Updated: Oct 13, 2020
How do you design a chatbot conversation that works for humans?
Kaveer Beharee, CEO at Ubiquity AI, says that developing a successful chatbot has less to do with AI or technology and more to do with creating an authentic conversational experience. In his third insight piece on creating better #AI chatbots, he outlines his own chatbot design strategies and the three main considerations that will help you move towards better machine-human conversations.
Artificial intelligence-based chatbots are becoming increasingly popular with companies finding new commercial applications every day. Some #airlines use chatbots to help you manage or change flight bookings while others help you choose the best combinations of toppings on your pizza order.
I don’t think that we’ve scratched the surface yet in terms of understanding how chatbots can change business and society, but I do know that the success and adoption of chatbots has less to do with the underlying technology and more to do with how chatbots are able to engage with people to create authentic experiences.
Think about chatbots being used by doctors to manage patient aftercare; schools using chatbots to help children who need extra tuition. Chatbots can significantly reduce the cost of engagement while improving the probability of successful outcomes.
The potential of chatbots was recognised more than half a century ago. The first chatbot #Eliza was developed in 1966 and was specifically designed to engage in goal-directed behaviour. The intention was good but the execution was woeful. Over the past few decades, chatbots were developed using rule-based (decision-trees) to best match a user conversational input with a pre-programmed response. The conversion flow was horrible and frustrating – like many chatbots today – as the chatbot response required a user to ask questions in exactly the same way the chatbot was designed to recognise.
Today, most chatbots still have the same problem but for very different reasons. As opposed to rule-based chatbots, the ‘artificial intelligence’ in chatbots is primarily used to understand user intent. In theory, this means that chatbots should be able to understand inputs and communicate significantly better than the old rule-based bots – but half a century later chatbots still don’t.
How do you create a chatbot conversation that works?
It is important to point out that many of the chatbot success stories reported by companies, including the examples mentioned above, do not require much user conversational input and are primarily designed to automate business processes.
I’ve posted previously on why chatbots are often failing to live up to their potential and why so many are so bad. For chatbots to become mainstream, I believe that the following four requirements need to be met:
Chatbots must be capable of understanding the diverse ways people communicate – the experience should be built around the person, not the chatbot.
Although we’re far away from this, chatbots should be able to pass the Turing Test (which means it must be able to pass for human to the user) – a person should feel like they are speaking to a person on the other side, not a dumb machine.
Chatbots must assume responsibility for being the single point of contact from customer initiation all the way to resolution, even if the chatbot escalates a query to the human. This means chatbots must be capable of closing the conversation loop 100% of the time. Is there anything more frustrating than a bot referring you to a call centre number after 10 attempts to try and explain a problem?
Chatbots must have well-thought out design parameters and capabilities that are well-understood by users to create a great user experience.
How do you make a chatbot conversation flow?
It may seem counterintuitive but these four elements are not primarily technology challenges. Rather they are conversation design challenges: namely how will your chatbot take a customer through a journey that is pleasant and maybe even fun, purposeful and with a clear destination in mind.
AI chatbots work in the same way as rule-based chatbots. This means that the AI chatbot will try to match a user input with an intent file in its system and respond accordingly. So, the key is building a chatbot that encourages authentic communications with the user but not to the extent that it overwhelms the system.
Let’s look at three key considerations for designing a compelling chatbot conversation architecture that will flow.
Three steps to better chatbot conversation design
1. Understand chatbot technology dimensions and begin with the end-goal in mind
The three technology dimensions for chatbots are:
correctly understanding the context of the question or request as a means of identifying the user’s intent. People intuitively understand context (most of the time); AI will not (until sometime in the future where generative systems will be capable of generating responses from scratch)
retrieving the correct answer that fulfils that intent from the most appropriate knowledge base or source of information
conversational interaction – critical for ease of use, general user experience and platform retention.
To plan for each dimension appropriately, begin visualising the chatbot’s desired outcome, ie, define the purpose of your chatbot by clearly defining the chatbot’s goals. By defining what the desired outcome from a customer’s perspective looks like, it becomes much easier to work backwards to how and when the conversation is likely to be triggered and what the conversation structure (and construction) will look like to reach the desired outcome/s.
For commercial applications, your chatbot will most likely be a closed-domain system where your AI chatbot’s knowledge base is built upon a knowledge management system for the purpose it is being built. Alternatively, if you are building, for example, a chatbot translator then you would most likely build an open-domain chatbot that accesses multiple sources of information. The latter is very rarely viable in a commercial application.
2. Understand the role of taxonomies in chatbot design
#Taxonomy (from Greek ‘taxis’, meaning arrangement or division and ‘nomos’ meaning law) is the science of classification according to a pre-determined system with the resulting catalogue used to provide a conceptual framework for discussion, analysis or information retrieval (source: DPCI).
There are many chatbot technologies out there but the most viable options are taxonomy-powered chatbots that are much better at understanding and standardising user input.
Broadly, there are two kinds of taxonomies that are critical for a smooth-functioning chatbot:
domain knowledge – such as background knowledge, industry entities, industry/relevant consumer topics, disambiguation of user inputs
structured knowledge – such as customer details, question types and question topics.
Many off-the-shelf chatbot technologies come with basic operational taxonomies prebuilt, but a fair amount of work is still required to design conversation branches built upon both domain and structured knowledge that addresses the customer’s query.
I’m sure that there are more sophisticated chatbot conversation design tools out there but I’m old school and use draw.io to map out a chatbot conversation flow design. In the chatbot conversation flow template, I identify associated taxonomies and taxonomy nodes, ontologies and knowledge for each step of each anticipated conversation all the way to resolution.
3. Designing a chatbot conversation – getting the questions right
Chatbots are really good at answering straightforward questions such as: ‘What is my outstanding balance?’ But they are poor in addressing complex questions such as: ‘If I settle my outstanding credit card balance this month, can I apply for a second card?’
Customers demand clear and accurate answers to queries, which means a chatbot’s overall success or failure will ultimately be determined by its ability to improve the relevance and accuracy of its responses.
Therefore we need to employ some strategies for getting the right answer:
#Disambiguation and follow-up – if a chatbot doesn’t understand, it needs to further interact with the customer and ask clarifying questions.
Topic taxonomy – can be used to index known questions and their answers, which will enable the chatbot to reformulate an original question in different ways and make it easier to find a correct answer. The best approach in this regard is leveraging knowledge across the organisation including recorded customer complaints, call centre operations, interviews with sales and customer service representatives, and so on. In my experience as an AI chatbot designer in the fintech space, short-term limited live pilots are the most viable way to understand how customers are likely to interact and engage with your chatbot.
Set use parameters that customers understand – develop and communicate to customers clear and accurate criteria and parameters for determining the type of questions that can be addressed (number 4 in my requirements to enable chatbots to go mainstream). There is definitely a trade-off between the extent to which you enable natural free-flowing conversation with your customer on the chatbot platform and the design parameters that will enable your chatbot to understand and respond to the customer’s intent. In this regard there are several ingenious conversation design strategies – mostly revolving around the chatbot pre-empting, prompting and leading a customer in articulating a query, or simply offering a customer query or service category tabs – which may disrupt the flow of the conversation but nonetheless may be crucial for the chatbot to fulfil its task successfully.
…remember that the purpose of chatbot conversation design is establishing the questions that trigger conversations and processes that lead to resolution.
Keep it simple, keep the scope narrow and manage user expectations upfront.
This blog was first published on Firehead's website. Firehead is the the go-to place for European recruitment, consulting and training in AI and digital communications.