The impact of generative AI on companies has been one of the most discussed issues in recent years, with promises of Transform whole sectorscreate new Business opportunities And, above all, throw uncertainty about The professional future of millions of workers.
However, a recent MIT report has brought to light a less optimistic reality: most of the great experiments with generative the generative in companies, do not achieve The expected resultsalthough it is being invested in them huge amounts of money. In other words: a lot of noise and few nuts.
The “failure” of AI in companies. The figures are clear. According to the study ‘The Genai Divide: State of Ai In Business 2025’ Prepared by the MIT, only 5% of the generative the pilot projects in large companies reach a positive and measurable impact on their income. The rest, an overwhelming 95% of the projects, fails to effectively transform any essential part of the organization or obtain a return on investment, which doubts the global fever for adopting AI accelerated.
It is important to underline that the report considers “failure” the absence of economic return derived from the transformation of the company. That does not mean that the project of I do not provide productivity increases of its employees or in the improvement of their products, as can be inferred from the term “failure” used in the study, such and as they detail in Futuriom.
Chatgpt does not transform, but it is very helpful. As detailed in the study, the problem of poor transforming success of AI projects in companies is not related to the quality of AI models themselves, but rather what researchers call “learning gap”, both for tools and for companies themselves.
The authors of the study emphasize that failure lies in poorly oriented business integration. The generalist models, such as chatgpt or co -pilot, are effective as assistants to Improve productivity of employees on a personal level. However, they fail to adapt to the specialized workflows of companies, and stagnate when they try to replicate on a larger scale. That is, they work very well in the demos, but when they must face the variables of Day -to -day processes fail.
According to the report, 80% of the companies considered using a generalist model to perform a certain task in the company, of these, 50% finally launched a pilot project, and 40% of them were successfully implemented, which is a very high percentage. However, Its effect is not transformative since it is raised as a tool to support employees and does not generate a measurable direct return (it is an indirect return in productivity).
The key is the specialized AI agents. Instead, the AI agentsmuch simpler and specialized in performing a single task, if they manage to transform the area in which they are already applied They imply automation and make human supervision unnecessary. So I explained it to Fortune Aditya Challapally, the main author of the report: “Some large companies and startups have seen their income from zero to 20 million dollars grow in a year because they choose the problem to be solved well, they execute precisely and collaborate with partners who really use their tools.”
As detailed in the report, 60% of the companies surveyed investigated the possibility of implementing this model. Of those, only 20% tested it in a pilot project. Only 5% ended up being implemented and completely automate that task. It is what the study considers “success.”
The AI Stop the Accelerator, does not change the car. The analysis shows that more than half of the budget allocated to generative in companies is invested in projects for sales and marketing, in which the investment made is not being recovered in a measurable way (which does not imply that they are a failure).
The authors emphasize that the true return on investment comes from the automation of internal processes and back-offce. In those cases, the investment in AI begins to recover immediately since all operating expenses are annulled with the automation of the process and everything is return. In other words, what MIT researchers have discovered is that while betting on support tools is how to step on the accelerator of a car to go faster, automate the tasks With ia agents It’s like changing the entire car. The Gartner consultant, It is less optimistic And he estimates that in a few years 40% of those “cars” will crash.
The report also indicates the difference between the solutions created by companies and the tools acquired from specialized suppliers. Only 33% of internal systems achieve some success, while external solutions reach positive rates around 67% of cases.
The “Shadow AI” and the impact on work. The researchers detected what they called “Shadow AI” or unofficial use of tools such as chatgpt by employees. That evidences both the interest in using AI to reduce the workload and the Lack of a corporate strategy Clara for AI. For this reason, the measurement of the real impact of AI in productivity and the benefits it provides is still a pending subject for companies that wish to justify their investment.
Finally, they emphasize that, according to the low successful figures in the implementation of AI as a means of automation and its implementation as an assistant, the disruption in employment is already visible in areas such as areas such as customer service and administrative tasks. However, thanks to that role as “assistant” and not as “performer”, instead of finding us Before millions of layoffsthe trend between companies that have integrated is not to cover the positions that are vacant, especially concentrating on works that were previously outsourced for being considered of low added value.
GIPHY App Key not set. Please check settings