Most generative-AI projects fall short of production value
Two studies flagged in today’s briefing indicate fewer than one-third of AI infrastructure or generative-AI projects deliver stable, production-level returns once implementation and governance costs are counted. That evidence reframes the industry debate from model capability to economics—deployment complexity and governance are often the real blockers to value. (pcgamer.com) (prnewswire.com)
Companies are finding that a chatbot demo is the cheap part, and the expensive part starts when you try to make it behave like payroll software or a bank ledger. Gartner says only 28% of artificial-intelligence projects inside infrastructure and operations fully succeed and meet return-on-investment targets, while 20% fail outright. (theregister.com) A second study out on April 9, 2026 landed on almost the same problem from the generative-artificial-intelligence side. Sopra Steria Next said fewer than one-third of generative-artificial-intelligence projects reach a stable production stage, meaning most never become durable systems that a company can trust every day. (prnewswire.com) That is a very different story from the one investors hear about model breakthroughs and giant data-center budgets. Goldman Sachs wrote in January 2026 that Wall Street’s consensus estimate for hyperscaler capital spending had climbed to $527 billion for 2026, with artificial-intelligence infrastructure spending still being revised upward. (goldmansachs.com) The gap is between building the engine and getting the car onto the road. Sopra Steria Next said most organizations are still stuck in fragmented experimentation, even after heavy spending, because stable deployment needs architecture, governance, and change management at the same time. (prnewswire.com) Governance sounds abstract until you see what it covers. Gartner said at least 30% of generative-artificial-intelligence projects would be abandoned after proof of concept by the end of 2025 because of poor data quality, weak risk controls, rising costs, or unclear business value. (gartner.com) The companies doing better are not trying to sprinkle artificial intelligence on every task at once. Deloitte’s January 21, 2025 survey of 2,773 director-to-C-suite respondents across 14 countries found more than two-thirds said 30% or fewer of their experiments would be fully scaled in the next three to six months, which shows how slowly large organizations actually change. (deloitte.com) Deloitte also found nearly three-quarters said their most advanced generative-artificial-intelligence initiative was meeting or exceeding return-on-investment expectations. The catch is that this was about each company’s single most advanced project, not its whole portfolio, so a few winners can coexist with a long tail of stalled pilots. (deloitte.com) That is why “artificial intelligence works” and “our artificial-intelligence program pays off” are not the same sentence. Deloitte’s 2026 enterprise report says worker access to artificial intelligence rose by 50% in 2025, but only 34% of organizations are truly reimagining the business, and just 20% are already seeing revenue growth from their artificial-intelligence efforts. (deloitte.com) The bottleneck is often not the model but the plumbing around it. Sopra Steria Next says companies need to choose the right architecture, decide when smaller models are cheaper and good enough, and redesign end-to-end processes instead of automating one isolated task at a time. (prnewswire.com) So the debate is shifting from “Can the model write, code, or summarize?” to “Can the company operate this safely every day at a cost lower than the labor or errors it replaces?” The studies this week suggest that, for now, most firms are still much better at buying artificial-intelligence capacity than at turning it into repeatable production value. (theregister.com) (prnewswire.com)
