How to Navigate the Bot and Conversational AI Maze

Satya Nadella, CEO of Microsoft, recently said, “we’ve seen two years’ worth of digital transformation in two months.” Accelerated digitization is taking place in all industries, and it’s triggering an influx of customer service inquiries—either for assisting customers along their new digital journeys or for handling exceptions. For contact centers already wrestling with capacity issues because of COVID-imposed remote operations, it’s the perfect storm. All are turning to automation and artificial intelligence (AI) to offload agents.

 

Thankfully, conversational AI and bots technologies have matured a lot and provide many options. The bad news is that there are over 200 vendors in overlapping, loosely defined categories. Additionally, the industry is collectively guilty of under-representing the amount of effort required to deploy and scale these solutions. It pushed conversational AI to the peak of inflated expectations before the pandemic burst. For practitioners, exploring options often feels like navigating a maze. Below are three lenses to help make better sense of the space.

 

But first, we all like to start our explorations by doing our own research, and picking a category is the logical first step. Alas, several cover the market with no agreed-upon definitions and a lot of overlap. The market started with virtual customer assistants (VCAs) and virtual agents (VAs), geared at customer service but never settled on any of these words. Chatbots, which initially were used to refer to simpler solutions for chat, became used as well for VCAs and VAs. It’s now frequently shortened to” bot,” a word that originally referred to scripted automation.

 

When applications expanded into marketing, sales, and Ecommerce, three new categories emerged: conversational marketing, conversational commerce, and conversational Sales. The differences between them have remained slim, with most vendors addressing the three. To add to the confusion, a few players have adopted conversational commerce as the uber-category name, while Forrester and Gartner coined the term “conversational AI.” Do you follow me?

 

The first dimension to explore is the technology packaging, either as a platform or an application. Many vendors love the word “platform” and feel compelled to use it regardless of the actual software abilities. Understanding if an offer is a platform or an application requires a bit of digging on your end.

 

Platforms bring great flexibility and might appear the safest way to future-proof your investment. It’s not the panacea, though. Versatility comes at the cost of requiring more substantial investment, in particular in terms of training. Applications come pre-packaged for specific use cases, allowing faster starts. Selecting a platform involves evaluating the flexibility you need, such as being able to choose your Natural Language Processing (NLP) engine; or having access to a broad set of communication channels. It’s especially important to assess not just the application programming interface (APIs) but the associated tools, specifically, for instrumenting and testing applications.

 

Conversational applications must be configured and trained for specific intents. But “intent” is a misleading word. We construe its meaning from the English definition, connoting a broad set of intentions. However, in the world of conversational AI, these are very narrow and represent each a specific question that can get mapped to a transaction in the back-end. For the application to be meaningful, you must train it on many intents. Each requires capturing different ways someone can ask a question, which involves a lot of training and tuning. It’s a much larger effort than most people think, and often an unpleasant surprise for practitioners. In many industries, it entails teaching the application a specific terminology. Think of mortgage or healthcare. Thankfully, many vendors have been packaging generic “intents.” I recommend looking for providers that cover your application domains. You must assess how industry-specific your use cases are and orient your research accordingly.

 

This landscape includes an immense spectrum of solutions, from very simple to highly sophisticated. It’s critical to target ones that commensurate with the complexity of the problem you’re tackling. Also, conversational AI technology is best deployed one step at a time. You will want to think through its adoption in terms of a journey, learning to walk before running. For that, I like to use the following framework for assessing needs as well as calibrating the type of software to consider:

1. The first level doesn’t leverage AI and relies instead on scripted logic. It’s the original meaning of the word “bot.”
2. The second level uses unsupervised learning to scan FAQ and knowledge bases. It surfaces the best responses to questions without manual training and requires a solid database of responses. Because it uses unsupervised learning, the effort is minimal. It typically offers the top three answers, letting the customer choose one or get transferred to an agent. This level is sometimes called an answer bot.
3. The third level is the first one involving supervised learning and focuses on tagging questions for triaging rather than trying to answer them. It requires a small set of variations per question and can be deployed quickly. These capabilities are growingly getting incorporated into chat or digital customer service solutions.
4. The fourth level addresses single-intent questions. It requires more learning, which is where most of the industry stands today, with different levels of breadth.
5. The fifth level includes intent switching. The AI engine can detect a different intent in the course of the conversation and pivot to the new context. This level is being spearheaded by the most advanced providers.
6. The last level, multi-intent, is the next frontier for the industry.