On August 28, 2025, a series of reports and industry commentaries sent shockwaves through the business and tech worlds by revealing an uncomfortable truth: the vast majority of generative artificial intelligence pilots in major companies are failing. According to a comprehensive study published by MIT’s NANDA and cited by Fortune, over 95% of these projects deliver zero return on investment. The findings, based on 150 interviews with industry leaders, a survey of 350 employees, and an analysis of 300 public AI deployments, paint a sobering picture for executives and investors alike.
“Some large companies’ pilots and younger startups are really excelling with generative AI,” said Aditya Challapally, the lead author of the MIT report. Startups led by nineteen- or twenty-year-olds, he noted, “have seen revenues jump from zero to $20 million in a year.” Challapally explained that these rare successes come down to focus and execution: “It’s because they pick one pain point, execute well, and partner smartly with companies who use their tools.”
But for most organizations, the dream of digital transformation through AI remains just that—a dream. The MIT research highlights a critical gap: generic AI tools like ChatGPT perform admirably for individuals but fall short in enterprise settings, where they struggle to integrate with existing workflows and adapt to business-specific needs. As Challapally put it, “They don’t learn from or adapt to existing workflows.”
The scale of this failure is staggering when set against the backdrop of the AI investment boom. Over the past year, AI startups attracted around $250 billion in venture capital, according to Fortune. Yet, only 5% of those projects created measurable value. The report’s author, who leads a robotics company serving industries like energy, defense, manufacturing, and mining, described the current state as a gold rush: “Investors are throwing shovels at anyone who promises they’ll strike digital gold. But most of these prospectors don’t know what they’re digging for, let alone how to build something that lasts.”
The core problem, according to the MIT report and echoed in Fortune’s commentary, isn’t the sophistication of AI algorithms. It’s the data—or rather, the lack of it. “AI isn’t failing because of weak algorithms. It’s failing because of weak data,” the robotics CEO wrote. Companies are seduced by sleek presentations and dazzling demos, but beneath the surface lies a fatal flaw: poor data quality, missing data, and a lack of ground truth. “An algorithm is only as good as the fuel you put in. And right now, most companies are trying to drive Ferraris on empty tanks.”
This data gap is especially pronounced in traditional industries, where analog processes and siloed information are still the norm. The commentary draws a vivid analogy: “Most people in tech don’t realize how analog the physical world is, and how complex the systems are. Edison once said, ‘Vision without execution is hallucination,’ and that is what AI is doing in these pilots.”
However, there are notable exceptions. Maisa AI, a one-year-old startup, has taken a different approach and is gaining traction. With $25 million in seed funding led by European venture firm Creandum, Maisa launched Maisa Studio—a model-agnostic, self-serve platform designed to deploy digital workers trained with natural language. Unlike “vibe-coding” platforms that simply generate responses, Maisa’s system uses AI to build the process behind the response, a concept they call “chain-of-work.”
Maisa employs technologies such as HALP (Human-Augmented LLM Processing), which engages users to clarify their needs while digital workers outline each step of the process. To address the notorious problem of AI hallucinations, Maisa developed the Knowledge Processing Unit (KPU), a deterministic system that aims to ensure reliability and accountability. CEO David Villalón told TechCrunch, “We are going to show the market that there is a company that is delivering what has been promised, and that it’s working.”
Maisa’s enterprise-focused strategy means its customer base is still relatively small, but the company is planning rapid growth. With the new funding, Maisa intends to expand its team from 35 to 65 employees by the first quarter of 2026 to meet anticipated demand, beginning in the last quarter of this year. The startup hopes to position itself as a more advanced form of robotic process automation, offering productivity gains without the rigidity of traditional rule-based systems.
Meanwhile, established financial giants are also betting big on AI, albeit with a different focus. On the same day as the MIT report, Citi Wealth announced the deployment of two new AI platforms—Advisor Insights and AskWealth—aimed at accelerating client communications and streamlining adviser workflows globally. According to Citi, Advisor Insights is a dashboard that delivers market commentary, portfolio updates, and event notifications to wealth management staff, drawing on research from Citi Wealth’s Chief Investment Office. The platform is currently in pilot mode for advisers serving Citigold and Citi Private Client customers in North America, with plans to expand to North America Private Bankers in the fourth quarter of 2025 and to international teams in early 2026.
AskWealth, on the other hand, is a conversational assistant powered by generative AI, providing instant responses to queries from service teams and advisers. The platform has already completed its global rollout. Joe Bonanno, Head of Data, Analytics & Innovation at Citi Wealth, emphasized the real-world impact: “These platforms will save hours of time for our advisors, bankers and service teams while reinforcing for clients the personal and high-touch experience that is a tradition at our firm.”
Andy Sieg, Head of Wealth at Citi, was equally bullish: “These new AI-powered tools are gamechangers for Citi Wealth. They give our advisors sharper insights, streamline how we work, and open new possibilities for serving clients with speed and precision.”
Citi’s approach reflects a broader trend in the financial sector, where institutions like Morgan Stanley, UBS, and Goldman Sachs are racing to integrate AI tools to boost adviser productivity and enhance client service. Citi Wealth tested its platforms in Asia before expanding access to other regions, using pilot feedback to refine performance. The bank has also invested in machine learning for credit risk assessment, fraud detection, and customer service automation, positioning AI as central to its competitive strategy.
Yet, as the MIT and Fortune reports make clear, technology alone is not a silver bullet. The commentary warns that the so-called AI bubble won’t burst because AI is useless, but because too many companies try to skip the hard work of data creation. “The winners will be the ones who understand that owning the data means owning the AI. Algorithms will come and go. Data endures.”
The geopolitical stakes are high, too. The race for AI supremacy is increasingly viewed as a race for energy and industrial capacity. Nations that digitize their industrial bases and secure energy resources, the commentary argues, “will win this century.” For companies and countries alike, the message is clear: without a solid foundation of high-quality data, even the most sophisticated AI will fall short.
As the dust settles from these revelations, one takeaway stands out for every CEO and boardroom: before signing off on the next AI pilot, ask one simple question—what data will this AI actually use? Because, as the MIT report and industry voices warn, without the right data, the promise of AI will remain just that—a promise.
