Since, we hear a lot of people, especially bankers, believe that chat is the new universal UI (like we once passionately believed), we thought it might be useful to talk about our experiences and learning.
Back in early 2015, we thought that chat would become a universal UI because people were really using chat like crazy on their mobiles for P2P messaging. We thought we could build a personal assistant that helps them get things done over chat.
We built our chat process team over a period of 3 months from March to June 2015. By June 2015 – we were probably the biggest C2B (consumer to business) chat apps in the world. Probably bigger than all our Indian competitors and American counterparts like Operator and Magic combined. We were doing more than 70,000 chat sessions a day (not messages, chat sessions). The idea was that the manual chat would provide enough training data for building out a great AI bot.
Initially, we had a great response – customers hadn’t seen an experience like this before and loved the novelty.
Learning on chat
The only problem with chat was that customers who tried it weren’t coming back. So we tried harder, optimised first response times, average response times, our knowledge base, canned responses and built better dashboards for monitoring all of these (all the while scaling the backend for the chat volume). That still didn’t work.
We built AI-based chat bots and also experimented with re-imagining chat (make it into a series of simple clicks with graphic UI elements) but the results were not very encouraging. We thought that at some point users would get trained to use chat in the right way. It now seems foolhardy to even think like that but we really thought that chat UI will truly work, if only we worked hard enough. Here are the top 3 reasons why chat doesn’t for transnational use cases:
- The empty text box is not at all intuitive for such use cases like P2P chat where the user has a high degree of clarity and certainty on what he wants/needs to type
- AI never gives an exact answer to complex questions, there are always multiple answers ranked in order of importance. But with a chat UI providing multiple options becomes a challenge and in many cases makes chat redundant and painful. Imagine if you are allowed to use only the “I’m feeling lucky option” on Google Search (one of the world’s most powerful AI/ML implementations)
- Chat is unidirectional i.e you mostly keep moving forward while chatting but transactions require back and forth, check multiple options, filter results, compare products etc. making it very difficult to do it on chat
The hard truth (for us)
Finally somewhere in Sept 2015 it became obvious that the problem was chat.The customer would try it once out of curiosity but never come back. Other ways (simple UI) was just way faster and on the internet people will ALWAYS choose what is faster.
People often confuse the debate with of AI/ML and chat bots simply because most companies (Fintechs) have been falsely peddling one as the other. We will talk about REAL usecases of AI/ML in banking sometime later in another post. But does it really make sense to chat when you can simply tap once and book services (no matter what you think, our user data was very obvious on this. The answer is no.
This Dan Grover (WeChat product manager) article nailed it and before that it was Connie Chan’s (A16Z partner) tweetstorm. Very recently Facebook also announced that it would be shutting down Facebook M – it’s chat based personal assistant.
Chat or chatbots may not be the panacea that everyone touts it to be. It’ll always be simpler to just tap and get things done. For some things like customer service or advisory services, a chat UI might still work well. The key point is that bankers must not get swayed by press releases etc. but look at the actual impact of implementing such a thing because when chat fails, it is a waste of money and leaves you with dissatisfied customers.