Most generative AI pilots at companies fail to deliver business impact, according to a new report from MIT’s NANDA initiative. Despite rapid adoption, only 5% of pilot programs show measurable gains in revenue. The findings point to widespread challenges in integrating AI into enterprise operations.
MIT’s study draws from 150 executive interviews, surveys from 350 employees, and reviews of 300 public AI projects. It highlights a stark contrast between successful use cases and those that never take off. While some startups are thriving with generative AI, larger organizations are struggling to convert early tests into real value.
Aditya Challapally, lead author of the report, said younger startups, some led by founders in their teens, have grown revenues from zero to $20 million in just one year. Their success comes from focusing on a single business problem, executing effectively, and building the right partnerships.
Integration issues, not AI models, driving failures
In contrast, the report found that many large companies fall short not because of model performance but due to integration issues. Generic AI tools often fail in enterprise settings because they can’t adapt to complex workflows. Challapally noted that tools like ChatGPT, though useful for individuals, stall in business environments without customization.
The report also points to poor resource allocation. More than half of enterprise AI budgets go toward sales and marketing applications, yet the highest returns are seen in automating back-office tasks. These include reducing outsourcing, cutting third-party costs, and streamlining operations.
Vendor partnerships outperform in-house development
Adoption methods also play a role in outcomes. Companies that purchase AI solutions from specialized vendors and form strategic partnerships see success roughly 67% of the time. In contrast, firms that build tools internally succeed only about a third of the time. This trend is especially relevant in financial services, where many companies still try to develop proprietary AI systems.
According to Challapally, companies often avoid discussing failures, but the data reveals a consistent pattern: vendor solutions tend to deliver more reliable results. The report recommends enabling frontline managers—not just centralized AI teams—to lead implementation efforts. Selecting tools that integrate smoothly and evolve with changing needs is also key.
Workforce impact and the rise of agentic ai
Workforce disruption is already taking shape. While mass layoffs are not widespread, companies are choosing not to refill roles in customer service and administrative functions. These roles are typically those that were once outsourced and are now being phased out as AI systems take over.
The report underscores the growing use of unsanctioned tools such as ChatGPT within organizations and the challenge of measuring AI’s true impact on productivity. Some forward-looking companies are now piloting advanced agentic AI systems capable of learning, remembering, and acting within defined parameters. These developments may signal the next phase in enterprise AI evolution.
