The conversation about AI and jobs has been polarized between two camps since 2022: those predicting mass displacement in the near term and those insisting AI is just a productivity tool that will create as many jobs as it eliminates. The most useful evidence comes not from economists modeling long-run effects but from specific companies that have made concrete decisions about how AI changes their workforce — and are willing to report the actual outcomes.
Klarna is the clearest case study for genuine displacement at scale. The Swedish fintech company reported in January 2024 that its AI customer service agent — built on OpenAI — was handling the equivalent of 700 human agents' worth of customer queries per month. By mid-2026, the company's customer operations headcount has declined by approximately 30% from its 2022 peak, and Klarna's CEO Sebastian Siemiatkowski has been unusually direct that the company intends to remain smaller as it grows rather than return to its peak headcount. The AI is not reducing quality by Klarna's metrics: customer satisfaction scores have held stable, and resolution time has improved. The 700-agent figure represents a real cost reduction that has affected real employees.
IBM's approach has been more cautious and more explicitly managed. CEO Arvind Krishna stated in 2023 that IBM expected to pause hiring for roles that AI could automate, primarily back-office functions like HR and finance. IBM did not announce layoffs of existing employees in those functions — it committed to natural attrition. The practical effect has been that IBM's HR and finance operations have handled growing work volume with flat or declining headcount over two years, with the difference covered by AI and automation tools. IBM's public position is that this frees employees for "higher value" work, though what that higher-value work consists of has not been specifically defined for affected employees.
Duolingo's decision in April 2026 to reduce its contractor workforce for content creation by 10% was explicitly connected to AI-generated curriculum content. The company stated that AI now handles initial content generation for new language pairs and exercises, with human editors reviewing and refining rather than creating from scratch. Duolingo's learning content creation has historically been expensive because it requires native speakers with pedagogical training — a relatively niche combination. AI reduces the creation bottleneck, but the editorial layer remains human.
The pattern across these cases: AI displacement is real, concentrated in functions with well-defined, high-volume repetitive tasks, and happening on a timeline measured in one to three years rather than decades. The functions where displacement is slowest — client-facing roles requiring relationship management, strategic decisions requiring institutional judgment, creative work requiring brand voice — remain human-dominated, but the boundary between automatable and non-automatable is moving inward with each model generation.