2023.04 The Taiwan Banker NO.160 / By Ingrid Chang
What will the intersection of finance and ChatGPT look like?Banker's Digest
Over the past two months, the ChatGPT generative predictive model, which now boasts one-hundred million active users, has elevated the development of AI. With everyone discussing what kind of new spark this technology will bring to their professions, the Taiwan Banker magazine organized a “ChatGPT + financial sector” panel hosted by TABF Deputy Director Dennis Lin, who invited two experts on AI: Song Ming-yuan (Microsoft Taiwan, Senior Enterprise Application Business Associate) and Jiang Meng-hong (Microsoft Taiwan, Cloud Technology Consultant). Financial professionals were including Liu Mei-ling (All Win Fintech, President), Liu Pei-wen (First Bank, Vice President), Wu Jen-yu (Hua Nan Bank, Digital Operations Department Manager), Zhang Xiao-wei (Taiwan Cooperative Bank, Electronic Finance Department Associate), Cai Yu-chen (Taiwan Business Bank, Digital Finance Department Associate), and Liao Jia-hong (Sinopac Holdings, Director of HR) were also invited. (In order to safeguard the privacy of each bank, respondents’ names are replaced with letters in the following interviews.) Lin, TABF: ChatGPT is easy to use and knowing how to use it is key Associate Director Lin kicked off the discussion by mentioning that TABF’s chair had sent him and several other managers to Academia Sinica to attend a six-month course on AI in 2018. As someone without an IT background, he found the course to be rather difficult and thought AI technology was high-level. However, after witnessing the emergence of ChatGPT this year and trying it out himself, he found it easy to use, and is now interested in understanding how to apply it. Lin wants to “use it, not be used by it.” At the beginning of the event, Song Ming-yuan said that that Microsoft had a special vision, and invested in OpenAI, helping it develop ChatGPT. In 2019, it signed an exclusive cloud supply contract with OpenAI, which would use Microsoft Azure to provide computing power for model training. In exchange, OpenAI would supply its services to the Azure cloud platform. This demonstrates Azure’s ability to meet bespoke, high-level information security needs. Three major advantages for digital marketing and embedded finance A: Everyone is curious about the differences between generative AI and its predecessors. Behind ChatGPT’s chatbot are different models, including the famous Da Vinci model, which has advanced from version 3.5 to 4. Early models collected training data from all over the internet, which was then labeled by researchers. Prior to the labeling, version 3.0 was highly sequential, while versions from 3.5 onwards added many more labels, allowing the AI to focus on causal relationships and relevant dialogue. For example, a bank who wants to produce audit reports or internal bank documents (like legal affairs) can use these labels. This is simply sentence training. As one can imagine, keywords are extremely important in order to allow the AI to speak accurately, especially when it comes to training. Giving the model accurate information and then fine tuning it repeatedly according to the situation enables it to provide progressively more accurate answers. Because of this, it is important know how to provide accurate keywords to accomplish a desired task. Accurate tags are extremely important. It is also worth making note of OpenAI’s Codex function. In the past, when banks wanted to develop digital services, they needed talent who understood both finance and technology. Finding such skills was always a problem. Now, however, AI can help write programs. Previously, a Hong Kong InsurTech company called OneDegree hoped to develop some AI services to be integrated with Microsoft systems, but linking even internal life insurance systems together is complex, much less internal systems to external partners. Using OpenAI’s Codex to turn ideas into code, however, they were able to allow each business line to develop relevant forms and workflows. In the past, it was necessary to provide IT with comprehensive request forms, but now, 80% of code produced through speech is usable, instantly increasing efficiency and competitiveness. Banks previously had experience using AI. According to bank data, physical personnel serve approximately 20-30% of clients, which leaves 70% without a way to access more in-depth services. Accordingly, banks made frequent use of AI customer service systems. After one bank used a verbal AI customer service system, they realized that its success rate for outgoing calls was higher than they thought. For example, when calling to remind customers about bills due, most of them paid what they owed. On the other hand, when customers initiated the call, the system could only process simple questions, and there was no way to upsell. This is one of the current system’s blind spots. If a bank wants to replace previous AI customer service systems using OpenAI, it might first input all its information into the AI language model. From there, it could produce sentences that make more contextual sense via fine-tuning. Vectorizing its information and clients’ questions, storing them in a database, and then allowing AI to decide which of the answers comes closest will support the development of contextual finance. ChatGPT has several advantages compared to past AI customer service models. The first is its heightened understanding of coding and generative ability, as well as its ability to understand and tolerate input, which allows it to produce expert answers in most areas of knowledge. Second is its incorporation of ethical principles, allowing it to discern which information is malicious and consequently refusing to provide an effective answer. The third is its ability to produce continuous dialogue based on past inputs, elevating the model’s interactive experience. Banks can also adjust the questions and answers to produce different genders, age ranges, speaking styles, and languages for different marketing effects. OpenAI has already released the GPT-4 model, whose most important feature is not just written dialogue, but also ability to interpret images. Additionally, its explanatory ability and accuracy substantially surpass GPT-3.5. What is more, banks can control its persona to match their desired image. The number of token inputs has also greatly expanded to some 30,000, equivalent to 50 pages of text. The text input API tool has been released first, and paid