Klarna and the Illusion of AI Agents: Why Humans Are Coming Back

⚡

Key Takeaways

1In May 2025, Klarna began rehiring employees after reducing its workforce by 2,000 due to the inefficiency of its chatbot.

2Klarna's AI deployment failed to fully replace human interactions, especially in emotional cases where listening is crucial.

3Intercom Fin, with a resolution rate of 67%, shows that chatbots can still excel in certain tasks, but with significant limitations.

💡Why it matters — Over-reliance on AI without human support can harm customer satisfaction and the reputation of businesses.

Klarna and the Return of Human Employees

In May 2025, Sebastian Siemiatkowski, CEO of Klarna, discreetly began rehiring human employees after initially touting the merits of an AI assistant capable of replacing 700 customer service agents. This decision came after Bloomberg reported that cost had been identified as a predominant factor in evaluating AI effectiveness, to the detriment of service quality. Between 2022 and 2024, Klarna's employee count dropped from 5,527 to 3,422, but the chatbot remained in place. However, the idea of a fully AI-automated customer service did not materialize.

The notion that autonomous AI agents would replace chatbots is misleading. In reality, this replacement only occurs in specific areas, particularly back-office tasks where there is no direct human interaction. Companies expecting complete automation of their support, sales, or engineering services may find themselves defending budgets against results that will not materialize.

Klarna: An Example of AI Limitations

Klarna's 2024 press release boasted 2.3 million AI-managed conversations in a month, across 35 languages, with resolution times reduced from 11 to 2 minutes. The customer satisfaction score (CSAT) was 4.4, slightly higher than that of human agents, which was 4.2. However, these figures have never been independently verified. OpenAI even used this case study to promote its capabilities, but it was still a chatbot, not an autonomous agent.

Gergely Orosz, in his analysis, pointed out that Klarna had essentially automated level 1 support, similar to the IVR systems of twenty years ago, but using natural language. The chatbot served as a filter, escalating complex issues to humans. Reports from May 2025 by CX Dive and CNBC converged on a unique picture of AI failures, particularly in emotional cases, where being technically correct was not enough. Compliance teams refused to allow the AI to autonomously close accounts, forcing Klarna to reintegrate a flexible human layer, hiring students and remote workers.

The deployment of AI at Klarna was the most cited of the LLM era, but it failed in the part of the job where a customer was online and cared about being there. This is not a story of agents replacing chatbots, but rather an illustration of the resistance of customer-oriented interactions to complete automation.

The Central Argument: Who Triggers the Work?

The debate is not between conversational and autonomous, but rather about who initiates the work. In some cases, chatbots remain more effective.

Chatbots Still Dominate Certain Tasks

Intercom Fin is a clear example of chatbots' ability to handle customer support. By the end of 2025, Intercom Fin reported a self-reported resolution rate of 67% across more than 40 million cumulative conversations, spanning over 10,000 business accounts, with a cost of $0.99 per resolved conversation. In comparison, the cost of a human agent is estimated to be between $5 and $10 per request. Although these figures are provided by the vendor, Teneo's cost analysis for 2025 confirms this order of magnitude, ranging between $8 and $15 per fully loaded human resolution.

However, it is crucial to note that "resolution" is defined by Intercom, and no public study has verified actual customer satisfaction. Resolution rates vary significantly between accounts, with some achieving only 27.6% while others exceed 80%. Companies that invested time in cleaning their knowledge base before launch generally achieve better results.

The unit economics of chatbots remain viable despite these caveats. Intercom Fin operates at scale without requiring human intervention, but that does not mean AI agents can replace chatbots in all situations. Companies must assess whether the additional costs associated with using AI agents justify the potential benefits.

Legal Precedents and Implications

In February 2024, the BC Civil Resolution Tribunal held Air Canada liable for incorrect advice provided by its chatbot regarding bereavement fares, resulting in damages of 650 CAD. Although the amount is small, the precedent is significant. Any system, whether conversational or autonomous, that provides binding information to customers exposes the company to legal risks. This explains why migrations to autonomous systems focus on areas where there is no direct interaction with customers.