Customer Service Week Musings: How does a machine know if it’s wrong?
By Laura Ludmany, Knowledgebase Engineer
There are many comparisons dealing with the main differences between humans and machines. One of the recurring points is while humans have consciousness and morals, machines can only know what they are programmed to, hence they are not able to distinguish right from wrong unless they are provided data to make decisions based on that information. There have been many discussions on the self-awareness of robots, which is a topic as old as Artificial Intelligence, starting from Isaac Asimov’s three laws of robotics, continuing to the Turing test and nowadays AI ethics organisations.
One thing is commonly agreed – bots need to be ‘taught’ morals, and to achieve this there could be two approaches, both having their advantages and disadvantages. The first one contains a loose set of rules, but plenty of space for flexibility; this system could always reply to questions. However, it could also result in many false positives cases and could go wrong on many levels. The other would mean more rules and a narrower approach. The system could answer a limited number of queries, however, with very few or non-false judgements.
What does this mean from the customer service and customer experience (CX) view and for virtual agents answering real time customer queries? If we narrow down our conditions, bots would deliver the right answers at most times. However, they could not recognise many simple questions, making users frustrated. The same can happen with the loose set of conditions: the assistant would easily deliver answers but could misinterpret inputs, resulting again in annoyance.
To solve this problem, we must use a hybrid approach: an AI tool can only be trained appropriately with real-life user inputs. While we can add our well-established set of rules based on previous data and set a vague network of conditions, the bot will learn day-by-day by discovering new ways of referring to the same products or queries through user interactions. Half of a virtual assistant’s strength is its database, containing these sets of rules. The other half lies within its analytics, which is an often-overlooked feature. What else could be better training for a CX tool than the customers leaving feedback at the very time an answer was delivered? Conversation surveys are not only important to measure the performance of the tool. They are also crucial for our virtual assistants to be able to learn what is wrong and what is right.
Our approach at Creative Virtual to reporting is to follow the trends of ever-changing user behaviour. We offer traditional surveys, which measure if a specific answer was classified as helpful or not by the user and if it saved a call. Sometimes, the specific required action or transaction cannot be performed through self-service options and the customer must make a call, or else, the answer has been slightly overlooked and needs to be updated – for these cases there is a designated comment section, so users can express themselves freely.
We all know from personal experience, that we can’t always be bothered to fill out long or detailed surveys – we are on the go and just want to find the information we were looking for without spending extra time to leave feedback. This is typical user behaviour, and for this we came up with different options for our clients such as star ratings and thumbs up and down, keeping the free text box, to make the rating simpler for users. The solutions deployed are always dependent on the requirements and preferences of our clients, which are in line with the nature of their business and their website design. For example, financial organisations usually go with the traditional options for their customer-facing self-service tools, but internal deployments often have more creative user feedback options.
What if, during a conversation, a virtual assistant delivered the correct answer to five questions, but two answers advised the user to call the customer contact centre and one answer was slightly outdated? Does this rate as an unsuccessful conversation, due to three unhelpful answers? To solve this dilemma, we have End of Conversation Surveys, which ask customers to rate the whole conversation, on a scale to 1-10 and choose what they would have had done without the virtual assistant. As always, there is a free text box for further communication from the customer to the organisation. These surveys show high satisfaction levels as they measure the overall success of the conversation, which can have some flaws (just as in human-to-human interactions), but still can be rated pleasant and helpful.
Let’s take a step further – how can the virtual assistant learn if it was right or wrong if none of these surveys have been taken up by the user? Is this valuable data lost? Our Creative (Virtual) analytics team have levelled up their game and came up with a solution! During voice interactions, such as incoming calls to customer contact centres, there is a straightforward way to understand if the conversation wasn’t successful, even if it wasn’t stated explicitly, as the tone might change or the same questions might be repeated. But how can we rate a written communication with our customer? There has been a specific platform developed, which sits on the top of our previously described survey layers. This platform classifies the whole conversation, with a carefully weighed several-factor-system, which can be tailored to our client’s needs, containing factors such as if there has been more than one transaction, whether the last customer input was recognised by the virtual assistant, if there have been negative user responses recorded, etc.. The primary ‘hard’ indicators remain the user-filled surveys, so this is just a nice icing on the cake, as our mature deployments show over 80% of successful conversation rates.
With our proactive approach and multi-layer analytics tool sets, we can be sure that our virtual assistants will learn more and more about what is right and wrong, to increase the customer satisfaction level continuously. However, I think no machine will ever be able to answer all questions correctly, as this would mean that deployments have stopped being fed up-to-date real-life data. Our world is changing rapidly as are our user queries. These cannot be fully predicted ahead, just analysed and reacted to appropriately. As long as AI tools serve customer queries, they will always face unknown questions, hence they will never stop learning and rewriting their existing set of rules.
As we celebrate Customer Service Week this year, we need to recognise the role customers play in helping to teach our AI-powered chatbots and virtual assistants right from wrong and the experts that know how to gather, analyse and incorporate that data to help train those tools. Check out our special buyer’s guide that explains why experience matters for using this hybrid approach to create reliable and always learning bots.