Learn from an unhelpful chatbot and build a better customer support system for your business

Uber Sri Lanka is in the news again, this time for turning a blind eye to the issue of an errant driver assaulting a customer over an innocent question about a cancelled trip. To make things worse, Uber Sri Lanka has done nothing to rectify the situation after the passengers took to social media, leaving its chatbots to spew pre-programmed messages instead. Understandably, this has infuriated many, and questions are being raised about the legality of Uber’s operation in Sri Lanka, where there is no formal channel for unsatisfied customers to raise their grievances.

Sri Lankan users have been voicing complaints regarding Uber, but to no avail/ Source: Twitter

Customer care has been hailed as an area which chatbots can excel, rendering customer support executives a thing of the past. However, as is visible with the Uber Sri Lanka debacle, a chatbot can sometimes do more harm than good.

In such a backdrop, what should you be aware of if you’re thinking about deploying a chatbot to solve your business problems?

Understand The Different Types of Chatbots

Chatbots can be categorised into two primary categories: ‘Transactional Chatbots’ and ‘Knowledge Chatbots. Before you deploy a chatbot in your business, make sure you understand what each type of chatbot can and can’t do. Deploying a transactional chatbot in a place which needs a knowledge chatbot is a recipe for disaster.

Transactional Chatbots are built and optimised to execute a limited amount of specialised processes which eliminate the need to talk to an expert or use more complicated UIs such as mobile apps or websites. In contrast, knowledge chatbot support thousands of processes and in some cases, can even make decisions for you.

Transactional chatbots are trained on top of structured data and can do a set of limited operations. Think of what a bank operator can do for you over the phone: verify your identity, block your stolen credit card, give you the working hours of nearby branches and confirm an outgoing transfer. The exact same functionality can be imparted to a transactional chatbot, which can then be deployed via Facebook Messenger, for instance.

On the other hand, knowledge bots are helping you both make a decision and execute it. In order to be able to “make a decision” the chatbot is usually trained with a vast amount of unstructured and structured data, and is trying to produce a response as an expert. Though not exactly bots, IBM Watson (and Deep Blue before it) are good examples of knowledge systems.

A matrix showing the differences between knowledge chatbots and transactional chatbots / Source: progress.com

Simple, But Not Stupid

For a chatbot to truly deliver value, it must simultaneously be simple to use, while not being too simplistic a.k.a. ‘stupid’ to turn away and/or frustrate users. This is a problem Uber Sri Lanka has to deal with. Remember that a chatbot is not sentient, and is only as good as its bank of pre-programmed scenarios. When a query doesn’t match a pre-programmed scenario, the bot gets stuck in a loop and users get frustrated.

When a chatbot gets stuck in a loop, user frustration is not too far away / Credits: progress.com

A casual scroll on Twitter is enough to get a sense of how quickly the average Sri Lankan Uber rider is frustrated by the misplaced efforts of the chatbot.

Disgruntled Uber users are not hard to find in Sri Lanka / Source: Twitter

This is where sophisticated Natural Language Processing systems come in. Microsoft, Amazon, and Facebook already provide state of the art Natural Language Processing (NLP) developer tools to help your bot understand user intents, so whenever you deploy a chatbot, make sure the chatbot is built to benefit from these powerful NLP engines.

Know What You Will Pay

Nobody likes hefty bills of eye-popping amounts. It pays (in this case, literally) to understand how your chatbot vendor will charge you for your chatbot.

Generally, enterprise-level chatbot platforms use one of the following pricing models:

Subscription: Build limited or unlimited bots for a flat monthly or annual fee. This is probably the most suitable model for an enterprise.

Pay per usage: The platform cost is based on chatbot usage and the number of API calls. This model can provide you with a higher degree of flexibility allowing you to scale gradually.

Pay based on performance: The platform cost is based on how the chatbot performs against goals agreed between the customer and vendor. For example, a successful interaction between the bot and the client, which leads to a signup, is considered as a paid conversation, regardless of the length of the chat. Under this model, enterprises will be charged accordingly to the goals achieved.

Don’t Underestimate Humans

Chatbots may be better and more efficient at a range of things, but they cannot beat a human at one thing: displaying empathy. Therefore, depending on the function of a chatbot, it becomes important to ensure that a human is available to step in and takes over a user interaction if things get out of hand. After all, even the most experienced support agent often requires a second opinion and your chatbot is no different. That is why it is recommended to follow a bot-human hybrid model, which allows a human to takeover an interaction depending on the complexity and criticality of the issue.

To understand this issue in detail, take a look at the two screenshots below. In both cases, the triggering keyword is “server failure”. In the first case, the user simply wants information about the precautionary measures in case of a server failure, while in the second case the user wants to report the occurrence of a server failure.

Same keywords, but different questions entirely. How the chatbots responds is important /Credits: Kommunicate.io

What are your views on chatbots? Who has deployed chatbots well and who hasn’t? Let us know your thoughts below!

Cover image credits: COPREUS