Klarna's own count put an AI assistant at 700 agents' work

The conversation

On 27 February 2024, Klarna published a press release stating that an AI assistant built on OpenAI technology had, in its first month live, held 2.3 million conversations, a volume Klarna describes as two-thirds of its customer service chat traffic. The release frames that workload as 'doing the equivalent work of 700 full-time agents' and says the assistant operates in 23 markets, 24 hours a day, in more than 35 languages. These are Klarna's own figures, released as a company announcement rather than an independently measured operations report.

What the documents show

The February release states that resolution time fell to under two minutes, compared with an eleven-minute average beforehand, and that repeat inquiries dropped 25%. It says the assistant is 'on par with human agents' on customer satisfaction score, without publishing the survey instrument or sample behind that comparison, and it estimates a $40 million USD profit improvement to Klarna in 2024. A follow-up release from 19 September 2024 expanded the assistant into product search and comparison across Klarna's pricing database, and it restates the same claim that within weeks of a January 2024 debut the assistant was handling the equivalent workload of 700 full-time agents. Both documents name Klarna as the source of every figure; neither is described as audited by an outside party.

The system boundary

Klarna's announcements describe what the assistant handles: customer service chats and, later, shopping questions such as product comparisons and price checks tied to a database the company says covers millions of items. Neither release describes what happens when the assistant cannot resolve a query, what threshold triggers a handoff to a person, or how many conversations end that way. The documents are built to advertise volume and savings, not to specify an escalation path, so a reader cannot tell from Klarna's own text where the boundary between the assistant and a human agent sits.

Where it fails

The releases do not define what counts as a 'handled' or resolved chat, so the two-thirds figure could include conversations the assistant only partly addressed before a person finished them. Klarna also does not disclose the customer satisfaction methodology behind the human-parity claim, and its profit estimate for 2024 was published before the year it covers had ended. A builder weighing a similar deployment should treat these as marketing disclosures rather than a specification.

• What definition of 'handled' or 'resolved' sits behind the two-thirds figure?
• What share of conversations were escalated to a human agent, and how was satisfaction measured for each group?
• Has Klarna published the actual 2024 profit outcome to compare against its estimate?

Klarna's releases are a company's own before-and-after account of a large deployment, useful as a record of what it chose to publish and when, but not a substitute for an independently verified operations study.