The Inner Workings of AI Systems

5 minute read

AI systems are not created equal. Some are more complex and capable than others. Read this article to find out how IPsoft’s Amelia’s brain works and how she helps customers resolve all kinds of issues.

AI systems come in many different forms. How they interact with people depends on how they’re programmed and what tasks they’re trained to perform. At the most basic level, companies program and train AI systems to perform tasks, make decisions, and interact with humans, all on their own.

We believe IPsoft’s Amelia is the most dynamic AI system thanks to her cognitive learning abilities, autonomic task management, and emotional intelligence. In this article, we’ll walk you through how Amelia’s brain works, how she can be used to resolve customer service issues, and how she can help employees do their jobs faster and more intelligently.

The Initial AI System-Human Interaction

The entry point into a conversation with Amelia is through what we call a user utterance. That user utterance can be delivered via a chat box on a website, a voice call either through a mobile or a home assistant such as Amazon Echo, within a chat app such as Facebook Messenger, or wherever a company would like customers to interact with Amelia.

Before Amelia ever speaks to a customer, we use the Business Process Model and Notation 2.0 specification to model a company’s business processes and ensure that Amelia does exactly what the company wants her to do in any given customer-facing situation. So, if a credit card is lost, if an account is being closed, or if a customer wants to open a new account, Amelia will have learned and mastered the company’s business processes before ever fielding a customer query.

By using stochastic spread activation Amelia can handle dialogue variance with no issues.

Once Amelia has been fully trained to perform a specific task, such as an insurance claims processor, a fraud detector, or a mortgage processor, she’s ready to interact directly with customers. When dealing with a customer, the first thing Amelia's brain has to process is which language she’s going to have to speak. We've leveraged machine translation algorithms to try and detect which language the customer's speaking, and have Amelia switch languages on-the-fly. If you say, "Hi," she says, "Hello." If you say "Como estas," she says, "Estoy bien, gracias." She’s fluent in 10 languages and can be trained to speak utterances and phrases in 100 other languages.

As with any other AI system that interacts via voice, Amelia uses natural language processing (NLP) to understand and speak in coherent and human-like sentences. This allows Amelia to sort through someone’s sentence to find meaningful terms and phrases that she’ll use to prepare her response. If you were to say, “I bought a car yesterday,” Amelia would sort out that “car” is the noun, and “bought” is the verb.

Because we primarily train Amelia to handle customer and employee-related issues, we’ve also added intent recognition to her capabilities. We use neural network algorithms to detect intent. If a customer says, “I lost my credit card yesterday,” Amelia will remember her training as a credit card replacement agent. She’ll know that in the case of a lost credit card the customer’s intent is typically to deactivate the missing card, get a new card issued and resolve any disputed charges. She takes basic data (Who, What, and When) and determines that you lost your Gold Card in Brooklyn last night. She’ll see that a charge was processed this morning in Connecticut and immediately recognize this charge as suspicious.

But what happens when someone enters an interaction with multiple intentions? For a typical chatbot that isn’t armed with cognitive intelligence, multiple intentions cause confusion. Amelia can not only register multiple intentions she can also triage them to ensure the most important processes are handled first. For example: If someone calls and says, “I would like to go paperless, but I lost my credit card yesterday in New Jersey, and I think there might be fraudulent charges on my account,” Amelia won’t handle those requests in order. She’ll triage and determine that fraudulent charges are the most important element of the conversation, reissuing a new card is the second-most important element, and going paperless is the least important element. There is no limit to the number of tasks Amelia can triage. She is able to track and remain aware of context. She can also switch context with you and be aware of the previous context for later in the conversation, or for future conversations.

Unlike simple chatbots that are trained to speak off of a script, Amelia is able to handle variance in dialogue. For example: If a chatbot is handling a credit card replacement, and in the middle of the conversation the customer realizes he was discussing the wrong credit card, a chatbot would get confused and need to go back to the beginning of the script to restart the process, or escalate the call to a human worker. By using stochastic spread activation Amelia can handle dialogue variance with no issues; she’ll only need to go back to the point in the conversation at which she confirmed which card was being used, redo that interaction, and continue the process without having lost any of the information she retrieved during the tangent.

Remember business process modelling? Amelia is also trained to handle the process while accounting for human variance. If the customer goes off script, and asks a question that has nothing to do with business processes, Amelia can adapt. So for example, if Amelia asks "Would you like me to mail you a new credit card?" And the person responds, "Is it going to cost me anything?" Amelia is able to take into account the context of the interaction, determine that “it” is shipping a new card, know that no cost is associated with that service, and inform the customer accordingly.

AI Systems, Emotion, and Speed

Amelia isn’t trained to respond in robotic, all-business phrases. While she’s processing things like intent and importance, she’s also figuring out the most appropriate emotional response. If a customer says, I lost my credit card and I see fraudulent charges, Amelia doesn’t respond by saying, “That’s great. Let’s get started.” She’ll respond by saying, “That’s terrible. Where did you last use it?” She can also use social talk to help make conversations seem more natural. Instead of saying, “That’s terrible,” she can be programmed to say, “That stinks.”

In order to deliver faster and more intelligent service, Amelia uses episodic memory of past interactions to figure out how best to resolve an issue. These memories don’t even have to be her own; she can read a transcript of a chat between a human agent and a customer about a previously lost credit card to determine the appropriate response. If the previous conversation implied that the customer preferred picking up the credit card from a retail location, rather than having it mailed, Amelia will use that knowledge to make that recommendation.

This is a very short introduction into how Amelia works as an enterprise-ready AI system. Behind every interaction that Amelia is trained to resolve, there are hours spent crafting her verbal and tactical responses to ensure extremely high resolution rates. These are complex processes that require trial and error and lots of brain-power. But once Amelia is implemented, she makes life easier for customers and employees, and she constantly improves her performance.


Previous Next

The Intelligent Contact Center

Companies have spent decades implementing Interactive Voice Response (IVR) systems in their call and customer care centers, but they've proven unable to keep up with customers' expectations.

In this white paper, we examine the benefits of an Intelligent Contact Center, where companies utilize Conversational AI-powered virtual agents to provide first-line resolution and support for customers, and augment human employees through AI and automation.

Download our paper to learn the benefits of this approach and why current IVR systems simply will not cut it in today’s hyper-paced digital landscape.

Learn More