The AI Job Market Reality Check
Matthew Bermango watch the original →
the gist
The anticipated 'white-collar bloodbath' from AI has not materialized, as companies realize that AI requires human oversight and that current enterprise adoption is hampered by integration bottlenecks rather than model capability.
The Disconnect Between AI Hype and Enterprise Reality
Contrary to early predictions from industry leaders like Sam Altman and Dario Amodei, AI has not yet caused the mass displacement of entry-level white-collar workers. Companies that previously conducted layoffs while citing AI as the primary driver were often bloated from over-hiring during the zero-interest-rate era, using AI as a convenient scapegoat for structural corrections. The current job market remains strong, and data from sources like Apollo Research shows no evidence of widespread AI-related job losses. Instead, Jevons paradox is in effect: as AI makes specific tasks cheaper, the demand for those tasks increases, leading to higher overall spending and employment rather than contraction.
The Bottlenecks to AI-Native Operations
While the frontier models are highly capable, most companies struggle to move beyond simple question-and-answer or basic automation. True AI-native company building, as described by firms like Y Combinator, remains a theoretical future state rather than a current reality for most. The primary barriers are not model intelligence, but the lack of internal change management and the existence of business bottlenecks outside of code generation. Shipping features is insufficient if those features are not packaged, marketed, and verified by humans who ensure the end user receives actual value.
The Cost of Frontier Intelligence
Companies are finding that relying exclusively on frontier models like Claude 3.5 Sonnet or OpenAI's latest offerings is becoming prohibitively expensive. High token usage without clear ROI has led firms like Uber to question their AI budgets. While frontier models are necessary for building complex software factories—such as the one Peter Steinberger demonstrated by closing 10,000 issues and 5,000 PRs in a week—most enterprise use cases do not require that level of intelligence. Cheaper, efficient models like DeepSeek are emerging as viable alternatives for enterprise, offering a path to scale without the massive overhead of frontier-model token pricing.