Most U.S. companies are seeing scant returns on their generative-AI experiments, according to a new report from the Massachusetts Institute of Technology’s NANDA initiative. Of the 300 public deployments examined, only 5 percent generated meaningful revenue acceleration, while 95 percent stalled with little or no impact on profit-and-loss statements. The study draws on 150 interviews with corporate leaders and a survey of 350 employees. MIT estimates that U.S. firms have poured between $350 billion and $400 billion into generative-AI projects. Yet more than half of those budgets go to customer-facing sales and marketing applications, even though the highest returns were recorded in back-office automation that eliminates business-process outsourcing and trims external-agency fees. Implementation strategy is emerging as a decisive factor. Companies that purchase specialized AI products and forge vendor partnerships succeed roughly two-thirds of the time, the report finds, while those building proprietary systems internally succeed only about one-third as often. The authors attribute most failures to a “learning gap” between generic AI tools and the unique workflows of large enterprises, rather than to regulation or model quality. The gap is already reshaping workforces: instead of mass layoffs, many firms are opting not to backfill administrative and support roles as they become vacant. MIT researchers say the next competitive frontier will involve agentic systems that can learn and act independently—raising the stakes for firms still struggling to turn today’s pilots into productive assets.