AI joins the conversation

Traditional sentiment analysis looks at random samples (usually less than 2% of all interactions) and rely on automated transcriptions with limited accuracy that are produced well after the call has ended. That can give chief customer officers a bird’s-eye look at satisfaction, but it doesn’t help an angry customer on the line with a problem that needs to be resolved now.

Conversational analytics perform real-time voice processing, eliminating the need for batch transcriptions and creating opportunities for service managers to address problems on the fly. One such platform—developed by Dialpad, a cloud communications startup—analyzes 100% of calls as they take place and generates predictions about future customer behavior. In just three weeks of testing with beta customers, according to the company, customer satisfaction scores rose 15%.

At CallMiner, a customer-experience analytics company, similar technology is helping integrate insights from contact centers with data generated by other customer service tools, such as user surveys. Sentiment analysis may suggest that one customer is happy, but does that correlate with what many other customers are actually saying? Are survey results biased one way or another? The new tools help answer those questions, and also help break down silos between a contact center and other areas of customer experience, says MJ Johnson, senior product marketing director at CallMiner. The new technology “exposes this kind of dark data,” says Johnson.

Conversational analytics can also make it easier to train and improve agents, says Temkin. “Contact centers are now using these tools for hyper-targeted agent coaching,” he says. “Companies can get signals from every call, so they can bubble up topics for every agent that needs help, as opposed to randomly listening to them.”

Eventually, says Temkin, companies will be able to focus on customer issues in enough detail that they can provide tools directly to agents, “and let them do self-coaching.”

In the insurance industry, Wall has helped identify another promising use case for conversational analytics—fraud detection. In 2021, Wall led a team, funded by the UK government, to use deep learning to extract hidden information from customer calls. The problem: Call agents typically receive only two weeks of training on identifying suspicious behavior before they are deployed to the front lines. If something seems fishy, says Wall, operators are often tasked with flagging it. Attempts to combat fraud have been exceedingly poor, with a large number of false positives flagged for review and as many cases of actual fraudulent calls missed.

Wall’s software, called LexiQal, allows NLP models to analyze a call in real-time and assign it a risk score between 0 and 100 as it unfolds. Anything above 80 is flagged for investigation.

Built on top of the widely used BERT machine learning model, LexiQal can pick up on some 30 markers in a conversation that can indicate fraud. The model goes beyond looking for certain words, and looks at how often they appear and when they are clustered together, as well other indicators like overtalk and “dead air.”

Proof of LexiQal’s value, says Wall, is that the AI model was able to generate new markers that hadn’t been part of its original expert training data.

LexiQal is now commercially available following a technology transfer to British tech company Intelligent Voice, but Wall says that tools like this could be more valuable if companies partnered together to share training data. “There’s nobody joining the dots because companies don’t share their data,” she says. “We could be progressing a lot faster if we had access to it.”Those challenges won’t stop efforts to further develop the technology. Conversational analytics may eventually provide value beyond the world of customer support. A universe of applications exists beyond the call center, says Harm de Vries, staff research scientist at ServiceNow. He imagines the technology soon being used to empower citizen developers who can create software and run analytics reports using their voice instead of traditional programming tools.

“We’re looking at cases where you can take data scientists out of the equation and empower people to drive analytics themselves,” says de Vries. “Not everyone has the data science skills needed to write this kind of code.”