How to Optimize the Voice AI Experience
See more videos from CX Connect 2021 on the CRM YouTube channel.
Read the complete transcript of this clip:
Meghna Suresh: It's a lot besides just the AI that can make the difference between success and failure. For example, AI algorithms don't just get smarter over time. By default, it's necessary to build out the plumbing and the pipeline to extract error cases, to relabel the data retrain models. And this is something that we've packaged up into our continuous learning product, but it's really important to think about any time you design an AI system like this one. Make sure you're actually putting that investment into capturing the error cases and programmatically retraining your models. The other area where we've made really deep investments is in analytics, and even beyond that, in A/B testing. It's really important, when you think about a system like this, that it's not a black box, and this is one of the fundamental design principles that we operated with from the beginning.
We don't want to build a black box system. It has to be highly explainable. Every decision that our thinking machine makes needs to be explainable. We need to be able to dig into it to understand why that decision was made. And so we invest a lot in the analytics instrumentation that that helps us understand what's going on under the hood. And we've also built out a lot of UI that allows you to go into every conversation, every transcript, and really dig deep. Our system allows you to run A/B tests and you can say, "Well, which version of the greeting worked better in that particular call flow?" And it's really important to set the system up in this way.
Lastly, one of the things that we've thought about really long and hard in terms of providing that fast time to value is, what are the best ways to scale some of these best practices that we come upon? As we do more A/B testing, we find a certain type of greeting works really well, or this is the most effective way to get a customer to engage before we escalate them through an agent. Is there a way to take all those learnings and wrap them up in such a way that the next time we have to make that decision, that learning is carried forward. That's where we really think a lot about pre-built conversational components, which in the Replicant context, we call Powers. But the way that we think about this is, in order to scale conversational AI, you have to really think hard about conversational design.
In this next slide here, we'll talk about how we package up conversation design such that it's not something that we do on a case-by-case basis for every thinking machine, but we're actually able to scale our learnings.
Most importantly what we do when we think about conversation design, is we think really deeply about the user context, the motivations, the pain points in every type of call flow. And we really attempt to meet the caller where they are. We then abstract all of these learnings and these pre-built components, and allow those to be carried forward.
These are just a few examples of what's possible. This just represents a small selection of what's available in our component library, but we've got pre-built call flows, essentially, for a lot of the most common customer use cases. And this is a really easy and fast way to scale and bring about that fast time to value. Take, for instance, something like data collection. We've thought about the perfect way to collect an address. So our address collection entity extractor can do a lot of things. If we're talking on the line with a caller, we need to capture their address either to authenticate them or to geolocate them. And let's say that user gives us a pretty fuzzy input, like cross streets, or they say an address, but it's actually not the complete form of that address. There's a lot of intelligence that we've baked in to interpret those vague inputs like cross streets or to match against a known address database to figure out what's the address that that user is actually trying to talk about, and is it valid?