We used to believe that the future would have flying cars. Now, two decades into the 21st century, it turns out that the best the future had to offer was regular cars that are kind of like taxis except (here comes the selling point) you just happen to be able to hail them using … your phone.
Disappointing, I know.
Perhaps we will get over that disappointment eventually but the next one is already looming. It has to do with digital butlers — the kind Tony Stark uses in “Iron Man”, for example. Such films awakened an expectation in us that by now we would be having witty conversations with these assistants or, like in the film “Her” we might even fall in love with them.
The reality is the chatbots that we have at present have been known to malfunction so bombastically that, in some instances, they became pretty racist and sexist.
Thankfully, that is not the norm. Chatbots are actually turning into a productive part of the economy. And the good news is that we haven’t even begun to tap their true potential. So it’s worth considering exactly what that potential is and how we can get chatbots to fulfill all of our superhero-inspired dreams.
The state of the machine
In essence a chatbot is artificial intelligence software that aims to meet a consumer’s needs through algorithm-based responses. They generally incorporate machine learning elements which allow them to remember commands and preferences such that they incrementally become more like humans as time progresses.
Chatbots are already well used. According to Gartner they currently account for a third of customer-service interactions — that number is estimated to go up to 85 percent by 2020.
Interestingly, many consumers don’t know that they have interacted with a chatbot, indicating that the “illusion” of human interaction is being successfully created. A recent survey conducted by PointSource found that 54 percent of consumers knew they had used an AI application in the last 12 months. Considering that 62 percent of them actually had done so, it would seem that around 8 percent were unaware that they were using one.
But don’t get too excited quite yet, dear developers, the survey indicated some less optimistic trends as well. 51 percent of consumers still expect frustrations with chatbots’ inability to comprehend requests and about 44 percent question the quality of information provided by these digital conversationalists. Over half still prefer a human interlocutor. Clearly there is still a lot to be desired in chatbot functionality, which begs the question: “What needs to happen to get things up to scratch?”
Follow the success
To really get past the hump that developers are facing will require them to think about their target audiences’ relationships with AI, specifically with regards the success stories. The technology is effectively being used in companies like Starbucks (to place an order for a drink or food) or on apps like Lyft (to request a ride).
These examples illustrate what users love and hate about chatbot experiences. Consumers tend to like them in situations in which rote information is required, eg. requesting data about pricing or a delivery status. Their enthusiasm is diminished in scenarios in which a nuanced approach to a complex problem is required. Until now, that is.
The cavalry’s here
Understanding how to improve chatbots begins with knowledge of how they learn. Basically they adapt their behaviour based on the training data used to develop the algorithms. In other words, a chatbot is initially as good as the training scenarios used in its development. This is where crowdsourcing becomes significant. By diversifying the kinds of input a chatbot is exposed to, it becomes possible to add to their spectrum of possible responses.
Training AI works best if the models are based on data that matches what they are trying to predict. That is to say, if you are trying to model responses to customer complaints, for instance, it is best to obtain thousands and thousands of such queries. The more subtle the differences, the better. Crowdsourcing makes obtaining this data really easy. And best of all, as your accuracy needs to grow, the size of the sample can increase accordingly.
What’s more, users don’t need to be particularly “tech-minded” to garner data from the crowd. Various platforms, like Crowd Guru, do the work for you. Some of the latest machine learning algorithms can even give an indication of their confidence in the predictions. If a model’s self-reported accuracy is unsatisfactory, you can enlist the help of the crowd to increase the data input quality and create a positive “active learning” loop.
Secondly, when it comes to improving chatbot performance more work needs to go into personalization of the bots. It starts with giving them names but doesn’t end there. Facial or personality traits go a long way towards making consumers feel like they are interacting with a human.
And finally, there is also the element of surprize to contend with. Rob Harles, managing director at Accenture Interactive, recently stated in an interview with TechRepublic that the introduction of chatbots is a lot like the advent of ATMs in the 20th century. “Many thought the ATM would replace banks and tellers, but in reality, it simply created a new channel,” Harles said, “That is what chatbots will deliver for brands—a new communication channel.”
This channel will have its own opportunities and challenges. Yes, chatbots may reduce the need for customer service representatives, but it will require specialized personnel to deal with inappropriate or unrelated questions that chatbots are not yet able to handle. They may need to be able to deal with personal questions, current affairs, not-safe-for-work content, and seemingly random inquiries.
The bottom line is the “future” has arrived. As always, it’s a little different to how we imagined it, but probably now more than ever developers are best poised to precisely steer it in the right direction. Let’s hope they don’t disappoint.