95% of AI implementations in business processes fail

The Massachusetts Institute of Technology has concluded that the use of generative artificial intelligence in most cases does not bring companies the expected results. According to the study, in 95% of cases such projects do not lead to an increase in income.

Despite the high interest and rapid spread of new models, only about 5% of pilot programs demonstrate a quick effect. The bulk of initiatives are stopped, with virtually no impact on profit and loss indicators.

The report entitled “Distributing GenAI: The State of AI in Business 2025”, prepared by the MIT NANDA initiative, is based on 150 interviews with managers, a survey of 350 employees and an analysis of 300 public cases of technology implementation. The document clearly shows the difference between successful examples and projects that never reached practical implementation.

One of the study’s authors, Aditya Challapalli, told Fortune that companies that focus on a single problem, invest their time in solving it, and build strong partnerships with tool providers are the ones that succeed.

Most companies have a different story. The problem, the researchers say, is not so much the models themselves, but rather “learning gaps”—both during initial system setup and as they are adopted across the organization. Executives often blame regulatory or productivity constraints for failures, but MIT believes the root cause is poor integration into workflows. One-size-fits-all tools like ChatGPT are great for individual users, but their flexibility is lacking in a corporate environment that requires customization.

The data also shows that more than half of generative AI budgets go to marketing and sales, while MIT sees the greatest return in back-office automation – eliminating outsourcing of business processes, reducing costs for contractors and streamlining internal operations.

The method of application is of particular importance. Buying solutions from specialized vendors and working in partnerships yield better results, while in-house technology often ends in failure. This is especially noticeable in financial services and other highly regulated industries, where many firms are trying to build their own systems.

The report also notes changes in the employment structure: mass layoffs have not yet been observed, but companies are increasingly not filling vacant positions, especially in customer support and administrative positions. Most often, the changes concern tasks that were previously outsourced.

Special attention is paid to new experiments with AI agent systems that are able to learn on the fly, remember, and act autonomously within given frameworks. According to the researchers, it is these technologies that can become the basis for the next stage of applying artificial intelligence to business.