insufficient targeting, uncontrollable compliance, and difficulty in integrating into standardized enterprise sales processes.

[0321] Its comprehensive technical effects are significant: it greatly expands the automated coverage of sales email marketing scenarios and improves the accuracy of matching email content with user needs; through an interpretable quantitative scoring system and a fully traceable control mechanism, it ensures the objectivity of marketing decisions and the compliance of content generation; it achieves seamless integration with the enterprise's existing customer management system and sales follow-up process; while significantly reducing the cost of manual email writing by sales representatives and improving customer follow-up efficiency, it significantly improves email open rates, response rates, and sales lead conversion rates; through dynamic priority scheduling, it achieves optimal allocation of marketing and sales resources, maximizing the return on marketing investment, and provides a solid and feasible technical path and practical paradigm for B2B enterprises to build a precise, efficient, controllable, and implementable intelligent email marketing and customer follow-up system.

[0322] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the disclosed embodiments of the present invention is limited to these examples; within the framework of the embodiments of the present invention, the technical features of the above embodiments or different embodiments can also be combined, and there are many other variations of different aspects of the embodiments of the present invention as described above, which are not provided in detail for the sake of brevity. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of the present invention should be included within the protection scope of the embodiments of the present invention.

Claims

1. An intelligent email system based on multimodal user value analysis, characterized in that, include: The user value scoring subsystem is used to process multi-source heterogeneous user data, complete standardized preprocessing and construct scoring context. Based on the preset five-dimensional feature system, weight configuration rules and scoring model configuration, it generates a structured scoring instruction Prompt_M. After inference by a large language model, it outputs a quantitative user value score with multi-dimensional analysis, and simultaneously generates and outputs structured user intent recognition results. The email generation engine receives the user value score and user intent recognition results output by the user value scoring subsystem, and combines them with historical communication records and preset business rules to generate special prompts for composing emails. The email composition prompts are submitted to a large language model for inference to generate personalized email content. The personalized email content is then subjected to automated compliance review and quality verification to complete the closed-loop management of content quality. as well as An automated triggering and sending control subsystem is communicatively connected to the user value scoring subsystem and the email generation engine, respectively. The automated triggering and sending control subsystem is used for triggering control and sending scheduling. The email generation engine is configured to perform the following steps: Aggregates multi-source input information: Receives user value scores, user intent recognition results, and historical communication records of the target user output by the user value scoring subsystem; wherein, the user value score includes the total user value score, user stratification level, and detailed scores for each dimension of the five-dimensional feature system; the user intent recognition results include the user's core intent type and key topics; Generate email communication strategy: Based on the aggregated multi-source input information and combined with the preset strategy generation template, generate structured email communication strategy prompts; submit the email communication strategy prompts to the big language model to perform inference and obtain a structured email communication strategy including communication actions, intent type matching, and core communication strategies. Generate final email composition prompts: Integrate the structured email communication strategy, the aggregated multi-source input information, and the preset email composition rules to generate final email composition prompts; the preset email composition rules are part of the preset business rules.

2. The system according to claim 1, characterized in that, The five-dimensional feature system includes basic attribute dimension, user interaction dimension, behavioral dynamic dimension, demand urgency dimension, and historical email dimension, among which: The basic attribute dimension stores the user's static attribute information, which includes company size, industry, job role, and region. The static attribute information is used to match a preset ideal customer profile. The user interaction dimension stores user interaction behavior data, which includes the number of occurrences, frequency of occurrence, and session duration of multi-channel interaction behaviors such as page visits, page downloads, telephone communication, online consultation, and social media comments. The interaction behavior data is used to quantify the user's interaction activity and interaction depth. The behavioral dynamic dimension stores content tags corresponding to the content browsed and downloaded by the user. These content tags are used to identify the user's interests, role attributes, and behavioral motivations. The urgency dimension of demand stores high-value interactive behavior data of users. The high-value interactive behavior data corresponds to preset strong purchase intention behaviors, which include repeatedly visiting product pages, repeatedly visiting pricing pages, downloading technical documents, adding items to the shopping cart, and visiting contact information pages. The high-value interactive behavior data is used to identify the urgency of users' purchase intention. The historical email dimension stores user feedback data on historical emails. The feedback data includes email opening behavior, reply behavior, reply content keywords, and sentiment tendency. The sentiment tendency is determined as positive, neutral, or negative based on preset keyword rules.

