Imagine a customer service agent who can do virtually no wrong. She typically answers questions in about three seconds, with an astonishing 94 percent accuracy rate.
She puts callers at ease, offering them information about related issues without making them listen politely through an entire spiel or pressuring them to buy — a key reason many customers do not like chatting with customer service reps.
She’s unruffled by verbal abuse, but she’ll tell an irate caller firmly but politely that certain language is not acceptable. As for her own language, she’s fluent in English, Spanish — and many other languages too.
Her name? Well, it’s anything you want it to be, although she has been known to go by “Sarah.” She — if you haven’t already guessed — is an avatar-based virtual agent developed by MyCyberTwin, a provider of customer support reliant on artificial intelligence.
Users of the technology include NASA and other government agencies, and major banks such as the National Australia Bank. Within the next 30 days, a large U.S. financial institution will add “Sarah” to its workforce, MyCyberTwin cofounder Liesl Capper told CRM Buyer.
Virtual agents are gaining acceptance among companies and by consumers as the technology improves, but that acceptance is coming slowly, Capper said.
That’s partly due to early tech problems that limited deployment of artificial intelligence in a customer service setting.
“One problem was scale — you need to put a lot of content into the ‘brain’ for it to deliver a high rate of accuracy. No company will put a virtual agent on its website if it only has an accuracy rate of 65 percent,” Capper pointed out.
Another problem, at least for the implementing companies, was the huge expensive involved in the earlier generation of technology. “Pilots can be in the high millions of dollars,” she said.
MyCyberTwin addressed these issues with its tech approach. As a result, Capper noted, “we can do a pilot that is in the several hundred thousand dollar range.”
Self Learning Pays Off
The tech issues were addressed through a combination of using deep self-learning technology based on fuzzy logic and natural language programming instead of pretraining a virtual agent from scratch.
“The key element is that the virtual agent is able to make an intelligent decision and then give itself a confidence score about that decision,” explained Capper. “It self learns when the customer accepts the answer.”
In this fashion, the agents gain personality as well as domain expertise.
Not every customer wants an agent with personality, though, Capper noted. The National Bank of Australia, for example, wanted its avatar to be stripped of any anthropomorphic characteristics.
Other clients like the more personable agents. Either way, the nonhuman, yet still intelligent, agent has met an interesting need identified by financial institutions or any company looking to up-sell or cross-sell during a service call: that is, people sometimes just don’t want to be bothered to make nice chit-chat with an agent.
“Really, a lot of people just want to get in and out of these types of conversations and don’t want to be rude by interrupting,” Capper said. “So instead of asking for rates for another product, for example, for fear it will lead to a long pitch, they just don’t ask.”
But they will ask the virtual agent, she said.
“In side-by-side comparisons of calls, people asked the virtual agents twice as many questions,” she said, noting that conversion rates were often twice as high as a result.
These agents also earn the grudging respect of some callers, as they are able to put the unruly and unreasonable in their place in a calm rational manner — a difficult task for us humans.
“The avatars have self-learned how to handle someone who gets abusive or inappropriate,” Capper said. “We have found that the submissive avatars don’t work.”