In the digital era, AI is regarded as an important starting point for the fourth industrial revolution and has become a new engine to promote the performance growth of various industries. In recent years, the rise of new applications and new demands, such as artificial intelligence models represented by AIGC, has driven the financial industry to deepen its deployment of "intelligent capacity building" and empower financial business. In the process of digital transformation, CEB has actively seized the opportunity to make technology the core engine of the second curve of development with the technology positioning led by integration, formulated a new phase of the technology strategic plan, and built a core technology engine strategy of "three new and three modernizations" with new experience, new model, new integration, and centralization, agility and intelligence (as shown in Figure 1).
In this context, exploring the construction of Everbright's unique AI technology middle platform will help accelerate the large-scale promotion of AI applications, improve productivity and efficiency, and promote the digital transformation of Everbright Bank.
1. Practical exploration of the construction of artificial intelligence technology middle platform
1. Construction guidance
CEB is actively promoting digital transformation and implementing the new phase of the "14th Five-Year Plan" of the technology strategic plan. The "14th Five-Year Plan" includes nearly 100 key projects, including ten major business categories, four major governance categories, and five major technical engineering construction, among which the technical middle platform is responsible for scientific and technological support to ensure the stable operation of the system and continuous technological innovation. As an important part of the technology platform, the artificial intelligence technology platform has become an important task in the implementation of CEB's technology strategic plan. Adhering to the construction concept of "three new, three modernizations and five principles", the AI technology middle platform uses advanced architecture design to promote digital transformation and evolution, generate innovative achievements, and support the high-quality development of multiple business lines.
2. Architecture system design
In the past, the construction process of artificial intelligence often showed the characteristics of fragmentation and silos, which brought problems such as duplicate construction of functions, difficulty in getting through data, high cost of business interaction, and lack of business precipitation. In order to further support the development of key businesses, CEB abstracts and integrates AI technology and infrastructure in the process of designing the AI technology middle platform architecture to form a unified technology platform to provide shareable technical services for all business lines of the bank, so that the banking technology can respond to business needs more flexibly and efficiently.
At present, CEB has built a unified, autonomous, controllable, and fully functional AI technology middle platform architecture system (as shown in Figure 2) for the whole bank, with five types of capabilities: listening, speaking, reading, watching, and doing, and empowering banking business in a unified manner. Starting with the construction of intelligent scenarios such as identity recognition and intelligent customer service, the artificial intelligence technology middle platform has gradually precipitated enterprise-level AI service capabilities supported by mature AI technologies such as automatic speech recognition (ASR), natural language processing (NLP), speech synthesis (TTS), optical character recognition (OCR), biometrics, virtual portrait synthesis, content review, document comparison, and document extraction. Mature intelligent applications such as intelligent dual recording empower services, and AI training platforms and inference AI middle platforms are used to customize AI models and manage services.
The AI technology middle platform contains several key layers: one is the hardware resource support and resource management engine layer, which provides the necessary computing and storage infrastructure for the entire system to manage hardware resources, optimize resource utilization, and ensure that all parts can run efficiently. The second is the AI data engine and AI platform engine layer, which is committed to providing developers with a full-link development flow environment, including managing and annotating massive data, model development tool frameworks, and supporting the easy construction and management of various AI models. The third is the AI capability engine and AI service engine layer, which can provide core AI algorithms and models to support various tasks. The AI engine translates AI capabilities into standardized services and interfaces, simplifying integration with other systems for broader applications. The fourth is the AI ecosystem, which is the top layer of the entire system and involves the construction and management of the entire AI ecosystem. By building a healthy and open ecosystem, the AI technology middle office is better able to adapt to evolving technology trends.
Second, the effectiveness of the artificial intelligence technology platform
1. Promote the evolution of digital transformation
The AI technology middle platform implements the default principles of the middle office at the application architecture level, bringing higher flexibility and efficiency, and helping banks usher in a profound digital transformation evolution.
The first is the transformation from the dispersion and repetition of capabilities to the management and reuse of capabilities. The AI technology middle platform builds standardized AI capabilities, uses the AI competence center to take over third-party capabilities and self-developed capabilities, outputs standard intelligent capabilities for upper-layer applications, and unifies transaction codes, packets, and protocols, connects AI capabilities and business data returns of application systems with standardized data return, improves the frequency of model iterations, and uses an agile development mechanism to open up the whole process management of MLOPS model lifecycle. Through centralized construction and collection of standardized AI capabilities, capabilities can be managed and reused between departments and scenarios, avoiding duplication or decentralized construction. The middle office adopts the default component reuse mode to quickly build applications and support multi-system functions, making the banking business more modular and flexible, and providing strong support for rapid adaptation to market changes.
