With the development of retail business, banks generate massive voice data through channels such as phone and WeChat. The traditional quality inspection methods have exposed problems such as low execution willingness of branches, insufficient manpower at the head office, weak complex semantic recognition, high costs, and inability to empower the business. Against this backdrop, Cloopen Ai collaborated with a city commercial bank to create an intelligent quality inspection system based on large models. Through five core innovations, it comprehensively resolves the problems of traditional quality inspection in terms of coverage, accuracy, efficiency, and cost.
Innovation Point One: Multi-tenant and multi-level user capabilities, achieving data hierarchical control
The Cloopen Ai large model quality inspection system has multi-tenant isolation capabilities. It can independently divide task allocation, data storage, and report statistics dimensions based on different business scenarios such as marketing, operation, and customer service, ensuring that data from different scenarios do not interfere with each other and meeting the requirements for refined management.
At the same time, the system supports three-level user configuration of head office - jurisdiction bank - branch. Through strict permission control, it ensures that each quality inspection personnel can only view data that matches their own responsibilities (such as branch personnel can only see the quality inspection results of their own branch, and the head office can coordinate the data of the entire bank), achieving a hierarchical monitoring model of "data hierarchical visibility and management responsibility at each level", which not only ensures data security but also improves management efficiency.
Innovation Point Two: Multi-modal full coverage, completely eliminating compliance blind spots
Traditional manual quality inspection relies on sampling, with a coverage rate of ≤ 5%. Over 95% of voice data is in an unregulated state. The Cloopen Ai system breaks through channel barriers and supports unified access of data from multiple channels, achieving the collection of customer interaction information without omissions, ultimately achieving 100% full quality inspection.
Whether it is voice data from telephone marketing, text conversations from online customer service, or audio and video information from video customer service, all can be included in the quality inspection scope, completely eliminating the compliance blind spots of traditional quality inspection, ensuring that the service quality and compliance management of the entire bank have no blind spots.
Innovation Point Three: Large and small models collaboration, breaking through the bottleneck of complex semantic recognition
In response to the complexity of quality inspection tasks, the system adopts the "efficient collaboration of large and small models" mode to balance accuracy and cost:
- Small model handles basic tasks: responsible for sensitive word detection (based on the preset word library to match illegal text), emotion recognition (analyzing voice features to determine the speaker's emotion), volume / speech rate / silent interval monitoring (ensuring the smoothness of the call) and other basic rule-based quality inspections, quickly completing standardized verification;
- Large model tackles complex scenarios: focusing on logical contradictions (such as "the customer confirms the risk first and then asks if it is guaranteed") and implicit risks (such as ambiguous responses from the customer during follow-up), using natural language custom rules to conduct context-related analysis of the entire conversation.
This mode reduces costs while increasing the accuracy of complex semantics to over 90%, with a missed detection rate of less than 2%, achieving a qualitative breakthrough compared to the 50% missed detection rate of traditional NLP quality inspection.
Innovation Point Four: Model autonomous iteration and optimization, continuously improving quality of quality inspection
The system builds a "automatic detection - manual recheck - model iteration" closed-loop optimization mechanism to ensure that the model capability dynamically improves with business needs:
- Automatic marking of high-risk items: The system detects call data in real time and automatically marks suspected violations, semantic contradictions, etc., high-risk content;
- Manual recheck correction: High-risk items are pushed to the manual recheck team for correction of model misjudgment results;
- 3Incremental training iteration: Regularly incorporate the results of manual corrections into the model training library for incremental training to optimize the model algorithm;
- Appeal review to ensure fairness: Supports agents to appeal the quality inspection results, and compliance experts complete the review within three days to ensure transparent and fair determination.
After this city commercial bank launched the system, the model accuracy rate increased from 90% to 95% after three iterations, with a 40% reduction in misjudgment rate, continuously meeting the actual business needs.
Innovation Point Five: Private deployment + industry adaptation, meeting financial-level security and business requirements
Considering the sensitivity of banking data, the system supports private deployment, with all data stored on the bank's own servers, meeting the strict requirements of the banking industry for data security and privacy protection, and avoiding the risk of data leakage. Meanwhile, through domain pre-training and fine-tuning, the model is deeply adapted to the business scenarios of city commercial banks: special optimizations are carried out for bank marketing script norms, compliance clauses (such as anti-money laundering reminders), and key nodes of business processes (such as core issues of account opening follow-up), ensuring that the quality inspection rules are highly matched with the actual business of banks, and avoiding the problem of "general model not adapting well to local conditions".