Enterprise AI Transformation Programmes (2025–2026)

Narrative

The single most cited data point in this space, the claim that 95% of generative AI pilots show no measurable P&L impact, originates from MIT's Project NANDA "State of AI in Business 2025" report rather than a peer-reviewed paper. The report is based on 52 executive interviews, surveys of 153 leaders, and analysis of 300 public AI deployments, and its lead author Aditya Challapally has acknowledged that companies were often hesitant to share failure rates, a methodological caveat that critics have used to question the durability of the headline number. The same report describes a "GenAI Divide" in which over 80% of organizations have piloted tools such as ChatGPT or Copilot and nearly 40% report deployment, yet enterprise-grade custom systems are abandoned at high rates, with 60% of firms evaluating them but only 5% reaching production.

The most methodologically rigorous empirical counterpoint to vendor-driven adoption narratives comes from METR, whose July 2025 randomized controlled trial found that when experienced open-source developers were allowed to use AI tools on real repositories, they took 19% longer to complete tasks, even though they had forecast a 24% speedup and, after the fact, still estimated they had been roughly 20% faster. This "perceived versus actual performance" gap, drawn from 16 developers completing 246 tasks, has become a touchstone finding cited across subsequent productivity literature, though METR itself now labels the result historical given how quickly frontier tools have changed. METR's parallel benchmark work, HCAST (Human-Calibrated Autonomy Software Tasks) and RE-Bench, underpins its "time horizon" metric, which the organization says has grown exponentially with a doubling time of roughly seven months since 2019, a trend it continued tracking into 2026 with an expanded 228-task suite.

Economic literature from NBER working papers offers a more measured empirical middle ground. Bick, Blandin and Deming's survey work found that as of late 2024, nearly 40 percent of the U.S. population aged 18 to 64 uses generative AI, with adoption at work proceeding as fast as the personal computer's early diffusion, while Brynjolfsson and colleagues' firm-level studies of customer-service deployment show productivity gains concentrated among newer, less experienced workers. A large three-experiment field study spanning nearly 5,000 developers at Microsoft, Accenture and a Fortune 100 firm found AI-assisted developers increased weekly task completion by 26.08%, with junior developers showing larger gains than seniors, complicating any single "success rate" narrative.

Governance and framework literature on arXiv has grown rapidly but remains dominated by proposed maturity models and architectures rather than longitudinal outcome data. Papers such as the Agentic AI Governance Maturity Model note that industry surveys report only 21% of enterprises have mature governance models for autonomous agents while 40% of agentic AI projects are projected to fail by 2027 due to inadequate governance, and validate their five-level framework, grounded in NIST AI RMF and ISO/IEC 42001, through simulation rather than field deployment. Complementary work maps ISO/IEC 42001, NIST AI RMF and the EU AI Act as complementary rather than substitutive frameworks, with ISO providing certifiable management-system framing, NIST providing risk-function structure, and the EU AI Act providing binding legal requirements within EU scope. Separate empirical work on regulatory operationalization finds that while data governance and cybersecurity practices are relatively mature, significant weaknesses persist in continuous lifecycle governance and oversight of autonomous and agentic AI systems, with explainability requirements inconsistently implemented despite regulatory mandates.