3. The system according to claim 2, characterized in that, The user value scoring subsystem is configured to perform the following steps: Multi-source data aggregation and preprocessing: Acquire multi-source heterogeneous user data, including static user information, interaction behavior sequences, historical communication records, product and target profiles, and historical rating data; complete data standardization and integration, deduplication and invalid data filtering; construct a complete rating context including the above data; perform feature encoding and vectorization transformation on the structured data, behavior sequence data, and unstructured text data in the rating context to generate input features that can be processed by large language models; Dynamic scoring instruction construction: Based on a preset five-dimensional feature system, weight configuration rules, and scoring model configuration, a structured scoring instruction Prompt_M is automatically generated by combining the input features. The scoring instruction Prompt_M is a dedicated instruction used to constrain the large language model to perform user value scoring inference, clearly defining the scoring dimensions, weight rules, scoring processing logic, and output format constraints. Among them, the five-dimensional feature system is the basic framework of the scoring dimensions, the weight configuration rules are the scoring weight constraints for each dimension and subdivided behavior, and the scoring model configuration includes scoring logic decomposition rules, large model inference constraint rules, and output format standardization rules. The weight configuration rules include preset basic weight allocation rules between dimensions, industry-adaptive dimension weight adjustment rules, event weight coefficient rules for subdivided behaviors, and weighted calculation rules for deep interactive behaviors for user interaction dimensions, and sentiment tendency scoring adjustment rules for historical email dimensions. Large Language Model Inference and Scoring Output: The scoring instruction Prompt_M is submitted to the large language model for inference to obtain a user value score in a preset structured format that conforms to the configuration constraints of the scoring model; the user value score includes a quantified total user value score, as well as the original score, weighted score, and score reason analysis based on objective facts for each dimension in the five-dimensional feature system; User intent recognition: Based on the rating context, key user behaviors are identified and analyzed to generate structured user intent recognition results; wherein, the key user behaviors include searching keywords, accessing or downloading key content.

4. The system according to claim 1, characterized in that, The email generation engine is configured to perform the following steps: Full context analysis: By analyzing the generation instructions, user background information, email communication strategies, and historical communication records in the email composition prompts using a large language model, the core objectives and constraints of email generation are clarified. Constrained content generation: Based on the parsing results of the large language model, personalized email content that matches the user's core needs and fits the email communication strategy is generated in accordance with preset content compliance requirements, format specifications, and communication style constraints. Standardized structured output: The large language model outputs a structured email result according to the preset format requirements agreed upon in the email composition prompts. The generated email result includes at least the email title, email body, and content summary.

5. The system according to claim 4, characterized in that, The email generation engine is configured to perform the following automated compliance review and content quality closed-loop control steps: Multi-dimensional automated review: Based on preset content review rules, multi-dimensional verification is performed on personalized email content generated by the large language model. The multi-dimensional verification includes at least key information matching verification, brand compliance verification, content compliance verification, and format standardization verification. Approval Processing: If the personalized email content passes all checks and is deemed qualified, the email will be pushed to the sending queue; manual quality annotation of the email content is also supported, and the annotated data is used to optimize and iterate the generation effect of the large language model. Closed-loop management of failed verification: If the personalized email content fails verification, analyze and locate the reason for the failure, and execute the following branch processing: Automatic branch regeneration: Generate optimization constraint instructions based on the reasons for the failure of the validation, combine the original email composition prompts to regenerate personalized email content, and perform automated review again until the content passes the validation or the preset maximum number of retries is reached. Manual intervention branch: If the maximum number of retries is reached and the verification still fails, or if the content involves a pre-set complex processing scenario, the content will be routed to the manual processing stage. The manual correction of qualified email content or the manual supplementation of generation instructions will be received to complete the quality control of the email content.

