95% of Corporate AI Projects End in Failure
A new MIT report, published by the NANDA initiative, shows that even though companies are investing heavily in generative artificial intelligence, most of their efforts end in failure. Based on 150 interviews with executives, a survey of 350 employees, and an analysis of 300 public AI deployments, the study reveals a clear difference between successful cases and those that remain at a standstill. According to Aditya Challapally, the report's lead author and a researcher with MIT's NANDA project, only about 5% of AI pilot programs achieve rapid revenue growth, while the vast majority—95%—have no significant financial impact on a company's bottom line.
This failure is not caused by the poor quality of the AI models themselves, but rather by the so-called "learning gap"—a gap in learning that affects both tools and organizations. Generic tools such as ChatGPT work very well for individuals thanks to their flexibility, but fail in corporate environments because they neither learn from nor adapt to workflows. The report, discussed by author Sheryl Estrada in an article for Fortune, emphasizes that companies often blame regulations or model performance, but the real problem lies in flawed integration into business processes.
Why Pilot Programs Fail and Where the Money Is Lost
One of the key findings is the poor allocation of budgets. More than half of generative AI budgets go to sales and marketing tools, even though back-office automation delivers the greatest return on investment (ROI) by reducing costs related to external agencies, outsourcing, and operational processes. According to MIT data, internally developed AI systems fail almost twice as often as those purchased from specialized vendors. Nevertheless, many companies, particularly in regulated industries such as financial services, insist on building their own proprietary systems.
Aditya Challapally noted in an interview that companies are often reluctant to share their failures. "Almost everywhere we went, businesses were trying to build their own tool," he said, but the data shows that purchased solutions deliver more reliable results, with a success rate of around 67%. By contrast, in-house development succeeds in only one-third of cases. Another problem is so-called "shadow AI"—unauthorized tools such as ChatGPT that employees use without oversight, increasing security risks and reducing control.
Successful Cases and What We Can Learn from Them
Some companies, however, excel with generative AI. For example, young startups led by nineteen- or twenty-year-old founders managed to increase their revenue from zero to $20 million (approximately CZK 450 million) within a single year. According to Challapally, they achieved this by focusing on a single pain point, solving it well, and forming smart partnerships with companies that use their tools. Large companies that have succeeded often empower line managers rather than central AI labs and choose tools that integrate deeply and adapt over time.
The report also highlights ongoing changes in the labor market. Rather than conducting mass layoffs, companies are leaving vacant positions unfilled, particularly in customer support and administration. These changes are concentrated in roles that were previously outsourced because of their low value. In addition, advanced systems such as agentic AI are emerging that learn, remember, and act independently within defined boundaries, indicating the direction of future development.
Technical Limitations and the Future of AI in Business
Additional information from MIT research shows that generative AI has a technical limitation—models can perform specific tasks accurately, but they lack a coherent internal model of the world. For example, they may provide accurate navigation instructions in New York but fail when streets are closed or detours are required. Effectiveness depends on the quality of user queries (prompts), and automatically rewriting prompts sometimes leads to worse results when they do not align with the user's intent.
In software development, AI can increase developer productivity by up to 55% under controlled conditions, but rapid and careless deployment increases technical debt, particularly when integrating with legacy systems, leading to scalability and stability issues. MIT emphasizes that organizational readiness, careful integration, and realistic expectations are essential to realizing the potential of generative AI. Although the Federal Reserve expects AI to ultimately increase productivity, current failures highlight significant challenges in both the technology and its application.
This MIT report serves as a warning to companies: without the right approach, generative AI will remain merely an expensive toy with no real benefit. If you want to succeed, focus on adaptation, partnerships, and back-office automation—that is the path to genuine results.
