ASAPP’s business is centered on using technology to enable agents to efficiently and effectively serve customer needs. To do so, we need to have a system that understands (and is always learning) what it takes to solve a customer’s needs, such that we can predict what an agent should say and do.
So the kernel of what we are trying to do is to deeply understand how a customer issue is resolved. We hear so much in the press about companies using AI to solve customer issues through chatbots…but when you inspect what they are actually doing, they only account for half of what is needed to truly understand how an issue was solved. That’s because knowing what was said needs to be paired with knowing what was done. And that’s where the vast majority of these technologies fail.
Conversational AI — natural language technology that understands what a customer is saying and can respond back — is not enough to create an efficient problem-solving engine. If we only look at the conversation, we miss critical information and don’t get the whole picture. It’s essential to know the entire state of the problem, the actions that an agent takes, indications of which actions contribute to the solution, and to have a machine learning method to improve, adapt, and inform future interactions.
Thus, an AI system that truly helps solve a customer problem needs:
- Natural language processing to understand a customer’s words and intentions.
- Action-monitoring technology to have a full picture of what agents are doing outside of the conversation.
- Accurate data signals, analytics, and metrics that indicate if the system generated predictions of words and actions helped resolve an issue — or not.
- Feedback-powered machine learning algorithms to 1) predict and recommend the next thing an agent should say or do during each interaction and 2) optimize the system accuracy to positively increase key metrics.
Each of these parts contribute to a customer service solution that far exceeds what can be achieved with Conversational AI alone. It is critical that each of these components work together to predict the best solution. Without coordination, the conversation may at first glance appear intelligent but may not actually lead to a positive outcome that helps a customer. But when the system learns from agents’ behaviors — for example the action they take with a knowledge base, CRM, or other backend system — we can reinforce all components that gave a correct answer, and teach the components that gave a wrong answer to do better in the future.
It’s essential to know the entire state of the problem, the actions that an agent takes, indications of which actions contribute to the solution, and to have a machine learning method to improve, adapt, and inform future interactions.
Aaron Isaksen, PhD
This approach gives a fuller view of all the information available, which is far more than what is said or typed, and enables a company to truly optimize performance. A system that uses this approach learns how to map language to actions, which helps build models to predict how an agent should respond — in words and actions — to best address the customer’s issue. If the environment changes, the system observes when its predictions are no longer effective, and automatically learns and adapts to those changes.
Ultimately, the goal is not just to understand the customer, but to use technology to most efficiently and effectively help them. Conversational AI is not enough. We must also understand what actions are necessary to solve the problem, execute the correct actions, and learn to adapt and improve the entire system. These components, working together as an entire system, are the path to building AI systems that optimally solve customer problems.
Aaron Isaksen, PhD is Senior Director of Machine Learning Engineering and Technical Site Lead for the ASAPP NYC office. His career spans natural language understanding, embedded artificial intelligence, interactive UX, and extensive startup leadership. He has engineering degrees from UC Berkeley, MIT, and NYU and his doctoral research focused on using AI to optimize for human experience in interactive software systems.