Our AI Future: Tone Optimisation Will Help AI-Powered Technologies To Interact More Naturally With Humans
For AI based systems, such as chatbots, that interact with a company's clients the customer experience is very important - if people become frustrated they will opt for a different service provider. South African startup FinChatBot helps financial institutions implement systems that retain customers and increase customer satisfaction, and one of the tools in its arsenal is tone optimisation for natural-language processing.
This is the second in a series of articles highlighting interesting technologies and concepts that caught my attention at the inaugural AI Expo Africa, which took place in Cape Town last year. The first article covers disruption
(and volumetric holograms), this one looks at customer experience, the third looks at synergy
(and automation), and the fourth looks at new business models
(and the insurance industry).
At AI Expo Africa Antoine Paillusseau, a co-founder (with Romain Diaz) and the CEO of FinChatBot, which develops chatbots for financial institutions, discussed three points that are crucial for ensuring clear communication with customers: how the company works to improve customer experience by changing how chatbots are scripted; using different AI-powered communication methods (Facebook Messenger
, in app, WeChat
, SMS, and WhatsApp
) to communicate with its clients' customers because different generations want to communicate in different ways; and using natural-language processing (NLP) targetted at various age groups to communicate with people using a language style that they find more comfortable.
"We use AI and chatbots to improve the customer experience," Paillusseau said. "The real breakthrough for the industry is the chatbot. It's already a 24/7 tool. It's much more cost-effective than any human intervention."
FinChatBot likes to "define what can we solve and then build the conversation" in order to satisfy the customer in less than two minutes. "So instead of spending 14 minutes on the phone with a call-centre agent you spend two minutes chatting and you get insured, you get access to your account balance, you get a lot of services that you can imagine on the financial-service side," Paillusseau said.
In the presentation he emphasised that businesses need to "to identify what the most performing, impactful elements of their business are and then to draft the conversation from this". The best performing ones are usually quotations and sales.
He cited the example of a product the company developed for one of its (presumably insurance) clients in which, instead of asking a customer to input lots of information about his car, he is simply asked for the car's licence-plate number.
FinChatBot then contacts TransUnion on the back end, pulls up the make, model, and year faster than the customer would be able to input the information himself, and asks "Is it your car? That's what we have in our records. Is it your car?" The customer just has to confirm that it is.
"It's what we call a thumb-free type of conversation - you can do the whole conversation with one thumb," Paillusseau said. "It's extremely user-friendly, it's quick, it's fast, and you get to the service that you want in a very short amount of time."
Businesses can access the analytics from interactions to see what questions are working well and experiment with new questions to see what's more efficient, thereby improving the customer experience.
The company's most interesting AI experiments, however, have to do with natural-language processing and tailoring the chatbot experience to the preferences of the customer by adjusting the language that the chatbot uses when it talks to a person.
"With chatbots we hear a lot of discussion around NLP, NLU - natural-language processing, natural-language understanding - and it's a bit scary for clients because you need to have a huge amount of data - classified data - in order to make this work.
"One area that is very powerful is the sales process. We were quite surprised: the first interaction we have with our clients is using their call centre script to create a transcript for the chatbot experience. We were super surprised because they only have one script for whomever they talk to so, whether you are 25 or 60, it's the same script, the same type of questions.
"Consequently we [are developing] tone optimisation using AI. Our cue is being able to ask [questions in] different tones and trying to understand the best way to make [the customer] feel comfortable in order to sell better. Having one script for everyone is so against the very basic sales principle. You need to adapt your speech. You need to understand the needs of your client in order to sell better. And that's exactly what we are trying to do. Our goal is to say chatbots are maybe more human than the call-centre agent - it's a bit scary, Skynet is around the corner - but it's true. We can really try to optimise the conversation from an audience perspective, such as age and gender. Being able to target a woman, 25, from Cape Town and adapting the speech to 60-year-old man from Johannesburg, you're definitely not going to have the same experience but, in the end, you still have the same objective of selling a product."
Above: The Millennial approach uses casual language and emojis.
"To give you a quick example, here is a special tone that we implemented:
'Howzit Antoine, I'm Holly the cute robot who's going to help you get a quick and easy quote for your car insurance. Sound good?'
'Let's do it!'
'To get this baby going we need to share your ID number with us.'
We can't use it for everyone - we found out that the older you get the less emojis, memes, gifs you're willing to read in order to trust your financial-service provider."
Above: The Baby Boomer approach uses much more formal language and no emojis.
"Another one, much more formal:
'Good day sir. I'm Holly, your automated agent. How can I assist you?'
Etcetera, etcetera, same process. This one is for an older generation and it works for specific audiences. The reason that our best performing one is the Millennial tone is you can imagine that a lot of Millennials are on messaging apps and using chatbots and are [receptive to using] those chatbots because they know they can access information in a very short amount of time. We have an average time of five minutes. To give you a point of comparison with the call centre you need 40 minutes for the same experience so you save pretty much an episode of Game Of Thrones to get to the same objective. With the tone optimisation we get up to four times higher conversion rates. It's extremely, extremely high compared to any other medium and ways of interacting with an audience."
Of course, there are always anomalous customers so I asked Paillusseau whether FinChatBot is able to detect if someone doesn't like the informal tone, or vice versa, and wants something different.
"We are working on it," Paillusseau said. "We don't have this capability implemented yet but we are trying to trigger a data [point] that is the time needed to answer the question. If the time [taken to reply] is too long then we can swap the tone for a more mutual way of chatting and see how it goes. It's clearly one of the things we are looking at."
FinChatBot's approach to using AI both to solve problems and ask questions and present information in styles that are more accessible to more customers helps to make the experience more pleasant for customers, which increases their satisfaction and, consequently, the customer-retention rate when compared to the drop-off rate that call centres experience. It also reduces the costs of running call centres as a lot of the workload traditionally assigned to human agents can be offloaded onto the bot, which is able to complete the tasks successfully by keeping the customers properly engaged throughout the process.
The company's success has resulted in it securing a number of major banking and insurance clients that have quietly implemented a customised version of the bot - which is officially named "Holly" - for their communications systems and this has not gone unnoticed. FinChatBot, which is only three years old and has been incubated by Far Ventures
in Cape Town, secured, late last year, a reported
US$500 000 venture-capital investment (at the time around R7 million at the November 2018 exchange rate of R14.00/US$1) from Kalon Venture Partners
, which is based in Johannesburg, South Africa, and Compass Venture Capital
, which is based in Mauritius.
This month it was also nominated
for two Southern Africa Startup Awards
- Best Fintech Startup and Startup Of The Year - the winners of which will be announced at the end of November 2019.
For people to become more accepting of AI systems with which they have to communicate, which is going to become increasingly prevalent as businesses automate more of their customer-facing tasks, whether customers want it or not (generational gaps very much come into play here, although they are not the only consideration), the systems have to be approachable and users need to feel comfortable using them. Tone optimisation is one method that will help with this as has been demonstrated by FinChatBot's early successes - using technology that is still in its infancy.
Antoine Paillusseau: LinkedIn
Mandy J Watson was a media guest at the 2018 AI Expo Africa conference. The 2019 conference is being held in Cape Town at the Century City Conference Centre on 4 and 5 September 2019.