The Taiwan Banker

The varying business nature and regulatory requirements of financial holding companies subsidiaries often makes coherent AI strategy implementation a significant challenge. It is difficult for independently operating subsidiaries with disorganized AI infrastructure and talent allocation to effectively pool technical resources, which can lead to inconsistent progress. To gain a gain a deeper frontline understanding of their implementation strategies and effectiveness, the Taiwan Banker magazine team visited senior executives responsible for overall AI development within several financial holding companies. McKinsey & Company estimates that the full implementation of AI will generate 14%-24% in new profits for the global banking industry – a figure that could expand to 60%-80% within three years. To capitalize on this trend, Taiwanese financial holding companies have deeply embedded AI into their operational processes and encouraged employees to actively explore the possibilities of generative AI. Each has developed unique AI development strategies based on its group structure, resources, and customer base. For example, Cathay Financial Holdings emphasizes group integration. Cathay’s Digital Data & Technology (DDT) Center was established in 2016 to vigorously promote AI and the cloud, integrating the group’s technology resources, coordinating projects across subsidiaries, and establishing unified technical standards to avoid duplication of investment and achieve group synergy. A cloud migration plan was launched, migrating internal systems to the cloud in phases, while also adopting multi-cloud and hybrid cloud strategies, systematically assisting subsidiaries to migrate their systems to the cloud based on use cases and their complexity. Regarding AI development, the GAIA Generative AI Technology Framework was developed to achieve group consensus. AI projects in all subsidiaries can be deployed on the same architecture, enabling the sharing and reuse of models, data, and tools, which significantly shortened its development and deployment time. Cathay cultivates a shared talent pool across the group. Data and digital departments within each subsidiary also hold monthly networking with the DDT Center, ensuring that frontline needs are promptly reflected in the GAIA framework. This culture of breaking down silos has minimized resistance for group-wide innovation, allowing subsidiaries to collaborate in unison. To make AI tools more accessible and accessible, several financial holding companies have developed enterprise-grade AI chatbots for internal use, supplemented by educational training. For example, E.SUN Financial Holdings launched GENIE, its own internal generative AI platform which is open to all units and levels of staff, as an enterprise interface for connecting to external LLMs. The management team tracks the number of unique users and average usage of the platform each month, showing steady growth. By the end of May, over half of all employees had used GENIE. E.SUN implemented several measures to promote this tool. First, it established a prompt template library to provide sample questions for a variety of topical areas, flattening the learning curve for new users. Second, it held bank-wide practical workshops tailored to specific business needs, demonstrating GENIE's usage and case studies. It observed that every time a department completed training, its platform usage rate increased significantly. Employees are using it to automate report drafting, organize lengthy documents, and search internal knowledge bases. Generative AI is considered a smarter form of Robotic Process Automation (RPA), handling unstructured tasks difficult to automate with traditional RPA. KGI Financial Holdings has made generative AI available to employees for internal use, subject to approval and application. The security team monitors usage to ensure data security. Shortly after its launch, employees quickly identified new ways to improve efficiency, with document translation being the most popular application. While English documents previously required outsourced translation, most employees now prefer to use AI followed by human review. Many report that the quality of AI-generated translations rivals that of human professionals, with even more precise grasp of specialized terminology. Furthermore, AI also delivers its results instantly. Staff are starting to use AI to help produce first drafts of promotional copy, which are then polished and finalized by human staff. Previously, writing from scratch required extensive brainstorming, but now AI can quickly generate multiple versions for reference, allowing for personalized marketing copy. Furthermore, wealth management or market analysis departments are using AI to quickly read and summarize reports. For example, they collect multiple external market research reports and input them into AI, allowing the system to summarize key points for daily decision-making. Because these applications don’t directly face clients, employees even compare the summaries of different AI models (such as OpenAI and Google Gemini) to cross-validate the reliability of the information. These cases demonstrate the subtle integration of AI tools into employees’ daily workflows, making many tedious tasks more efficient. Innovation in the underlying models and technologies is essential to seize AI momentum. For example, SinoPac Financial Holdings’ R&D team has been working on “customer genomics” since 2021. Using deep learning, it maps customer behaviors and characteristics into latent vectors, moving beyond explicit fields like age and assets to capture their underlying “genes,” including transaction history, product usage, and digital footprint. This technology aims to uncover customer preference patterns that are invisible to traditional rules, enabling differentiated marketing and service delivery with