6. The system according to claim 1, characterized in that, The automated triggering and sending control subsystem is configured to perform the following triggering control steps: Rule reception: Based on preset multi-dimensional triggering rules, continuously receive user behavior events and user status changes throughout the entire process; the multi-dimensional triggering rules include at least user behavior event triggering rules, scheduled task triggering rules, user status change triggering rules, and historical communication response triggering rules; Trigger determination: When the received event or state change matches the trigger condition of any trigger rule, the trigger validity is verified and the target user to be followed up is identified; Process Initiation: A trigger command is sent to the user value scoring subsystem to trigger the user value scoring subsystem to pull multi-source heterogeneous data of the target user, start the user value assessment and user intent recognition process, and simultaneously send a start command to the email generation engine to trigger the personalized email generation process.

7. The system according to claim 1, characterized in that, The automated triggering and transmission control subsystem is configured to perform the following transmission scheduling steps: Email Receiving and Queuing: Receive qualified email content that has completed closed-loop content quality control from the email generation engine, and store the emails to be sent into a preset priority queue. Dynamic priority determination: Based on preset priority rules, a sending priority is assigned to each email to be sent; The priority rules are determined based on at least the user tier level corresponding to the user value score, the urgency of the triggering event, the timeliness requirement for email follow-up, and the response status of historical communications. Sending scheduling and control: According to the assigned priority, the system performs scheduling management on emails to be sent, supporting three sending modes: real-time sending, preset timed sending, and best arrival time sending; for emails that fail to be sent, the system performs resending operations according to preset retry rules until the emails are successfully sent or the maximum number of retries is reached. Sending closed-loop management: Record the sending status and result of each email, and complete the closed-loop control and data archiving of the entire email sending process.

8. A method for generating intelligent emails based on multimodal user value analysis, applied to the system described in any one of claims 1-7, characterized in that, Includes the following steps: Trigger control: Through the automated trigger and sending control subsystem, user behavior events and user status changes are received based on preset trigger rules. When the trigger conditions are met, the entire process of user value assessment and email generation is triggered. User value assessment and intent recognition: The user value scoring subsystem processes multi-source heterogeneous user data, completes standardized preprocessing and constructs a scoring context, and generates a structured scoring instruction Prompt_M based on a preset five-dimensional feature system, weight configuration rules and scoring model configuration. The scoring instruction Prompt_M is submitted to the large language model for inference, outputting a quantitative user value score with multi-dimensional analysis, and simultaneously generating and outputting structured user intent recognition results; Email prompt word generation: Through the email generation engine, the user value score and the user intent recognition result are received, and combined with historical communication records and preset business rules, special prompt words for email composition are generated. Personalized email content generation: Submit the email composition prompts to the large language model for inference to generate personalized email content; Content quality closed-loop control: Automated compliance review and quality verification are performed on the generated personalized email content to complete the content quality closed-loop control and output qualified email content that has passed the review; Email scheduling and automated sending: The automated triggering and sending control subsystem receives qualified email content and, based on preset scheduling rules, performs real-time or timed scheduling management of emails to be sent through a preset priority queue to complete the automated sending of emails.

9. The method according to claim 8, characterized in that, The five-dimensional feature system includes basic attribute dimension, user interaction dimension, behavioral dynamic dimension, demand urgency dimension, and historical email dimension, specifically: The basic attribute dimensions correspond to the user's static attribute information, which includes company size, industry, job role, and region, and are used to match a preset ideal customer profile. The user interaction dimension corresponds to the user's interactive behavior data, which includes the number of occurrences, frequency of occurrence, and session duration of multi-channel interactive behaviors such as page visits, page downloads, telephone communication, online consultation, and social media comments. This data is used to quantify the user's interactive activity and interaction depth. The behavioral dynamic dimension corresponds to the content tags associated with the content that the user browses and downloads, and is used to identify the user's interests, role attributes, and behavioral motivations. The urgency dimension of demand corresponds to the user's high-value interactive behavior data, which corresponds to the preset strong purchase intention behavior. The strong purchase intention behavior includes repeatedly visiting the product page, repeatedly visiting the pricing page, downloading technical documents, adding to the shopping cart, and visiting the contact information page, which is used to identify the urgency of the user's purchase intention. The historical email dimension corresponds to user feedback data on historical emails. The feedback data includes email opening behavior, reply behavior, reply content keywords, and sentiment tendency. The sentiment tendency is determined as positive, neutral, or negative based on preset keyword rules.

Images