The global AI economy has a credibility problem that is not evenly distributed. ~95% of companies that invested in generative AI in 2025 saw no measurable financial return. ~42% percent abandoned most AI initiatives. Copilot has a 5% conversion rate from a 400-million-user Office base after more than a year of availability. These numbers, from MIT, BCG, and Microsoft’s own disclosures, describe the horizontal AI layer, the general-purpose agentic tools that were supposed to transform enterprise workflows in 2024 and 2025. They are failing that thesis, and the market is beginning to price the consequences.

But these same numbers obscure a different story running underneath. In healthcare, legal, and financial services, vertical AI agents are delivering 3:1 to 5:1 ROI with documented business outcomes, 12-to-14-month payback periods, and revenue growth rates that defy the broader AI commercialization crisis. The market is treating “AI” as a homogeneous bucket. The commercial evidence is almost entirely concentrated in vertical, domain-specific agents, and the structural reasons why explain why that gap will widen before it closes.

Academic Research (Published 2020)

This paper investigates the extent to which CEO, industry, firm, year, corporate parent, and business segment effects contribute to variation in the performance of public US companies classified by NAICS industry codes between 2010-2018. Applying several statistical models, the paper finds that 32.9% of segment profit variation is associated with business segment effects with negligible year effects (0.11%), similar to the findings of prior literature. This analysis also finds that corporate parent membership plays a larger role and industry and CEO effects play a smaller role in profit variation than previously suggested. These results have potential implications for the fields of strategic management, financial economics, and others, but several considerations, 1.) comparability and external validity of results, 2.) lack of performance-level mechanisms of causal inference, 3.) reliance on variation as a tool for generalized linear regression, and 4.) autocorrelation, represent key limitations to the interpretation of these results.