greater precision. This system is currently being implemented for a variety of digital finance scenarios at SinoPac Bank and its subsidiaries, including SinoPac Securities, covering various aspects of sales (such as fund and insurance promotions), differentiated customer service, and personalized digital experiences. It is highly scalable, and also has the potential to be applied to areas such as risk management and fraud detection. Because each “gene vector” represents the profile of customers’ normal behavior, any abnormal deviation could indicate a risk event such as fraudulent activity. Similarly, the bank can also predict which new account holders are likely to become dummy accounts and conduct preemptive screening. The “Customer Genomics Marketing System and Method” has been granted both utility and invention patents. It analyzes the transaction behavior and digital footprints of over 5 million customers within the group, reducing the time required for targeted digital marketing by 40%, and is expected to save approximately 1,500 hours of marketing work annually. Furthermore, the monthly average credit telemarketing approval rate increased by 75% compared to manual screening, and the average approval amount per list increased by 31%. Furthermore, the activation rate of SMS messages with dormant securities accounts increased by over 30%. Due to a surge in demand for mortgages, Bank of Taiwan launched a Smart Mortgage Valuation Calculator, allowing potential borrowers to enter their property information and quickly calculate their property valuation and loan amount. Employees can also enter property information into the internal system when processing applications, and A1 provides instant valuation for credit approval. Because valuation is directly related to loan underwriting and risk control, the model utilizes comprehensive data and rigorous valuations. Bank of Taiwan has developed models for each of Taiwan’s 22 counties and cities, segmenting property types such as elevator buildings, mansions, apartments without elevators, townhouses, and townhouse land. Over 100 valuation models have been trained to cover diverse regions and property types. Regarding model training and deployment, the bank has iteratively added machine learning operational capabilities, enabling periodic training and automated deployment. To ensure that valuations are calibrated to market fluctuations, the models are retrained and deployed each month using the latest transaction data, ensuring that employees receive the most up-to-date and accurate valuations for each inquiry. Employee feedback and performance evaluations demonstrated positive results during the implementation phase. A later bank-wide survey revealed that over half of users believed the model significantly benefited their business. In the past, the life insurance industry often relied on external consultant physicians to review medical diagnostic certificates (such as X-rays) when processing claims, but the physicians were typically only on-site one or two days per week, often resulting in waits of weeks or even months from submission to approval. This was particularly true for cases involving fractures and chest imaging, where long wait times became a significant pain point for claims processing, impacting both claim processing timelines and customer satisfaction. To improve processing timelines, Yuanta Life Insurance implemented A1 medical imaging interpretation, which uses deep learning to automatically identify key features in X-rays. For example, for fractures, the model can determine the presence of the fracture and whether it meets the policy’s claim criteria. For chest X-rays, it can detect lesions or abnormalities. The model is trained on images and physician interpretations from previous cases, annotated by medical experts and data scientists, and Yuanta purchased GPU servers specifically for model processing. Using AI, the time needed to interpret and summarize chest X-rays was reduced from an average of two days to approximately two minutes. Images of limb fractures were also reduced from several days to just 10 minutes. AI has proven to accurately and reliably process the majority of routine claims, with only a few difficult cases requiring manual processing. Cathay General Insurance is also actively integrating AI into its insurance value chain with the SmartClaim claims assistant. Using a LLM for natural language modeling plus a lexical analysis module, it accurately interprets judgment documents, provides adjusters with loss estimates for physical injury claims, and establishes quantitative compensation guidelines, which has sped up loss assessment time from an average of 30 minutes per case to just five minutes. SmartClaim also utilizes a Social Network Analysis (SNA) association model to identify rare suspicious patterns, potentially identifying fraud networks and high-risk nodes. Since its launch, it has successfully detected and prevented over 890 high-risk insurance fraud cases, worth NT$29 million. Furthermore, Cathay’s CarTech Intelligent Auto Insurance Value-Added Service for corporate fleets collects vehicle driving data through connected vehicle devices, and uses AI to analyze driving patterns. Cathay General Insurance has pioneered a new model for value-added auto insurance, helping corporate fleets identify driving risk indicators, improving high-risk behaviors and safety systems, and reducing accident rates, which in turn helps Cathay better fulfill its corporate social responsibility of damage prevention. From unifying cloud infrastructure and cultivating cross-subsidiary talent to empowering generative AI platforms and implementing scenario-based applications, the core operations of financial holding companies are already experiencing combined effects of efficiency and innovation. The author is an assistant researcher at the Institute of Financial Research of TABF