It was an intimate conference featuring leaders and decision makers from enterprise brands and ecosystem players including Accenture, Audiocodes, AWS, AstraZeneca, Booking.com, CM, Conversation Design Institute, Deutsche Telecom, ING, Nestle, Novartis, Rabobank, SwissCom, Vodafone, and more.
The biggest takeaway was how far the conversational AI space has progressed — how advanced the enterprises are in pursuing conversational AI. We are well past the days of basic FAQ chatbots.
The presenters discussed more advanced solutions for better customer experiences, as well as the significant resources and efforts put into developing high quality solutions — building out teams, optimizing models, labeling data, designing for modalities, handling nuances in language and dialect, and more.
Tobi the chatbot
One of the most interesting presentations was by Vodafone.
Vodafone has over 400 people working on their Tobi chatbot — spread all around the world in smaller teams of ~12 people.
Tobi is an omni-channel “super assistant” — it not only handles customer service related to Vodafone, but users can also shop for groceries, play games, and more. There are over 1,000 different journeys. The goal is to be part of a customer’s daily life — Tobi is “everywhere for everyone.”
On the customer service side, Tobi handles 90% of issues, and contact-center calls have decreased 25%. The goal is to reach 99% containment and 85% resolution by 2025.
Vodafone’s journey started with FAQ chatbots, moved to natural language understanding (NLU) models, and is moving more to conversational AI experiences that are intuitive, can predict intent, and be proactive.
Trust is a key factor. The team wants Tobi to be an assistant users can trust will get the job done. As part of that, it is important to be credible and reliable. For example, only being on pages where it knows how to answer, and having a person be present if an escalation does occur.
Nuances of conversational AI design
Effective Conversational AI design requires domain expertise. While it is not that hard to build a basic chatbot, it can be challenging to build a high quality, highly effective experience that understands the user, responds appropriately, and responds in a way that satisfies the user — especially across multiple modalities and languages.
The presenters dove deeper into these concepts, providing best practices for optimizing NLU and automatic speech recognition (ASR) models, as well as conversational dialog design, leveraging data, and analytics.
My own presentation was focused on the best practices in conversational AI, based on the experience of processing 90 billion messages across a wide variety of chatbot and voice assistant use cases.
The power of words
There is an art to conversational AI. Dialog design is a cross team, collaborative process — combining linguistic, business, and technical skills. It is important to be concise, build trust, and be empathetic. How a question is asked, can influence the answers.
Personalization is important
Leveraging context and personalization are key. The more information you know about the user, the less the conversational AI has to ask. The team at Swisscom gave great examples of this. For example, if they know a subscriber only has a landline and not a mobile, the interactive voice response (IVR) system will not suggest the option to cancel a mobile line.
A benefit of conversational AI is that the user does not have to adapt their way of thinking to how a business process is done. The Swisscom team also described how older IVR systems fail because they apply internal processes to their customers. For example, if a customer calls and indicates their Internet does not work, an old IVR system may treat the call as a tech issue and transfer the customer to the tech department. However, if one incorporated personalization and context, it may have seen the customer has not paid their bill and that is why the Internet was shut off, and thus forward them to billing.
While personalization is a good thing, it is also important to be mindful of privacy laws and use the data only for the intended purposes.
Leverage data and iterate
The presenters discussed their conversational AI journeys to improve the quality and effectiveness of their solutions — leveraging data to make improvements.
A common issue is that out-of-the-box ASR and NLU services may need tweaking for the use case or region. The presenters discussed capturing portions of voice calls, transcribing the data, analyzing for accuracy, and iterating. Data labeling is an important part of this process. With voice in particular, there are regional dialects and accents that an off-the-shelf solution may not be initially set up to handle, so the models need to be enhanced. For example, with SwissCom, one-third of callers will speak in Swiss German, versus High German, so the models need to be customized.
Similarly, with text-based communication, one needs to look at the analytics, including mishandled and unhandled Intents, to iterate on the NLU models. Some of the presenters even use more than one NLU provider, depending on the region.
Overall, the World Chatbot Summit was a great conference. It was nice to be able to meet with and learn from industry experts in the space. I am excited to see how advanced the efforts by enterprises are in conversational AI — to truly deliver high quality, highly effective experiences.
Arte Merritt leads Conversational AI partner initiatives at AWS. He is a frequent author and speaker on conversational AI and data insights. He was the founder and CEO of the leading analytics platform for conversational AI, leading the company to 20,000 customers, 90B messages processed, and multiple acquisition offers. Arte is an MIT alum.