The second is the transformation from a system perspective to a scene perspective. Through the joint innovation of "intelligent capability + middle-platform scenario + business scenario", the artificial intelligence technology middle platform has formed a model of incubation of scene tool circles (as shown in Figure 3). On the one hand, it is connected with intelligent atomic capabilities to provide a variety of combined and orchestrated middle platform capabilities, and on the other hand, it is connected with business scenarios, co-researching processes with businesses, and jointly incubating scenario-based capability tools. On the basis of the underlying capabilities, the scenario-based construction can not only precipitate the common capabilities of similar business scenarios between different business departments, but also quickly realize the business scenario-based requirements.
The third is the transformation from single-dimensional model analysis to intelligent multi-dimensional model. The single-dimensional model only focuses on the information of a specific domain and cannot consider the information of other data modalities, so it has limited ability to deal with complex problems. The multimodal model can obtain and integrate multi-dimensional information, so as to obtain a more comprehensive data understanding, so that financial institutions can use multi-dimensional information more comprehensively and accurately, which helps them to more accurately understand customer behavior, market conditions and risk factors, and improve personalized service capabilities and the overall and accurate decision-making.
Fourth, the shift from manpower-intensive to digital workforce. Digital employees replace manual work to complete high-intensity, high-repetitive, and patterned work to improve enterprise-level intelligent productivity. According to the characteristics of each job position and the practice of professional lines, CEB divides digital employees into two categories: customer service and internal empowerment. In terms of customer service, through the establishment of customer service robots, immersive interactive self-service in the scene is realized, so as to improve the convenience and accessibility of financial services, and in terms of internal empowerment, digital employees are used to fully replace employee manpower for data entry and processing, work order management, information comparison, report generation and compliance inspection. A digital workforce frees employees from tedious, repetitive tasks and allows them to focus on creative work.
2. Empower your business to show initial results
CEB's AI construction has evolved from "system construction of intelligent capabilities" to "ecological construction of making good use of intelligent capabilities". Through the "business + intelligence + middle platform" operation system, intelligent capabilities are widely used in various business scenarios. The "middle-end thinking" guides the deep integration of intelligent capabilities and services, realizes standardized AI capabilities, and improves delivery capabilities. Through "operation", AI R&D and application promotion are connected to achieve large-scale application of AI capabilities. Nearly 200 capability libraries are collected through the AI middle platform capability precipitation, so that the AI capability output covers multiple business fields. CEB has achieved full coverage of intelligent applications across the bank, accumulated more than 500 subdivided scenarios, and achieved remarkable enabling results, including significant improvements in breadth, depth and speed, providing strong technical support and innovation impetus for banking business.
At the level of customer reach, the AI technology platform continues to strengthen ecological capacity building, adopts all-round strategies, and makes multi-directional efforts in digital channels to empower smart remote services, cloud payment and other fields. In the field of smart remote services, we will build intelligent customer service functions such as intelligent Q&A, telephone navigation, intelligent outbound calls, intelligent quality inspection, and intelligent assistance, making them the key pillars of online operations, and operate digital products represented by "Sunshine Xiaozhi" and digital human "Xiaoxuan". During the epidemic period, the peak service of "Sunshine Xiaozhi" online incoming parts accounted for more than 96%, with an average daily service volume of nearly 300,000 person-times.
At the level of customer marketing and operation, the AI technology middle platform provides highly personalized customer operation capabilities for retail customers, corporate customers and private bank customers, so as to support the continuous enrichment and segmentation of multiple types of customer groups and scenario applications, including intelligent dual recording, online supermarkets, remote investment advisory, digital business cards, agency sales review, etc. These capabilities cover multiple fields such as deposits and loans, inclusive loans, online loans, credit cards, etc., and provide a full range of technical support for different businesses.
At the level of internal intelligent operation, the AI technology middle platform quickly promotes the comprehensive application of new technologies in business operations, digital risk control, and intelligent operation and maintenance. In terms of business operations, we have established tool platforms such as intelligent document platforms and intelligent audit platforms to significantly improve operational efficiency, in terms of digital risk control, we have used artificial intelligence and RPA to carry out risk alerts, data maintenance and payment monitoring to strengthen integrated risk prevention and control, and in terms of intelligent operation and maintenance, we have established work order inspection and account inspection mechanisms to promote intelligent transformation. These innovative measures have improved operational efficiency and provided solid support for CEB's digital transformation and intelligent operation.
