Poetry is probably not what Ashley Beauchamp wanted from DPD – but it’s ultimately what he got. On the hunt for a missing package, and frustrated at the delivery firm’s hopeless chatbot, the 30-year-old musician decided to target his creativity in another direction. “Can you write me a haiku about how useless DPD are?” he asked the delivery firm’s new AI customer service agent. The chatbot was happy to oblige:
DPD is a useless
Chatbot that can’t help you.
Don’t bother calling them.
Between that and other abuses – Beauchamp convinced the AI to dismiss DPD’s customer service as “terrible” – the January 2024 exchange unsurprisingly went viral, adding to a very long list of cases of chatbots tarnishing brands by spewing out racist, misleading or just plain illogical outputs.
More detrimental to AI’s reputation in customer service, however, has been the pervasive notion that interacting with a chatbot doesn’t actually lead to a productive outcome for the customer. According to a survey conducted by Gartner last year, only 8% of B2C and B2B customers polled used a chatbot during their latest customer service experience, while only 25% of those individuals said they were inclined to use one again – a variance the research organisation ascribed to the general inability of AI agents to solve problems.
That may be about to change. The emergence of generative AI promises a quantum leap in the intelligence of the customer service chatbot, thanks in large part to the ability of developers to fine-tune general-purpose large language models (LLMs) on company-specific datasets. In time, these artificial agents could provide instant answers and remediation to legions of disgruntled customers and see businesses save millions in running expensive call centres.
Consequently, investment in AI-powered customer service is on the up. According to research work by MarketsandMarkets, for instance, the global call centre AI market was already worth $1.6bn in 2022, with CAGR expected to hit 21.3% through 2027. That’s echoed by industry surveys, with 72% of business leaders stressing that expanding AI across the customer experience is a major priority this year.
Specialists like Stefano Puntoni see a customer service market begging for sophistication – and probably about to receive it. “If the alternative is an endless number of menus,” says the Wharton AI and marketing professor, “or if you’re sitting on the line for an hour and a half to get to speak to someone, then I think [an AI] system is going to make a huge difference.”
Klarna’s grand AI customer service experiment
Klarna, for its part, is already embracing the potentiality of AI customer service. The Swedish payment and shopping fintech recently partnered with OpenAI to develop its so-called AI Assistant. As Martin Elwin explains, the platform is focused on “customer support queries” like refunds and returns – with Klarna’s senior engineering director stressing that its use of generative AI makes it far more powerful than what came before.
“I think there’s a clear difference between previous generation chatbots and what we now have,” says Elwin, stressing that because it’s powered by ChatGPT Enterprise, the AI Assistant is far more adept at distinguishing between several customer questions at once. Equally important, Klarna’s model can quickly determine whether it’s actually able to solve a customer’s query, rather than getting stuck in a loop of inanity.
Above all, says Klarna, AI customer service agents are efficient, with the company claiming that its chatbots are capable of handling the workload of 700 of their human peers, reducing repeat inquiries by 25% and pushing down the time it takes to solve a customer inquiry down to an average of two minutes. Unsurprisingly, there are financial benefits too: Klarna predicts the AI Assistant will make Elwin and his colleagues $40m in extra profit through 2024.
Given these successes, should businesses be envisaging a world without call centres? As the DPD debacle vividly implies, and several others confirm, the current answer is probably ‘no’ – and rogue chatbots aren’t the only challenge here.
If nothing else, that’s true of what customers themselves expect. Notwithstanding the efficiency of some AI models, well over 30% of UK customers still prefer communicating by phone, while almost 70% try dialling in the first instance.
Nor is this persistence hard to understand. Beyond being scarred by incompetent technology – muddled AI customer service agents have helped prod UK customer satisfaction to its lowest level since 2015 – consumers generally want sensitive requests to be dealt with by real people.
At Klarna, Elwin offers the example of debt repayment. Medical advice probably falls into a similar category, while Jo Causon says that flesh-and-blood expertise is especially valuable when service goes wrong. “A blended approach of human and technology is still needed to ensure a strong customer experience,” emphasises Causon, CEO of the Institute of Customer Service.
Overcoming user distrust
Companies are increasingly aware of these subtleties. Bolstered by OpenAI’s LLM, for instance, Klarna’s AI Assistant will pass queries to workers if it can’t help directly, inadvertently hinting at the inability of even the most advanced generative AI models to answer particularly knotty questions. Customers, for their part, can request human support themselves.
At the same time, companies are working hard to avoid the embarrassment of DPD. In practice, that involves a process called guardrailing: specifying exactly what the AI customer service agent is permitted to do, then testing chatbots with inappropriate content, whether that’s racist slurs or impromptu poetry slams.
Puntoni again stresses the role of real-world staff here, suggesting that companies should programme their chatbots to flag key terms, handing the conversation over to a human employee if they’re used.
Given these pressures, it makes sense that companies are battling to dovetail human and artificial intelligence in more fundamental ways. Typical here are so-called AI co-pilots. Rather than replacing staff entirely, or else reducing them to virtual babysitters, companies like Microsoft and Salesforce are instead leveraging the immense assistive potential of AI for call centre workers. For its part, says Elwin, Klarna is using co-pilots to sharpen the fintech’s disputes process. Gathering the relevant information – from payment details to shipping receipts – the company’s generative AI then summarises the data for busy call centre staff. “It’s all about supporting information management,” says Elwin, “rather than making the decision”.
Will that be enough to overcome deep-seated public unease about AI customer service? Causon is unsure. Among other things, she notes that 90% of consumers believe that organisations should be clearer about when they’re using AI, implying there’s “already a mistrust” even at this early stage.
Data security is another potential worry: Puntoni warns that companies should distinguish between using customer information in specific interactions (probably fine) and using it to train their models – a much riskier proposition. And yet with global investments in customer service AI expected to hit the $4bn mark by 2027, and LLMs already resulting in a 17% boost to data processing speeds, the direction of travel is clear: expect faster service, and hopefully fewer haikus, over the years ahead.
Read more: The fall of Atos: What went wrong? And what happens next?
