… and how it can transform how you handle complex customer queries
AI is being used everywhere to improve things such as customer service, medical diagnosis and animal conservation. AI can make decisions, decipher complex patterns and even learn to drive a car. But most types of AI, especially with Machine Learning, need a vast amount of data before they can “learn” to do anything.
For example, an AI system could analyse data on thousands or even millions of loans and assess which loans went bad. Based on this data an AI system can, with very good accuracy, predict whether a loan is likely to go bad based on the data in the thousands of loans it has analysed. AI can then keep on learning and improving the way it assesses loan applications.
And AI bots (basically webchat with a computer or robot) are being used a lot to manage queries from customers. Many of these bots are great at understanding English and answering questions on generic topics or even given facts like “What is the deepest Lake in England?” The bot could look up that question on the internet and based on the thousands of answers given, give a pretty accurate answer.
But what if you want the AI system to answer a specific question from a small set of rules or fixed information? And that you want the answer to be correct; indeed, it has to be correct. A simple example would be a list of branches and their opening times. You want to be able to ask the AI bot whether a particular branch is open at 6pm on a Thursday. The Bot doesn’t have lots of data to analyse – just a specific fact that it needs to look up. It also may get asked in many different ways such as “is the Branch Closed at 6pm”, or “what time does the branch close”. And there are many more ways of asking the same question.
We wanted to develop an AI chatbot that was able to be taught a set of facts that were contained in a document or set of rules and then be able to give accurate answers to any questions.
This began when a bank set us the task of developing a chatbot that could answer queries on its mortgage lending criteria. The basic challenge we were given was if a broker asked:
“My client wants to buy a flat above a chip shop”
Using large banks of data here wasn’t an option as there was only one answer and it didn’t even refer to a chip shop.
And we also had to develop the bot so that it could give different answers for different lenders and sometimes for a group of lenders.
Standard AI tools didn’t help so we developed our own but utilising a language developed for bots. And the result? Well you can ask that question and get the correct answer. And it will also answer if you ask about a cafe or say a newsagent.
But we didn’t stop there. The bot can also answer questions that involve calculations. For example, “Will you lend £400k on a £500k property on a 20 year term for a 40 year old” And we soon learnt that the bot needed to understand the context of what someone is talking about. If you are asking about interest only mortgages you don’t want all the next answers to be about Capital & Repayment.
The bot can now also fill in forms, send emails and look up answers or calculations in back end systems (such as calculate Redemption figures).
The possibilities are now unlimited as the bot can be “taught” answers on any topic such as Branch opening times. Or Building insurance, or savings information or……. In fact it can be taught to answer anything very quickly and easily.
But remember. It doesn’t learn, it needs to be taught! This is Machine Teaching rather than Machine Learning.