From AI Capability to Enterprise Returns

Enterprise AI capability and enterprise financial return are different outcomes, and between 2025 and 2026 they visibly diverged. This paper, the third in a series [1, 2], reviews 87 source records, most published in 2025 and 2026: randomized and quasi-experimental workplace studies, administrative-data studies, official statistics, large executive surveys, capability benchmarks and company disclosures. Each record is appraised separately for causal credibility and for proximity to net financial return, and no record in the review is strong on both.Five findings emerge. First, measured capability compounded while reported value did not: the length of software tasks frontier models complete with 50% success rose from about one hour to about twelve hours between February 2025 and February 2026, and the best published expert-comparison win-or-tie rate rose from 48% to 85%, yet the share of organizations attributing any EBIT impact to AI was 39% in 2025 and 37% in 2026, with high performers flat near 6%. Second, task-level causal effects are real but heterogeneous, ranging from 15% to 26% throughput gains in customer support and software to null or adverse effects elsewhere, and population-level time savings remain near 2% of US work hours, with Danish administrative data ruling out earnings effects above 2%. Third, firms report little realized value: about nine in ten of some 6,000 firms report no productivity effect over three years (realized about 0.3%, expected 1.4%), and 56% of 4,454 CEOs report neither revenue nor cost benefit. Fourth, the cost side is poorly measured: unit prices at fixed capability fall roughly five- to tenfold a year while volumes and capital spending rise, review-inclusive gains on an expert benchmark fall from 90 times faster to 1.12 times, and no company disclosure examined reports AI benefit and full AI cost on the same defined scope with a stated counterfactual. Fifth, returns are concentrated, and the strongest positive firm-level causal estimate, about 4% higher labor productivity, is conditional on complementary investment.The synthesis reframes the enterprise problem as one of conversion: capability is necessary and increasingly abundant, while value depends on workflow redesign, capacity capture and measurement. The paper revisits the earlier papers' claims, retains business-AI integration as a plausible organizational capability rather than a causally established role, and contributes an evidence map, a disclosure audit with a minimum disclosure standard for AI value claims, an evidence-gated decision framework, and a cluster-randomized design to test the integrator hypothesis.