ChatGPT Plus users have priority to try out the new version. The image input tool is currently only available to designated partners. Future prospects B: I believe that the financial sector will have three major avenues for expansion with ChatGPT. The first is conversational business. In had previously seen chatbots when attending an international conference, but they were unable to process Chinese. Fast forward to today, my first thought upon seeing ChatGPT was that it could be a big help for contextual finance. As the financial sector is aiming to cooperate with different industries, it is thinking of ways to find and connect every possible context with banking services. Previously, we relied on APIs to perform any kind of information exchange, which was quite burdensome. With ChatGPT, however, we can design conversation-linked commerce. For example, if a bank customer wants to go to Tokyo for six days, a travel itinerary or promotional packages from partners could be shown in the response, like travel groups and flights. After the customer selects an option, they can pay for those services, and even complete any insurance processes, truly incorporating finance into business contexts. The second avenue is ChatGPT’s potential to creatively empower bank staff. I have recently seen some foreign banks use ChatGPT technology to produce marketing materials. Historically, some business units needed marketing personnel to help them with design. However, as business units may struggle to write good marketing, and marketing personnel may not understand the product content so well, some marketing has not been so effective. However, ChatGPT will be of immense help for marketing plans. Its recently released code-writing capability could helping bank staff strengthen their technology skills. These should all be directions for human-machine cooperation, and we can expect even more employees to make use of ChatGPT for creative empowerment. Finally, ChatGPT could help with business insights. Banks’ internal processes are full of breakpoints. Information collection requires a person to determine its accuracy, and for the most part, current automated processes can only process structured data. Banks are also unable to spend large amounts of money on machine learning for the sake of analysis. In the future, if they are able to keep up with this new technology, AI should be able to quickly help them analyze big data. Even images and audio can be analyzed, not just written text. Banks regularly produce a lot of data, but most rely on employees to analyze it and follow the market in order to find business opportunities. However, a gap often exists with actual operations. The future is big data, and AI can be used to define and to improve the quality of decision-making, growing business opportunities. Many banks aspire to pair big data with AI, but they are constrained by resources, talent, and even technology. After seeing ChatGPT, however, they will instead feel that the future holds revolutionary changes. Will ChatGPT replace human work? C: After accepting that ChatGPT is a tremendous AI, professionals will naturally ask whether or not it will replace human work. We can imagine that future banking apps could include AI-generated tellers’ faces and voices. When a client asks how much money is still in their account, and how it could be managed, the teller will have several pre-set answers. Prototypes of such faces already seem imaginable. The pressure of personnel role transformations could swell in an instant for tellers and customers alike. HR managers are also nervous, and must consider how to guide bank staff even more quickly through this transformation. Impact on the banking industry D: Although legislation does not allow AI bank staff to immediately engage with clients, if this sort of generative model is further combined with audio and images, I think the industry impact will be enormous. However, we ought to make use of this trend as soon as possible. Our bank has already established a taskforce to make necessary preparations, but this process has been difficult. For example, how many employees can accurately input so-called business content into the AI? Who should do so, and how? Data definitions and regulation of its use are also enormously important. Internal use for now E: We have heard about the many benefits of ChatGPT’s, but when it comes to final responsibility, the financial sector cannot afford any mistakes. Even if fine-tuning allows ChatGPT to study every kind of financial operation, and AI can assist tellers, increase efficiency and reduce mistakes, the responsibility of selling a final product still lies with people. Right now, it seems unlikely that we can allow replacement by machines, and it remains to be seen how financial regulations may be relaxed. As for whether AIs can truly replace human staff, everyone still trusts real people. I personally think it is impossible in the short term to directly allow AI to interface with clients and consumers. Therefore, internal applications will be more appropriate at first. An AI could be aptly used as a digital assistant or secretary, helping draft paperwork or conduct preliminary surveys. In the end, however, humans will still need to analyze and digest the output. Pressure of transformation F: While it seems unlikely that AI will rob tellers of their jobs in the short term, the popularity of ChatGPT has caused them to feel pressure to transform. In the past, digital transformation was mostly initiated by upper management, but now, it can be initiated by general employees. We have seen that transformation has three important facets: business model, operating model, and management model innovation. The former two are difficult to direct from the top down. Basic employees also directly engage with clients on routine tasks on a daily basis. Transformation could be truly useful here. It is even more important that employees be allowed to experiment with the technology and discover its value. They should be allowed to play with ChatGPT immediately, and then we can ask them how it helps them write text and code. In conclusion, the financial sector has innovative perspectives on ChatGPT, as well as a few concerns. Even if we want to rely on AI to find business opportunities, human intelligence is still required in the back-end. AI relies on employee knowledge to build the content needed internally, which must have a sound basis. Therefore, producing grounded AI content and services will take hard work.