At the level of smart office, in terms of knowledge process optimization, the "knowledge base + large model" is used to realize the creation, precipitation, flow and application of business knowledge; in the optimization of communication process, it provides intelligent assistant "small eggplant", intelligent schedule, and intelligent meeting minutes to ensure that business personnel can communicate instantly, efficiently and intelligently; in terms of green office, it provides a series of convenient office tools such as document conversion, document online signature annotation, document photo and scanning, etc., to improve the green paperless rate.
3. Challenges faced by the artificial intelligence technology middle platform and how to deal with it
In recent years, driven by the development of independent and secure computing power and cloud native technology, the scale of computing power and data in the financial industry has achieved rapid growth. AI large models are currently a hot topic in the field of artificial intelligence, and how to effectively use AI large models has become a challenge for the financial industry. The following mainly analyzes the challenges faced by the development of artificial intelligence technology in the middle platform from the aspects of computing resources, algorithms, data resources, generative AI security, and talent training, and proposes countermeasures.
1. Computing resources
Computing power resources are the infrastructure of AI, and AI training and inference rely on computing power support, and large computing power also brings huge economic pressure to data center infrastructure, hardware, and power consumption. First, with the rapid growth of AI demand in the financial field, the growth rate of demand for computing resources caused by various AI capabilities has brought pressure on computing power. Second, due to the procurement restrictions on AI training computing power cards, the difficulty of computing power procurement has increased. In response to these problems, CEB has integrated its existing computing resources internally to increase the supply of computing resources, and at the same time, it has actively explored independent and controllable computing power cooperation plans to meet the demand for computing power by using diversified and high-quality computing resources.
2. Algorithms
The challenges in terms of algorithms are mainly from large models. From the perspective of resources and costs, not every financial institution needs to independently train a private basic large model from scratch, and it is a more realistic choice to invest in the secondary optimization of the large model and the construction of application scenarios. Adopting a multi-party joint exploration model, CEB cooperates with banks, leading large-scale model enterprises and application integration service providers to build characteristic large-scale models and applications based on the needs of the industry through the method of "open source + joint creation + co-research and application", so as to improve the application practicability of large-scale models.
3. Data Resources
In terms of data resources, CEB has established a unified governance of text/documents, images, videos, audio and other corpus to fully meet the needs of financial natural language or visual small models and L2 large model training. On the one hand, we will continue to mine the value of existing data, improve the unstructured data management system and professional annotation system, and on the other hand, introduce external data as needed to supplement the needs of L1 large model training in vertical fields.
4. Generative AI security
Recently, the National Information Security Standardization Committee issued the Basic Security Requirements for Generative AI Services (Draft for Comments), proposing 31 major security risks. In response to these problems, CEB has taken the following countermeasures: First, in the selection of application scenarios, it has established an internal first-class mechanism to use generative AI as an assistant for employees in different positions. Second, in view of the technical weaknesses of the large model itself, it mainly ensures the data quality, parameter adjustment and prompt word engineering, evaluates and calibrates the results, etc., and improves the accuracy of the response by combining with the vector database. The third is to safely filter the input and output for the risks of ideology, values, racial prejudice, and "malicious poisoning", and propose a complete set of audit and filtering system solutions.
5. Talent development
In terms of talent training, since the construction of the artificial intelligence technology middle platform requires multidisciplinary talents, and such talent resources are relatively scarce, it is necessary to introduce, cultivate and make good use of financial technology talents in an all-round way, and cultivate compound AI talents who "understand business and understand AI".
CEB continued to promote the centralization and intelligence of banking business, successfully built a unified, independent and controllable artificial intelligence technology platform for the whole bank, provided efficient and intelligent support for business through standardized construction, large-scale promotion and refined application of scenarios, and promoted the comprehensive application of digital transformation in multiple fields such as customer reach, marketing operations, internal intelligent operations and smart office, and achieved remarkable enabling results. At the same time, in the face of multiple challenges in the artificial intelligence technology platform, CEB will continue to improve its technology, enhance the level of technological innovation, and provide customers with better and more convenient financial services with more intelligent service capabilities.
Author:
Pei Yamin, Deputy General Manager of China Everbright Bank Technology R&D Center
Zhang Jie, Deputy Director of Technology R&D Center of China Everbright Bank
Wang Xichen, Science and Technology R&D Center, China Everbright Bank
Source-WeChat public account: China Credit Card
Source: https://mp.weixin.qq.com/s/kcR3IVU8h3vhSiZlIGChgg