Klarna’s AI experiment shows why virtual workers still need humans

Startups are increasingly experimenting with agentic AI systems as virtual co-founders that can handle coding, market research, customer service, and operations simultaneously—potentially replacing human workers across entire business functions. Real-world deployments reveal a complex reality where AI delivers impressive efficiency gains but often requires human oversight to maintain quality, forcing companies to balance automation ambitions with practical limitations.

What you should know: Klarna, a Swedish fintech company, conducted one of the most aggressive AI workforce experiments, reducing staff from 5,000 to 3,000 employees through strategic AI deployment.

• The company’s AI assistant handled the workload of 700 customer service agents, managing 2.3 million conversations across 35 languages in its first month.
• By mid-2025, Klarna began recruiting human customer service agents again after discovering AI produced “lower quality” outcomes despite being more cost-effective.
• CEO Sebastian Siemiatkowski acknowledged that “it’s critical that you are clear to your customer that there will always be a human if you want.”

Current AI capabilities: Several AI agents are now handling sophisticated business tasks that previously required human expertise.

• OpenAI’s Operator can interact with websites autonomously—filling out forms, making purchases, and completing tasks without needing special programming connections.
• Microsoft has integrated AI agents into its Microsoft 365 Copilot suite to automate routine work, generate content, and summarize meetings across Word, Excel, and PowerPoint.
• Cognition Labs’ Devin functions as an AI software engineer that can plan tasks, write code, test and debug software, and deploy it autonomously.

The deployment reality: Early AI implementations reveal significant risks when systems are deployed faster than organizations can properly test and manage them.

• McDonald’s discontinued its AI drive-thru system after it repeatedly added McNuggets to orders until reaching 260 items.
• Air Canada was ordered to pay damages after its chatbot provided incorrect bereavement fare information to customers.
• MIT’s “The GenAI Divide: State of AI in Business 2025” shows roughly 95% of generative AI pilot programs fail to deliver measurable returns.

Why this matters: The race to replace human workers with AI is revealing that successful implementation requires mature deployment strategies rather than aggressive automation.

• Failures typically stem from poor workflow integration, misaligned focus, overreliance on in-house development, lack of adaptability, and insufficient governance oversight.
• Companies achieving sustainable results focus on human-AI collaboration rather than wholesale replacement, designing intelligent ecosystems grounded in vision and strategic oversight.

What they’re saying: Industry leaders emphasize the nuanced reality of AI workforce integration.

• “I am of the opinion that AI can already do all of the jobs that we as humans do,” Siemiatkowski told Bloomberg, while acknowledging the complexities that statement obscures.