Basic decision-tree chatbots are easy to configure, but they are limited in how useful they are. For an AI chatbot to really add value, it needs to be able to converse with a person in a meaningful way with a level of understanding that goes beyond ‘yes/no’ or ‘choose 1,2 or 3’.
Conversational AI works using Natural Language Understanding, Natural Language Processing and Machine Learning algorithms to create three core layers of AI automation. Each layer looks at the conversation from a different angle to paint a full picture of what a person is saying. And once the bot understands, it can do something useful with that knowledge, and generate an appropriate response.
Let’s look a little deeper at the three layers of AI automation.
Reading and understanding each utterance
The AI chatbot learns to pick up what the person is saying within a context and what they want to do. The answer to this will determine the next step in the customer journey. For example, in a debt collection scenario:
Does the customer want to make a payment? Then go to Payments.
Does it look like they can’t pay? Then fill in an I&E form.
Through Machine Learning, the collections chatbot builds up an understanding of human language, but in the event they don’t understand, they fail gracefully and hand it over to an agent so that the customer is not stuck in an endless loop.
A ‘fail’ is also an opportunity for the bot to learn and it will know what to do the next time as the algorithm develops.
What elements in the utterance could/should be used
Entity gathering picks out the important pieces of data from within a longer sentence and stores it to use for various actions. Entities includes: dates, dates-of-birth, amounts, postal codes, etc. After picking them out, the system stores them in a standardised format that is used for further processing.
The benefit of this is that a customer can write out their answer in one go and doesn’t have to answer question-by-question.
Also, questions like ID&V and information gathering often fail if exact matching answers are required. However, good entity gathering gives 85-90% accuracy and the AI chatbot should know what questions to ask if it doesn’t understand what a person has said.
Reading between the lines
Once you know the intent and have gathered the entities, the AI propensity engine can determine what the underlying ‘truth’ is that the customer is communicating. It looks at things like: have they changed the way they say something, the words they use, are they more verbose?
For debt collection chatbots, they look for propensity to pay.
The propensity engine will pick up these clues and guide the direction of the conversation accordingly.
Have a look at this customer text.
"I'm really sorry I missed my payment. My job is on the edge. My hours have been cut. Maybe I can pay you thirty pounds next Wednesday."
Amount - £30
Date - ‘next Wednesday’ which gets formatted to an actual date.
Propensity to Pay - Customer is 'vulnerable'
With these three layers of AI automation working together, an AI chatbot can automate high levels of customer conversations without the need for an agent to intervene. Up to and beyond 75% of customer engagement can be dealt with by an AI chatbot, which delivers real operational savings in both cost and time.