Ending the chatbot’s ‘spiral of misery’

Some household gadget is misbehaving and you need help. Or you have a question about travel arrangements or insurance coverage. You go to the company’s website and a digital imp pops up in a small text window. “How can I help you?” it asks. Or you call a customer service number and a chirpy automaton asks the same thing.

Gamely, you go ahead, typing or telling the chatbot what you want. Its formulaic replies are off the mark. It doesn’t really understand you. Several wayward linguistic volleys later, you give up in despair.

That experience is so common that customer service experts have a name for it: “the spiral of misery.”

But there is good news. Customer service chatbots are becoming less robotic. And they are on a path to improve significantly over the next several years, according to researchers, industry executives and analysts, pulled along by advances in artificial intelligence. They will become more intelligent, more conversational, more humanlike and, most important, more helpful.

“Even now, there are times you sort of can’t tell it’s not a human,” said Bern Elliot, an analyst at Gartner, a technology research firm. “It’s not as good as you’d like, but it is moving in that direction. And innovation is occurring at a rapid pace.”

In research projects, AI has delivered amazing feats of understanding and producing language, known as natural language processing. AI software can write stories and poems, answer trivia questions, translate dozens of languages, and has even created computer programs. These projects typically have all but unlimited computing power and tap unlimited volumes of readily accessible data across the web.

Consumer digital assistant software, like Apple’s Siri and Amazon’s Alexa, also roams the wide-open web to answer questions.

But for most companies, everything is more constrained. Their customer information, needed to answer questions, is not on the web but resides inside corporate data centres. They have less data than the internet giants, and it has accumulated over years, stored in different formats, in different places. (AI algorithms struggle without ample data.) It’s more a geological dig than an internet scan.

Tackling that challenge has become an emerging and increasingly crowded market, called conversational AI. Big Tech companies like Microsoft, Amazon, Google and Oracle have offerings, as do smaller companies and startups including Kore.ai, Omilia, Rasa, Senseforth.ai, Verint and Yellow.ai.

The suppliers provide software tools that companies then customise and train on their own data.

This year, the business market for virtual assistants — aka chatbots — will grow 15% to more than $7 billion, according to a Gartner prediction. Some of those bots are designed to assist employees, but most are for customer service.

No company has made a more humbling and instructive journey to its chatbot technology than IBM. After its Watson supercomputer triumphed over human champions in the TV game show “Jeopardy!” about a decade ago, IBM set about applying Watson’s natural language processing to other fields. An early focus was the diagnosis and treatment of cancer, and IBM called health care its “moonshot.”

In January, after struggling for years, IBM announced it was selling off its Watson Health business to a private equity firm. A few days later, Gartner rated IBM’s Watson Assistant a “leader” in conversational AI for business. Watson has gone from cancer moonshots to customer service chatbots.

Today, Watson Assistant is a success story for IBM among its remaining AI products, which include software for exploring data and automating business tasks. Watson Assistant has evolved over years, being steadily refined and improved. IBM fairly quickly learned that a rigid question-and-answer approach, although ideal for a game show, was too limited and inflexible in customer service settings.

“The real world opened our eyes,” said Aya Soffer, a vice president for AI technologies at IBM Research.

The starting point for improvement, Soffer said, has been a deeper understanding of what happens in call-centres, working with other companies to mine and analyse many thousands of calls between customers and human agents. In dialogues, for example, tracking which questions and which follow-ups led to resolving a customer’s problem, she said, and what were the telltale signals of “conversations that went bad.”

Early chatbots were programmed with a predetermined set of questions and answers. But that led to dead ends if the software did not understand the questions. Today, Soffer said, much of the recent innovation lies in “teaching the system to understand and tease out a person’s intent.”

Creating software that can determine the essence of a person’s inquiry is a central challenge. “You assume there are only so many ways a person can say something, but you learn that is not really true,” said Bob Beatty, chief experience officer for GM Financial.

Initially, the financial services arm of General Motors had a rudimentary chatbot that simply delivered canned answers to a set list of questions. But it began working with IBM in 2019 to develop an interactive chatbot. GM Financial had a two-year plan to develop and roll out its chatbot, powered by Watson Assistant.

The coronavirus pandemic lockdowns in March 2020 meant a surprise acceleration of that timetable. Beatty sent home the 700 or so agents who worked at the company’s call centres in Arlington, Texas, and Chandler, Arizona. While rushing to equip the call centre agents for remote work, GM Financial, with emails and a notice on its website, steered customers toward its nascent chatbot rather than the phone.

The chatbot struggled at first. But the GM Financial developers and IBM engineers programmed in the ability to answer more and more inquiries — no matter how they were phrased — like, “What is my payoff amount?” or “Did you receive my March payment?”

Even simple questions require personalised answers that the software has to look up in a company database, though. At the start, the chatbot called Nanci (its name is within the word “financial”) was resolving less than 10% of customer inquiries. But within two months, the success rate rose to 50% — and is now at 60%, according to GM Financial.

So far, Nanci has been a text-only chatbot, but the company is adding a voice version. And it is working with IBM to automate more complex tasks like changing payment and due dates.

The main purpose of the chatbot technology, Beatty said, is to improve the customer experience and nurture brand loyalty for its parent company, General Motors. But the average call-centre inquiry lasts six minutes and costs $16, according to industry estimates. At GM Financial, many customer questions are now answered by the chatbot. In January, Beatty estimated, the company saved a total of $935,000.

So far, call-centre staff has not been trimmed. The technology, Beatty said, will allow agents to spend more time on difficult problems — for example, speaking to a customer who has lost a job and needs to extend a car lease or loan.

“That’s something a trained, empathetic team member can do in a way AI cannot,” he said.

For most businesses, a hurdle to progress with AI is not having enough training data. Modern AI software requires vast amounts of data to pore through to improve its accuracy — to learn, in its way. Some new AI technology may be able to overcome that obstacle by automatically generating more training data or to learn from lesser amounts of data.

Anthem, a major health insurer covering more than 45 million people, has no shortage of data, and it also has a technology staff of a few thousand including data scientists, AI experts and applications developers. IBM’s Watson Assistant is one of many tools Anthem uses.

Anthem shows what is happening now with AI-fueled chatbots — but also what might be possible in a few years. Its current technology, including its mobile app, is called Sydney and is 90% accurate in answering questions about copayments (“I’m getting a knee replacement. How much does my insurance cover?”) and medications (“Does my prescription have any drug-drug interactions?”), according to the company.

But the long-term goal, said Rajeev Ronanki, president of digital platforms at Anthem, is to use AI to sift through all its claims and clinical data to deliver personalised health advice. And other data: Sydney can even upload fitness tracker information.

There are, for example, more than 380 care and treatment options for people with diabetes, Ronanki said. What are the diet, exercise and medication regimens that have produced the best results for similar patients — by age, gender, other conditions and medical history?

That information could be delivered as treatment guidelines to a physician and as health advice to an individual through an increasingly intelligent and conversational chatbot.

AI, Ronanki said, can “help us move from reactive sick care to proactive, predictive and personalised health care.”

And a solution, perhaps, to the spiral of misery.

©2022 The New York Times Company