Are you still there? AI bots on the rise in customer service
PEOPLE WORRIED THAT ‘THESE TOOLS ARE MONITORING EMPLOYEES AND SQUEEZING THEM … AT THE EXPENSE OF EMPLOYEE SATISFACTION.’
WE FOUND THE OPPOSITE: EMPLOYEES ACTUALLY SEEM TO BE HAPPIER.
— ERIK BRYNJOLFSSON, STANFORD INSTITUTE FOR HUMAN-CENTERED AI
The first few months of the year are particularly busy for Wealthsimple Inc.’s customer service team: a two-month rush of questions leading up to the registered retirement savings plan contribution deadline is immediately followed by tax time questions.
But the company’s chatbot has helped manage the deluge. In the first few months of the year, it handled roughly 80,000 client questions monthly; customer support agents handled another 80,000 on their own.
“Some of our clients prefer just talking to a chatbot to get their questions answered,” said Sam Talasila, the company’s head of large language models (LLMS). “(But) any time our clients would like to speak to a human agent, that is readily available to them. By having the chatbot, we’re able to focus more of our agents’ time on these high-value interactions where we can get to know our clients (and) understand their hopes and dreams and goals.”
Wealthsimple’s chatbot is powered by Toronto-based Ada Support Inc.’s LLM technology and trained on Wealthsimple’s data, including the information in its online help centre and articles that its help centre associates have written.
Talasila said the company collects data on every interaction to make sure it was positive and had the correct intended outcome. Wealthsimple’s chatbot has led to a 10-point increase in customer satisfaction scores, according to Ada’s website.
And that’s just Wealthsimple’s experience. Companies of all sizes have quickly moved to deploy generative AI tools into their operations after the launch of Openai Opco LLC’S CHATGPT in late 2022. Customer service has become one of the most popular uses of the technology, with 85 per cent of chief executives surveyed by the IBM Institute for Business Values saying those tasks have become their top generative AI implementation priority. According to the report issued last August, six in 10 executives said that by the end of 2023, they would have invested in generative AI tools to support their customer service agents.
Mckinsey & Co. has estimated the use of generative AI in customer service functions could improve productivity at a value of between 30 per cent and 45 per cent of current function costs.
Companies including Lightspeed Commerce Inc., Bench Accounting Inc., Klarna Bank AB and more have launched tools such as Llm-powered chatbots or those that assist agents with customer calls.
In February, Klarna said its AI assistant had handled 2.3 million conversations within its first month — equivalent to twothirds of its customer service chats — and the bot has been able to resolve questions in two minutes on average, down from 11 minutes previously, and has led to a 25 per cent decrease in repeat inquiries. It also answers questions in 35 languages.
Lightspeed CEO Dax Dasilva said the company’s use of Intercom R&D Unlimited Co.’s Aibased customer service platform and a real-time translation service called Localise had successfully answered 48 per cent of customer questions about the company’s retail product and 37 per cent of those about its restaurant product.
Some are even predicting AI will significantly reduce or eliminate the need for customer service agents.
Tata Consultancy Services Ltd. CEO K. Krithivasan told the Financial Times in April that he believes the rapid advances in AI will lead to a “minimal” need for call centres within a year.
Things are already changing. Klarna CEO Sebastian Siemiatkowski told CBS News that the company now needed the equivalent of 700 fewer full-time agents from the outsourced customer service providers it works with.
Erik Brynjolfsson, the Jerry Yang and Akiko Yamazaki professor and senior fellow at the Stanford Institute for Human-centered AI and director of the Stanford Digital Economy Lab, said he sees augmenting the work of customer service employees as the most effective use of generative AI.
“AI can often make great suggestions and be very creative,” he said. “But you want to have a human making the final call.”
A National Bureau of Economic Research paper, co-authored by Brynjolfsson as well as Danielle Li and Lindsey Raymond from the Massachusetts Institute of Technology, found customer service agents equipped with a generative Ai-based conversational assistant resolved 14 per cent more issues an hour on average. Less experienced agents had a 34 per cent increase in productivity, though more experienced agents barely benefited. The study analyzed data from more than 5,100 customer support agents.
“The real power is from the previous call data. There’s millions of transcripts and a lot of tacit knowledge in there that no one has codified,” Brynjolfsson said. “Instead, the LLM gleans that tacit knowledge … and makes it accessible to less experienced workers.”
The study also found the AI assistant led to higher customer satisfaction and reduced employee turnover. The latter finding, Brynjolfsson said, came as a surprise.
“I was a little worried that the tool might create an electronic sweatshop,” he said. “Some people worried that ‘these tools are monitoring employees and squeezing them, and, yeah, they’re making them more productive but at the expense of employee satisfaction.’ We found the opposite: employees actually seem to be happier.”
Brynjolfsson said this could be a valuable benefit for call centres, where turnover rates are high and constantly hiring and training new employees is a large expense for employers.
Jackie Cheung, an associate professor at Mcgill University and the Canadian Institute for Advanced Research AI chair at the Mila Quebec AI Institute, said he sees a lot of potential in the near future for generative AI to “integrate a lot of information about a potential customer and be able to make predictions about what might be appealing or useful for (them).”
LLMS can also be useful for information seeking, he said, but companies deploying them that way need to design them in a way that ensures they won’t give incorrect information or mislead customers — a phenomenon called a hallucination. Examples of AI chatbot hallucinations have gone viral on social media, including a Chevrolet dealership in Watsonville, Calif., that seemingly sold a customer a car for US$1.
In the most high-profile example, Air Canada in February was held liable for incorrect advice its chatbot gave on the company’s bereavement fare policy. The airline was forced to pay Jake Moffatt $812 to cover the difference between its bereavement rate and the $1,630 he paid on a full-price round-trip ticket to Toronto and back after his grandmother passed away.
Brynjolfsson said these glitches underline why AI is most useful as a support tool for human agents, who can ultimately verify the information put forward and make a decision on how to respond to a customer.
Talasila said Wealthsimple has put strong guardrails in place to prevent hallucinations, including limiting what information the chatbot can access and the categories of queries it can answer, to prevent it from sourcing information from around the web or going off topic.
The company also uses information from customer interactions to continuously fine-tune its instructions for the chatbot and documents it can access. Insights from customers’ interactions with the chatbot have also directly led to new or forthcoming products.
“It’s a very active feedback cycle,” Talasila said. “Now we have the capacity to take a look at every single piece of feedback our clients have given us through this process of categorizing, subcategorizing. We can make sure we’re in a position to act on every single piece of feedback, which is very empowering for our product teams.”
YOU WANT TO HAVE A HUMAN MAKING THE FINAL CALL.