Something is wrong with this widget.
Calling customer service should fix that problem. But the first person you talk to may not be human.
Chatbots are becoming the first line of interaction in the customer service experience. Natural language processing is what enables the bot to understand what you are saying. Right now, that capability is pretty basic and limited, but that level of automation is about to scale towards autonomy. Companies have to get this right, since a good service call can fix resolve many service and product problems. Fail this and you burn your own brand.
What did you say?
Computers aren’t born speaking English. Natural language processing (NLP) enables them to do so by taking our spoken or written words, breaking them down, and rendering them into a computer language. Words have to be tagged by parts of speech. The completed sentence has to be crossed-checked against the grammar rules already encoded in the system. Finally, algorithms must be applied to comprehend words that sound alike but mean different things, or different words that may mean the same thing. And the AI has to be trained repeatedly to filter out mistakes.
Unfortunately for the computer, natural language can be ambiguous, thus adding to the challenge of using written or spoken words as an input. “Complex tasks, like playing chess, are extremely easy for a computer,” noted Cem Dilmegani, founder and CEO of appliedAI.com. Simple, everyday speech, on the other hand, doesn’t make big demands on the conscious mind. “It’s easy for us. It’s hard for the computer,” he said.
May I help you?
So how do you get a chatbot to do something?
Decision trees form the underlying guidance of a chatbot. The tree is nothing more than a series of questions. Each expected answer leads to the next expected question, until a solution is reached. AI can recognize the caller’s words and pick the right responses from the branches of the decision tree. This approach can handle the “routine stuff” in a typical customer service call. “AI is awesome responding to questions that are repetitive in nature,” said Ravi Raj, founder and CEO of Passage AI. “Something requiring intuition or creativity requires a human.”
The NLP/AI combination can judge whether the problem is outside of the decision tree, switching the call to a human who can resolve the extraordinary issues. Freed from routine, the human can act as the “value add” in the service call.
“Sentiment Analysis on user input (whether that is text or voice) can help in understanding whether they’re happy, unhappy or having a neutral experience.” said Mitul Makadia, director at Maruti Techlabs. “Conversation length is yet another factor that should be taken into consideration. If it takes the bot too long to resolve a customer issue, it is best to do a quick bot to human transfer.”
For Passage AI, the user can set a tolerance threshold, a point at which the chatbot will switch the call to a human, Raj explained. Sentiment analysis can act as the tripwire, again handing the call to a human if the caller expresses dissatisfaction.
“The key with any automated approach to customer service is providing value to users when you get it right, and not getting in their way when you get it wrong. “ explained Chris Hausler, Data Science Manager at Zendesk. “This provides delight and expedited service in the situation where we can provide the correct answer without otherwise delaying the resolution.” Zendesk’s Answer Bot already helps customers resolve most issues, but it is also “trained” to switch the call to a human should the automated service fail to provide a useful answer, he added.
Chopping down the decision tree
Lines of inquiry that take the caller “off branch” reveal the limits of the decision tree. Still there are other approaches one can use to expand the chatbot’s reach, using AI.
Start with machine learning, treating the service calls as data. “You can identify scenarios where the bot fails to answer queries based on the decision tree format, and train it further with the relevant intents and utterances in order to make sure that failure in the same instance is not repeated.” Makadia said.
The AI can also help the human handle the call. AI operating sentiment analysis can judge what problem the caller is experiencing, then make suggestions to the call handler about possible remedies, Dilmegani explained. A similar approach is used to augment sales calls, again offering counterpoints to the sales person to help convert prospects into customers, he added.
But the next challenge is to apply NLP and AI to produce a truly conversational chatbot that does not rely on a decision tree for guidance. “We’re at the point where basic conversational comprehension is there, but more complex interactions need to be curated in a workflow-like manner.” Hauser said. “ However the field is advancing rapidly and I would expect to see conversational bots that can handle multi-faceted complex queries in the coming years.”
Human error
Still, no matter how well a human can build a system, another human can find a way to screw it up. People are not always clear. “Anyone who is a bot developer, or is in charge of auditing conversation history, would have stories for days about how often humans tend to go off-track with the conversations.” Makadia noted.
“People often get caught up on little things such as unclear wording or flow which can lead to misunderstanding and disappointing outcomes for all involved.” Hauser said. Zendesk relies on user experience (UX) researchers as part of the product development team to understand and design for these cases, he said.
Just like human agents, even the most intelligent chatbots will need to be monitored for success. “If the customer is getting frustrated, there are a number of ways to see that,” Dilmegani said. “Go through all the answers…See how the chatbot translates certain sequences.”
Indeed, our experts all referred back to Microsoft Tay as a good example of how not to construct an intelligent chatbot. The system, activated in 2016, only lasted 24 hours, spewing racist and sexist rant after “learning” how to “speak” using unfiltered feeds from Twitter. Effective AI training requires humans to really pay attention to the data being used and the answers the model produces,correcting the mistakes before the system goes live.
One can train the chatbot to handle many scenarios, based on conversation history, and enable the system to reroute the conversation back to its original context, Makadia said. “Alternatively, you can also define a flow for random queries that come in. However, what works best is to define different conversations and responses based on multiple different personas that will interact with your solution.”
“The Holy Grail is 100 percent accuracy,” Raj said.
