Unsupervised learning is the only way bots can evolve beyond their original knowledge capability – unless there is a continuous feedback loop, bots cannot advance and be refined to help the customer in the best possible way. Unsupervised learning involves recognition of common human responses like, “Thanks, that was helpful,” or “No, that did not solve my problem.” Data on those exchanges can be fed to product managers who analyze the effectiveness of all of the bot’s answers to different questions, classify them, and then teach it more vocabulary.
Nuance, the creator behind Viv (Siri 2.0), is working to create a richer dialogue and illustrating context. Viv can recognize follow-up questions as a refinement of the previous question, for example. A bot like Alexa may know 50,000 words and may be able to apply them to a few hundred circumstances.1 By no means can chatbots converse at this point – this isn’t Iron Man, even though a Jarvis does exist that can remind you to do things.2
Current limitations notwithstanding, the upside in general is that bots can understand simple commands, perform tasks, and deliver help content within their own specialization. This has added benefits, like eliminating the need to surf for, download, and register with apps. In fact, there is even research that shows people are using far fewer apps than they download.3 With 2 million apps available and thousands more submitted every day, humans are feeling app fatigue.4
Chatbots still cannot expect the unexpected. Yet. And part of the shift in our understanding is that if we are giving up the human-to-human communication style and we accept talking to a machine with inherently limited ability, we should be getting something in return. That is, a heightened level of knowledge about us and an instantaneous recollection of our query history as well as information on our habits, preferences, and recurring instances.
Bots have their own intelligence level in their little microworld that they understand:
- They are not sentient beings, they are looking for markers to classify interactions as good or bad to enable improvement
- The best bots will absorb knowledge gained from interactions and incorporate it into the feedback used to teach it
- They should be able to recognize that time has passed and circumstances might be different even if the query is the same
- If they are really on the cutting edge, they can predict what you need
Alexa, for example, cannot predict yet – Echo is supervised learning only so far. And while Facebook Messenger has a whole bucket for search results labeled “BOTS,” its bot engine does not use deep learning.5 It is on an AI platform, but businesses will need to consider how their bots can evolve based on past interactions to better serve their customers.
Relationships and context
Talking to a bot instead of a human is a tradeoff we’re willing to make, but with the implicit understanding that the bot learns about us and retains that information. We seldom get to experience that with human support agents, so in that sense we will be much better served by bots than we are now by humans. The likelihood of customer satisfaction and a positive experience potentially increases since we have expectations that are more likely to be met.
Context will satisfy a customer, and that is where bots can add real value. Providing context in responding to customer needs will result in a more pleasant experience than dealing with a human and having to explain your backstory or hope they have a log of it somewhere. Context can be what we get back in return for accepting bots. And the feedback that is generated by the user’s verbal reaction to a given answer will inform the chatbot’s responses from then on as it builds knowledge.
Most chatbots assess the customer’s state, then provide content. It then assesses state again – that is, whether it has helped or not – and then provides content. And on and on until the query is resolved, whether the customer seeks help with an issue, more information about a product or service, or even a tutorial. And it all can be delivered with a quick reply, without having to take up your full attention with a phone call.
Even better – where businesses are concerned, capability should be embedded in the bot to recognize when it cannot help the customer, and send that person to a human for those higher level situations. By the same token, if the bot’s confidence in the answer it provides is high enough, it knows there is no need to send the customer to an agent.
As long as businesses incorporate the feedback with programming that recognizes positive and negative outcomes, enabling the right content to be delivered, bots will become more ingrained in our daily lives. At the moment, this feedback system is still being honed by developers. But the user must be the central focus of bots’ evolution, not the gimmick, and customer success the ultimate goal.
- Alexa is a chatbot embedded in a free-standing device by Amazon, called Amazon Echo. By calling to it with the name Alexa, a human can activate the chatbot and ask it questions or prompt it to do things like play music.
- See product introduction at http://hellojarvis.io; see also Eric Ravencraft, “Jarvis Is a Facebook Chat Bot That Can Handle Your Reminders,” Lifehacker.com, May 16, 2016 and Paul Morris, “Say Hello to Jarvis, a Facebook Messenger Chat Bot That Can Set Reminders for You,” Redmond Pie, May 22, 2016.
- See James Tiongson, Mobile App Marketing Insights: How Consumers Really Find and Use Your Apps, thinkwithgoogle.com and Ipsos MediaCT, May 2015.
- See “Number of apps available in leading app stores as of June 2016,” Statista.com; and “App Store Metrics,” Pocketgamer.biz.
- Deep learning is a branch of machine learning that focuses on building and train- ing neural networks through analysis of patterns of data input.