Block cuts staff after Jack Dorsey’s AI replacement plan backfires, layoffs hit

Block’s AI‑driven restructuring collapsed under the weight of its own assumptions about fraud detection. According to HumanPages.ai, the company eliminated hundreds of positions in early 2025 after Jack Dorsey announced that “AI could do the work.” Former fraud analysts, however, told The Guardian that the models in use could not replicate the contextual reasoning required to flag suspicious transactions, a shortfall that forced the layoffs to be reversed in many cases. The core problem, as the former employees explained, is that fraud detection at a payments platform is not a clean pattern‑matching exercise; it relies on tacit knowledge built from thousands of edge cases that never appear in historical training data.

The technical gap becomes clear when one examines the decision‑making pipeline. A typical fraud analyst at Block must synthesize merchant category, account age, geographic risk, and real‑time behavioral cues—variables that often lack explicit labels. HumanPages.ai notes that “the knowledge lives in people who have seen 10,000 edge cases, not in a model trained on historical data that fraudsters have already learned to game.” In practice, analysts make judgment calls under uncertainty, weighing incomplete information against evolving threat vectors. AI agents, even those equipped with large language models, struggle to “know what you don’t know” and to treat that uncertainty as a signal, a capability that remains uniquely human.

Block’s attempt to collapse the hybrid human‑AI workflow into a fully automated pipeline ignored the proven “90‑10” model described by HumanPages.ai, where AI handles the bulk of automatable tasks while humans intervene on the residual, high‑complexity cases. The report cites a compliance agent that can verify a business address via database lookups but cannot detect that a storefront has been shuttered for two years or that a phone number routes to a voicemail in a different state—tasks that a human completes in minutes for a modest fee. By eliminating the human layer entirely, Block removed the safety net that catches the 10 % of transactions that require nuanced interpretation, leading to a spike in false positives and missed fraud, which in turn eroded confidence among the remaining staff.

External coverage corroborates the internal fallout. Bloomberg reported that Block cut nearly half its workforce, with Dorsey citing AI as the lever that would let the company “do more with less.” The BBC echoed the scale of the cuts, describing them as “thousands of jobs” while highlighting the company’s public embrace of AI. Wired’s inside look at the rolling layoffs documented how the promise of performance‑driven automation turned into a “smokescreen” for cost‑cutting, noting that many laid‑off employees were still being consulted for ad‑hoc fraud reviews after the AI rollout failed to meet expectations. These accounts collectively illustrate a pattern: the AI tools deployed were insufficiently robust for the high‑stakes, low‑latency environment of payments risk management.

The episode underscores a broader industry lesson about the limits of current AI in fintech. While large language models excel at extracting structured information from text, they lack the ability to maintain and update the dynamic, context‑rich mental models that human analysts build over years of exposure to evolving fraud tactics. As HumanPages.ai emphasizes, “the system works correctly when the agent does the 90 % that is automatable and the human does the 10 % that requires judgment.” Block’s miscalculation—assuming that the 10 % could be eliminated without sacrificing accuracy—demonstrates the danger of over‑generalizing AI capabilities. Until models can reliably incorporate tacit knowledge and handle uncertainty as a signal, fintech firms will need to preserve a hybrid workforce that pairs algorithmic speed with human insight.