In a previous article “Human or Machine” I concluded that combining the simplicity of conversational interfaces – in the form of Messaging and AI powered chat bots – to drive understanding and prediction whilst also having humans in the loop to help with more complex tasks and make better informed decisions will be how enterprises drive step changes in both Customer Experience and Customer Experience Management (also known as the Colleague Experience).

In this piece I want to share some of the learnings I’ve been able to take away when designing and building conversational interfaces in the form of customer-centric chatbots using Messaging and AI.

Today most chatbots are rule based. This is because they are quick, easy and relative cost effective to deploy. Similar to how search engines work, keywords in the text are analysed and used to take the conversation flow down a pre-defined route. Where you have a visual UI (such as a website or mobile app) buttons can be used to help the customer complete a task in a couple of clicks. This is fine where there are a finite number of options for the customer (an old Apple UI guide for developers stressed no more than five to seven options was appropriate), but becomes confusing when the number of options is higher and the complexity is greater.

My team and I wanted to move from rule based chat bots to something that was AI powered, that could could help answer a wider range of questions, something that was more helpful – thus driving out more benefits for both the customer and the business. From our user research we uncovered a set of findings that we defined as our ‘Minimum Viable Threshold’ for Customer Experience with bots. These are the areas of customer experience that we felt would affect the perceived value and willingness of use for customers. It specifically outlines the minimum level of interaction requirements that users would expect during any form of interaction with an AI powered bot agent. For simplicity I’ve categorised the findings into 5 areas.

Understanding Intent:

• Your bot understands the question in multiple formats. This means having the ability to connect abstract data between what is being said and the intention of the customer.
• Your bot should be able to understand the question with spelling and grammatical errors.
• Your bot should be able to ask guiding, clarifying questions to determine the customers intent.

Consistent Answers:

Understanding context and the relationship between words and phrases is key to providing consistent, relevant answers. Just using keywords increases the risk of the bot misunderstanding the context and often results in a ‘confusion’ loop which leads to frustration, abandonment, lack of trust and a poor customer experience.

Relevance:

Your bot should be able to understand emotional sentiment such as urgency and anger in order to apply the appropriate treatment strategy, in most cases this would be a hand-off (by design) to a human agent for these emotions.

Conversation:

Customers feel more engaged and confident if the bot has appropriate social protocols such as greetings and closure (’thanks, goodbye’).

Customers want the tone of the conversation to be ‘human-like’ and therefore require the bot to have a high level understanding of natural language.

Failure Handling:

Customers want a safety net where they can be transferred to a human agent at any point. Best practice would be to integrate with a chat interface to provide live and/ or asynchronous chat.

Customers hate repeating themselves. So make sure you pass the context of the conversation from the bot to the human agent so they can quickly grasp what it is the customer is trying to achieve and are therefore better placed to help.

Devise a strategy for what happens when your bot doesn’t know the answer. Make a handover as natural as possible, remembering that the customers frustration of what went wrong will always be greater than the sense of delight when it went right.

Designing our bot by following the ‘Minimum Viable Threshold’ has enabled us to develop ‘human-like’ conversation flows that are achieving high levels of customer satisfaction. Using Natural Language Processing (NLP) enables a precise control of the flow of interaction through a single medium (i.e. speech or text) without needing any special affordances. And from every conversation it has with a customer, the bot can be trained to learn and improve its understanding. in summary, and to quote one of our customers…

“It has to be quick, accurate and intelligent. If not I wouldn’t bother”

In my next post I’ll look further at another piece of the jigsaw – the role that humans play and the importance of analysis and insight to drive good conversation design.

Terry Cordeiro is the Head of Product Management – Applied Science and Intelligent Products at 
Lloyds Banking Group and is a founding Steering Committee member of AITECHTalents.