The loudest chatbot success in customer service ended with people coming back. This is not an argument against AI. It is an instruction on how to use it.
I say this as someone who builds AI agents for companies and reads competitors' announcements to the end, including those from the second year.
February 2024. Klarna announces that its AI assistant held 2.3 million conversations in a month, does the work of 700 full-time agents, cut the time to resolve a case from 11 to 2 minutes, reduced repeat inquiries by 25 percent and will bring 40 million dollars of profit improvement. The whole market quoted those numbers for a year.
May 2025. Klarna's CEO says in an interview that AI delivered lower quality in complex cases and the company is again hiring people for customer service. It didn't withdraw the bot. It put a human next to it.
Klarna is an unusually easy case for AI: enormous volume, authenticated customers, cases built on structured financial data. If the bot couldn't carry complex cases there, it won't in a B2B company, in machine servicing or in a shop with made-to-order products either. That is not a flaw of the tool. That is the right place for the boundary.
The customer draws that boundary themselves. Gemius studied in 2025 how Poles shop online: 53 in 100 use AI tools in the buying process, but when they want to contact the shop they choose e-mail (31 in 100), phone (26) and chat with a human (17). Eight in ten people describe contact with a chatbot as positive or neutral. And still don't choose it.
What follows for a company that wants a chat on its website.
First: the chat should be the first line, not the only one. Repetitive questions, hours, delivery, returns, availability, instructions. When the case is complex or the customer wants to buy, a human joins the same conversation, with the whole history. Not "please write to this address", but the same conversation, the same screen.
Second: measure what Klarna didn't show in its first announcement. Not the number of conversations, but the share of cases with a confirmed outcome. Silence on the customer's side is "unknown", not "resolved". Independent benchmarks say the bot handles a median of 41 percent of cases on its own. The rest goes back to a human; the only question is how.
Third: tell the customer who they are talking to. When a human joins, the customer should see it. When the agent answers, too. The law requires it, but above all it is required by the trust Klarna had to rebuild.
That is how Nexly Agents works: the agent answers what is repetitive, and when it recognizes that the customer wants to buy or asks for a human, it hands the conversation to a consultant. Every conversation gets an outcome, including the one where the customer went silent.
A question for you: if your bot handled 2.3 million conversations, who would check how many of them ended well?
Want to see how the agent handles your customers' questions? Enter your website address and e-mail in the demo form. In 20 minutes you get an agent trained on your website's offer, ready to be quizzed.
Sources
- Klarna, press release of 27 February 2024; Bloomberg, interview with Klarna's CEO, May 2025
- Gemius, E-commerce in Poland 2025 (CAWI, n=1,629, July 2025)
- Zendesk ProductLab 2025; Zendesk CX Trends, Salesforce, Forrester 2025 (deflection benchmarks)