Bio
Our client is a Singaporean multinational bank with over $10 billion in revenue. In recent years, the bank is transforming to become more like a technology company and less like a bank, with a focus to make banking effortless and joyful. It enjoys patronage of over 12 million banking and credit card users. Furthermore, it offers research reports on 400+ Asian stocks for its retail and corporate investment banking clients. It has been voted the Best Digital Bank consecutively for 5 years.
Situation
The bank approached Alltius with two specific use cases.
In the first scenario, the bank aimed to enhance the information discovery process on its credit card portal by providing clear and concise responses to customer queries. This involved condensing vast amounts of information from various product and service web pages into direct answers to questions like "Which credit card is most suitable for someone with a $10,000 monthly income who spends on dining and travel?" or "Are there any joining fees associated with card X?". By streamlining the search process and delivering specific information efficiently, this enhancement would ultimately lead to higher web to lead conversions.
In the second use case, the bank sought tools to assist analysts in quickly finding information and generating instant insights by scanning through hundreds of analyst reports on specific industry trends. Analysts typically spend a significant amount of time sifting through multiple PDFs without a clear idea of where relevant information may be located. This often results in important material being overlooked, incomplete analysis due to time constraints, and excessive time spent on discovery and synthesis rather than conducting in-depth research. As a result, analysts of all levels would need to meticulously comb through numerous PDFs across hours to piece together a cohesive and comprehensive insight on a particular topic.
Both use cases would eventually need to be in the bank’s private cloud given the compliance mandates of the banking industry.
Solution
For the product discovery assistant, the bank efficiently set up a fully trained Assistant on Alltius’ web application, processing more than 3,000 product collateral pages from their public web documentation in just minutes. Through the web-widget channel, the bank successfully launched a chatbot that aligned with the brand on the very first day of the project. The team rigorously tested the bot with a substantial test set and after some adjustments, it was ready for deployment.
As for the report analyst AI assistant, the first goal was to analyse over 2,500 PDF reports, each containing a minimum of 20 pages and 40 graphs and tables. Alltius’ assistants have the capability to break down PDF content into text segments, convert tables into data frames, and interpret simple to moderately complex graphs into underlying datasets.
A newly developed customised skill was designed and plugged into the assistant to scan text, tables, and charts for specific answers and consolidate these responses into a comprehensive mini-report packed with valuable insights. These answers include detailed references to where they were derived from within the 2,500 odd files along with thumbnail previews of the corresponding charts and tables.
Results
The initial pilots yielded two highly effective assistants, one for end-users and the other for internal research analysts. Alltius demonstrated clear steps to set up an instance of the assistant building platform on a private AWS cloud.
The product recommendation and discovery assistant significantly reduced time to access information from 15 minutes to less than a minute, potentially increasing traffic to the enquiry funnel by double digit percentages.
The text, table, and chart interpreter for employees helped drastically reduce the time it took for human analysts and relationship managers to access information, from hours to minutes. This enabled them to share market insights with clients much quicker and more comprehensively.
The user-facing assistants boasted an impressive precision rate of 99% with a 95% recall rate, while the second assistant, which interpreted complex tables and graphs, achieved a 97% accuracy rate and an 80% recall rate. In both cases, the time required for training and launching was approximately 30 minutes for a dataset of 1,000 pages and files.
Way forward
The bank is now seeking to finalise the implementation in the private cloud through a close partnership model with Alltius after which they plan to expand the usage of the two assistants to accommodate higher web traffic and support more research analysts.
Additionally, recognising the value of the text, chart, and table interpreter, their data science team aims to integrate their research platform with Alltius' APIs to access both insights and raw deconstructed data from numerous PDF reports.
