Insight of the Week: 3 Insights From 3 Acquisitions in Conversational AI
In the last year, we've acquired three conversational AI companies. It's exciting to integrate them into a single practice. But it also provides unique insight into the way different teams approach similar problems.
I'm continuing to learn a lot, but already three key insights have emerged about how to build Conversational AI. In the real world.
1) Drag and drop makes for a nice demo. But not a scalable deployment
Most Conversational AI platforms have some kind of 'drag and drop' design / development environment. At the very least, there's going to be a website through which you configure your IVR or chatbots. These 'no-code' or 'low-code' environments make for a great demo. But every one of the companies we acquired ultimately found that these environments were constraining. So they found a way to bust out of the limitations of no-code / low-code and build an application with real code.?
The most interesting architectures, to me, are those that are somewhat 'database driven'... so the decisions about what to do next are driven by a table that lists a whole bunch of possible scenarios and what the IVR or chatbot should do next. The cool thing about this approach is that it's platform agnostic and scales easily to multiple languages and channels. It's early days, but I'm quite convinced that this is the way to build Conversational AI. And ultimately, to allow data - and eventually AI - to drive decisions about how to respond to customer requests. Because right now, AI is mostly limited to recognizing and understanding what customers say. Not deciding how to respond.
2) You can't just put in a few example phrases and expect the AI to do the rest
Modern Conversational AI platforms leverage huge background 'language models' developed by various commercial and academic research groups around the world. I talked about it in my previous email on the?mystery and magic of embeddings.
The idea is that these models capture the structure and meaning of language, and all we need to do is provide a few examples of what customers might say, and the AI will be able to understand all the variations. It works remarkably well. Until it doesn't. To get the most out of these platforms, you can't just blindly put in the example phrases and hope for the best. You've got to collect examples of what real customers say. And you need to identify where those examples are ambiguous. And where they might be unambiguous to a human listener but are easily confused by the AI models. All three of the companies we acquired found they needed to do a lot more legwork on the new AI platforms to get the most out of them. Often, it made sense to mix and match older, more constrained AI tech with newer approaches. Interestingly, this 'mix and match' approach is more easily supported when you have code, or ideally a database, driving the interaction, as I mentioned above.
3) Conversational AI platforms lack the reporting to drive long-term success and ROI
Reporting in Conversational AI platforms is universally bad. You can get at the data, one way or another, but to make sense of it is tough. All three companies found they had to build their logging and reporting tech to figure out how the IVR and chatbots were performing. Each of them did it differently. All of them would probably do it a bit differently in the future, given the more flexible analytics and reporting tools available now. Interestingly, again, the reporting becomes easier and more standardized when you have a code or database-driven interaction like I talked about in point 1. Why’s reporting so important? It’s the key to?managing your bot workforce.
领英推荐
So. You can't drag and drop your way to a great Customer Experience. It's OK to use a web console to explore platform capabilities and prototyping. But to build enterprise-scale applications, you need enterprise-scale software. That's not built-in a GUI. But it can have a GUI to configure it on the fly if that's important (and it often is).
You can't just put a few example phrases in and let the AI do the rest. You need to look deeply into how the AI succeeds and fails and work around that. And you need to build reporting.
The key point, I think, is that Conversational AI tech is immature. And constantly evolving. The best platform today won't be the best one tomorrow. And no one platform does everything great.
These things make success in Conversational AI harder. But the juice?is worth?the squeeze.?
It's not getting any easier to hire and retain agents. And customer expectations are going through the roof.
The cost of a bad experience has never been higher, but the cost of delivering a great experience hasn't gone down.
Conversational AI promises the ultimate balance of human agents and automation. But it's an immature tech. You need to make smart decisions about how to get the best out of it and avoid getting stuck with a platform that made for a good demo but just doesn't help you build and scale your?Bot Workforce.
Kerry
Kerry Robinson?is an Oxford physicist with a Master's in Artificial Intelligence. Kerry is a technologist, scientist, and lover of data with over 20 years of experience in conversational AI. He combines business, customer experience, and technical expertise to deliver IVR, voice, and chatbot strategy and keep Waterfield Tech buzzing.