This has been a fascinating week of discovery for me in the world of AI development.
It was my usual business trip to Yangon for my financial consultancy work after a 2 days period of brainstorming last week in S’pore. We had finally set up a digital transformation road map and game plan for our microfinance company for the next 12 months. In order to service 500,000 clients in 5 years from the current 30,000+, we need to have a quantum leap in innovation. The 8 pillars of innovation action plan we had mapped out – they all point to fintech developments that involve some element of AI evolution. I will elaborate on 2 of them below.
The first is about new chatbots. For most websites, we already have a chatbot button at the corner of the screen for user interaction and assistance. A majority of these are already automated with machines that handle up to 80% of the inbound traffic while the balance of the 20% are referred to actual human call centre personnel if the machine is unable to service further. In the old system, we have to program all the possible questions and answers into the database for the system to figure out the best response to each and every enquiry from a human user.
In the latest re-iteration of a chatbot, AI is added into the equation. What if we can set up a chatbot that is evergreen, learns on the job and stays relevant indefinitely as it builds up knowledge over time? This is possible now. AI helps the system to learn and then it uses statistical probability calculations to figure out the best reply. With NLP (Natural Language Processing) capability, the machine will be able to fully understand all requests and make an educated guess on the likely answer. Over time, it will get better at what it is tasked to do, as it collects more data and learns on the job.
The advantages of using an AI chatbot can also be channelled into many other areas within the company. What if we use this for internal purposes? Employee questions can be addressed and HR is able to outsource this function. For Marketing, they can use this to train new employees or do periodic training certification of staff members.
We are currently talking to a Canadian based company that has been doing this for a number of firms in North America. The initial discussions sound promising. We should be able to do more very soon. They have shared with us that the end to end process can take as little as 3 months, from the initialization of the project to going live.
The next pillar of innovation was on the development of a dynamic credit scoring model. Currently, we have a rules-based credit evaluation system, where only internally collected data from loan officers are inputted into the system. We have to constantly adjust the system as and when defaults increase. Even then, due to the size of the data, adjustments are still patchworks without any degree of certainty.
What if we can gather unlimited amounts of external data into this credit evaluation process, to find out what works and be able to fine-tune it at the push of a button? We can pull in data like geo-tagging of google map information on land size and acreage, weather reports – both current and historical. We can even gather social media data to determine if the repayment probability of a person is high…
AI excels in machine learning and with more data, the better it becomes. With unstructured data, it can detect trends that are invisible to human eyes. Using re-enforce and deep learning concepts, the AI will be able to run the process millions of times to optimize end results.
For our company, this will mean that we can approve more potential loans and lower/maintain the probability of default rates. We can then predict with a high level of confidence that these approved loans are safe, due to the analysis of internal and external big data gathered.
We had a conference call with a fintech company from South America on Wed evening that had such a credit scoring model. They had been successfully using AI to derive credit scores on a number of agriculture microfinance companies.
As they walked us through their deck, I realized that some of the terms they used were exactly the ones I had learned from my business analytics course! In one of our projects, we were taught the CRISP-DM (Cross Industry Standard Process for Data Mining) :
Using the SAS Enterprise software, we used a structured process to come up with the best model for implementation. Most of the work was in getting the data ready (steps 2 – Data understanding and 3 – Data preparation). Then one will tell the software to use the various models available (decision trees, clustering, neural networks etc.) to crunch the data. Finally, we evaluate all of them to determine the one with the best fit and highest score. This will be the one to implement.
In my project, I had 5 data fields which I used to construct a model that was for the approval of a short term loan. I only had about 3,000+ data in each field but the AI could have easily absorbed a million data points. Once implemented, we could periodically rerun the whole process again with more data or additional fields. Unlike the traditional process which can become outdated over time, this AI-based process has a longer shelf life.
We are now embarking on a live Fintech experiment and journey to transform a traditional business that is so ripe for digital transformation. If we can pull it off, we will set new standards for the rest of the industry to follow. It will enrich the end-users and benefit our bottom line as we optimized all areas using limited resources.
Leave a Reply