Artificial intelligence (AI) technologies are widely anticipated to revolutionize workplace efficiency, yet their impact on productivity remains muted according to recent research and industry reports. Despite significant adoption rates, many businesses have found that the technology’s promised time savings are offset by the need for extensive human oversight.
A study surveying nearly 6,000 senior executives from the Federal Reserve Bank of Atlanta, Stanford University, the Bank of England, and the Autonomous Technological Institute of Mexico revealed that roughly 90 percent of respondents reported no significant effect from AI on either employment levels or productivity despite 69 percent saying their firms actively use AI systems. This trend aligns with broader economic data, where major productivity improvements tied to AI have yet to materialize, as noted by Tonstern Slok, chief economist at Apollo Global Management, who described the productivity gains as “a forecast rather than an observation.”
One explanation for this disconnect is what experts term a “verification tax.” This refers to the substantial effort required from humans to check, correct, and validate the output generated by AI tools, effectively limiting the net efficiency gains. Christian Catalini, a researcher at the Massachusetts Institute of Technology, highlights that the primary constraint is no longer AI’s intelligence but the human capacity to verify its results. “The only AI that’s useful is the AI that is verified,” he said.
Surveys support this finding: a Workday poll of 3,200 employees and business leaders found they spent about 37 percent of the time saved by AI addressing errors or rewriting low-quality outputs. Similarly, Google economists surveyed 600 scientists and found while AI saved about seven hours per week on average, nearly 90 percent of those time-savers still devoted a portion of that saved time to verification, with 46 percent spending over 25 percent of their AI-related time on checking results. Furthermore, around 40 percent reported growing backlogs in experimental work, indicating that while AI accelerates hypothesis generation, downstream processes like physical testing remain time-consuming.
Industry examples further illustrate the issue. In software development, AI-generated code can be produced rapidly, but companies such as Amazon still require human reviewers to ensure functionality and safety. Collaborative research by MIT and the University of Pennsylvania found that although AI tools boost coding activity, productivity gains diminish during stages requiring human work like testing and integration.
In finance, some investment banks that have heavily invested in AI report increased workloads as employees must verify AI outputs and meet higher client expectations driven by competitors’ AI-enabled services. This phenomenon has prompted aggressive hiring despite automation advances.
Experts note that the verification burden varies by task complexity and the degree of subjective judgment involved. Jon Stoica, a computer science professor at the University of California, Berkeley, explained that while some problems are easier to check, others—such as contract evaluation—require nuanced human expertise. Attempts to use AI to verify AI have limitations, since shared blind spots between AI models can reduce effectiveness.
KPMG U.S. Chair and CEO Tim Walsh emphasized that human review remains integral to automated processes in his firm, cautioning that AI adoption involves trade-offs and is not a simple “switch” to instant productivity gains.
As AI technologies evolve, the bottlenecks presented by verification may shift, potentially requiring human reviewers to develop new, more advanced skill sets. For now, the promise of AI-driven efficiency continues to be tempered by the reality of substantial human oversight requirements.