AINEWS
Why Enterprise AI Projects Keep Failing Before They Reach Production
A Genpact and HFS Research survey of 2,000 executives blames the data, the processes, and the work underneath, not the models.
A Genpact and HFS Research survey of 2,000 executives blames the data, the processes, and the work underneath, not the models.
Forty-two percent of enterprise AI and analytics initiatives are already failing. The reason, according to the executives who commissioned them, has almost nothing to do with the models.
In a survey of more than 2,000 enterprise leaders across 16 industries and 14 functions, 85% said "enterprise debts" were actively limiting the value they could extract from AI, and only 33% of their data was actually ready to feed an AI system. The findings come from a study published June 15, 2026 by Genpact and HFS Research, the operations-services firm and an industry analyst with a long-standing commercial relationship with Genpact. That pedigree matters. The headline figure the release leans on, $18 trillion in "recoverable" enterprise value, is a constructed estimate, not a measured market.
What is more useful is the diagnostic vocabulary the study uses to explain why AI projects stall. Genpact and HFS group the underlying problems into four interconnected "debts" that companies owe themselves: data, process, technology, and talent. None of them is new, and that is the point. They are the slow-accumulating costs of running a business on patched-together systems and undocumented workflows, and they are now the binding constraint on AI returns.
The data debt is the most concrete. Only a third of enterprise data is structured, labeled, and accessible enough for AI systems to use. The rest is locked in legacy systems, buried in PDFs, or scattered across business units that do not share schemas. The process debt is the next layer: Genpact and HFS estimate that 40% of effort inside the average function is still manual or uninstrumented, meaning there is no event log, no API, no clean handoff for an AI to plug into. The technology debt is familiar, encompassing model sprawl, integration debt, and shadow AI tools bought by individual teams. The talent debt is the people side, not just AI engineers, but the process owners and change managers who can rewire the work itself.
The release frames the payoff for clearing these debts as roughly 8% faster annual revenue growth and 16% annual cost reduction, with the average function now spending about 13% of its budget on AI. Those are model outputs, not measured outcomes, and the study does not disclose its methodology, confidence intervals, or how it derived the $18 trillion figure. Over half of the leaders surveyed also said they have no funded plan to address the debts, which is the more telling number: the problem is identified, the budget is not.
For a chief information officer or operations lead reading this on a Monday, the practical question is not whether AI is overhyped, it is which of the four debts is binding inside the organization. Data readiness is the easiest to measure and the hardest to fix, because it usually means re-platforming a core system that nobody wants to touch. Process debt is the most expensive to ignore, because an AI project built on top of an uninstrumented workflow will quietly replicate the workflow's errors at machine speed. Talent debt, in the study's framing, is less about hiring more data scientists and more about giving the existing process owners the authority and the time to redesign their work.
The next signal to watch is whether the 85% who named the problem become the half who fund the fix. The release, after all, is published by the company that sells process intelligence, automation, and the consulting hours to clear exactly these debts. That is not a reason to dismiss the diagnosis, but it is a reason to read the $18 trillion as a forecast from an interested party, and to weigh it against the smaller, duller numbers in the survey that more directly describe how AI projects actually break.