MIT’s NANDA initiative just put hard numbers to what many leaders feel: most generative AI pilots are not moving the P&L. In “The GenAI Divide: State of AI in Business 2025,” the group reports that roughly 95% of enterprise pilots fail to achieve rapid revenue acceleration. The finding draws from 150 leadership interviews, a survey of 350 employees, and an analysis of 300 public deployments. MIT calls this failure rate the clearest sign of a growing divide between organizations that convert AI into operating leverage and those that cannot.
The report points to strategy and integration, not model quality, as the root cause. Generic tools like ChatGPT are flexible and effective for individuals, and that has fueled widespread shadow AI. But these tools stall in enterprise use because they do not learn from or adapt to workflows. Leaders also face a measurement gap. Many organizations cannot credibly tie AI usage to productivity or profit, which makes it easy to spin cycles on pilots that never pressure test against real outcomes. The cultural extremes are telling. IgniteTech’s CEO Eric Vaughan instituted “AI Monday,” where every Monday was reserved for AI projects rather than customer calls or budgets. Headcount figures reviewed by Fortune show that the company replaced nearly 80% of staff within a year. Intense focus is not a substitute for operational fit.
What works looks different from the default pattern. MIT finds that purchased, specialized AI solutions and partnerships succeed about 67% of the time. Internal builds succeed about one third as often. Despite this, “almost everywhere we went, enterprises were trying to build their own tool,” said Aditya Challapally, the report’s lead author. Purchased solutions delivered more reliable results, including in regulated sectors such as financial services. The reason is less about raw capabilities and more about integration and learning. Tools that embed in workflows and adapt over time compound value. And adoption practices matter. Teams that empower line managers, rather than confining AI to central labs, see better traction because ownership sits with the operators who manage daily constraints.
The budget mix also needs a reset. More than half of generative AI budgets are flowing to sales and marketing tools, but MIT found the biggest ROI in back office automation. The gains come from eliminating business process outsourcing, cutting external agency costs, and streamlining operations. That pattern shows up in the workforce data. Companies are increasingly not backfilling positions as they become vacant, with changes concentrated in jobs that were previously outsourced. This is what near term impact looks like when AI is treated as an operations program rather than a demo. It is also where measurement becomes tractable. Cost takeout and cycle time reduction are easier to track than diffuse top line promises.
The 5% success cohort is not imaginary. “Some large companies’ pilots and younger startups are really excelling with generative AI,” said Challapally. He points to startups led by 19 and 20 year old founders that have seen revenues jump from zero to 20 million dollars in a year. The through line is focus on products that learn from usage and fit tightly into a workflow. Advanced organizations are already experimenting with agentic systems that can learn, remember, and act within set boundaries. Whether you buy or build, the bar is shifting from chat interfaces to systems that change with the work.
Sector context underscores the stakes. In manufacturing, leaders are turning to AI to manage cybersecurity risks in converged IT and OT environments amid talent shortages and rising labor costs. Rockwell Automation’s State of Smart Manufacturing Report, which surveyed more than 1,500 leaders across 17 countries, ranks cybersecurity among the top external risks, second only to inflation and economic growth. One third of respondents have responsibilities that span both IT and OT cybersecurity. Among cybersecurity professionals, 48% cite securing converged architectures as key to positive outcomes over the next five years, compared with 37% of all respondents. This is precisely the kind of cross functional, line owned problem where purchased, workflow aware tools and partnerships have an edge over one off internal experiments.
If 95% of pilots stall, the remedy is not more pilots. Treat AI as an operations initiative with a P&L target. Buy and partner where the odds are better. Put line managers in charge of adoption. Focus budgets on back office automation where ROI shows up first. Choose tools that integrate deeply and learn over time. Do this, and your AI program is far more likely to join the 5% that actually move the business.