Failure or triumph: who will benefit from AI implementation

In mid-August, two events shook the world of artificial intelligence. OpenAI CEO Sam Altman recognized that the situation on the market had signs of an investment bubble. Shortly thereafter, a sensational report from the Massachusetts Institute of Technology (MIT) came out, claiming that 95% of attempts to implement AI in companies had failed. What is this - the beginning of the collapse of a multi-billion dollar industry and a repeat of the history of the dot-com bubble? More likely, the beginning of a new technology's collision with reality.

"Are we in a phase where investors in general are overly enthusiastic about AI? My opinion is yes. Is AI the most important development in a very long time? My opinion is yes, too," Altman said during a large interview with reporters, The Verge writes.

He compared the market reaction to AI to the dot-com bubble in the '90s, when the value of Internet startups skyrocketed before collapsing in 2000.

He added that he thinks it's "crazy" that some AI startups consisting of "three people and one idea" are getting funding at high valuations. "It's irrational behavior," Altman said. - Somebody's going to burn through there, I think."

Here he was probably alluding to competitors, in particular the startup Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, whom I wrote about in detail in July. Her company has received $2 billion from investors at a total valuation of $12 billion, and it doesn't even have a product on the market.

However, Altman does not seem to doubt the future of OpenAI. In the same interview, he said that the company has new AI models that it cannot launch due to a lack of computing power. Therefore, OpenAI is going to invest "trillions" in new data centers in the coming years, which will make "a bunch of economists desperate."

And then a new blow immediately followed. On Monday, August 18, MIT's "The Status of AI in Business 2025" study appeared. It begins with a catchy statement: "While companies have invested $30-40 billion in generative AI, our report reveals a surprising truth - 95% of organizations have not seen any return on investment. Only 5% of integrated AI pilots are generating millions of dollars in revenue, while the vast majority are stuck, with no discernible impact on profit and loss," the authors write.

It all looks alarming together, one can't help but admit it. Investors are spending billions on AI, while companies that should use this tool to make money for themselves and for investors in order to return their investments at a profit are stalling and cannot adapt it to their work.

Shares of AI companies have crept downward. AI flagship Nvidia has lost about 7% since the report was published through September 8, while another industry favorite, Palantir, is down about 12%.

The market reaction underscores growing concerns about the commercial viability of AI, Fortune wrote. "A major new technology has emerged that is sure to change the world and bring success. But some are confusing that with investment success," the magazine quoted Bridgewater Associates founder Ray Dalio as saying.

So it's bad? It depends.

Let's touch on the quantitative side of the MIT study for starters.

As the authors themselves write, it is based on a review of more than 300 publicly disclosed AI initiatives, interviews with representatives from 52 organizations, and survey responses from 153 executives gathered at four major industry conferences.

To be honest, I am not a statistical expert, but there are certain rules for conducting surveys, described for example here. The Australian Bureau of Statistics emphasizes that quantitative conclusions about the overall situation should be drawn when the selection of interviewees was done by random methods.

According to the U.S. Census Bureau, there are 5.9 million employer companies of all sizes and industries in the U.S. in 2022. The level of AI utilization as of June 2025 was estimated at 11.6%, which means we can very roughly estimate that the number of companies using AI exceeds 680,000. The authors of the study should have randomly selected a certain number of firms from this set and surveyed them about the state of affairs. This is not specified in the study, so it is quite logical to assume that the selection was based on the principle of "who agreed to talk", i.e. the sample is not random - it is a so-called convenience sample, and the results of its survey cannot be interpreted as the general state of affairs in the industry.

"You are in a hurry to get preliminary data on your idea. You reach out to your colleagues in the marketing department and collect information from them. This sample gives you initial data, but it doesn't reflect the views of everyone in the company," Survey Monkey gives an example of a convenience sample.

In other words, the eye-catching numbers cited at the outset of the MIT study cannot be extended to all companies implementing AI in some way.

"The claim that 95% of enterprise AI projects fail does appear in the report, but without a detailed explanation of how the calculation was made or the data behind it. Despite such a bold figure, the lack of transparency leaves room for doubt," BigDataWire writes.

Here, for example, a much larger survey by consulting firm BCG, which spoke to 1,000 executives around the world, found that 26% of companies are implementing AI with some success. "Over the past three years, AI leaders have achieved 1.5x revenue growth, 1.6x shareholder return, and 1.4x return on invested capital. They have also achieved outstanding results in non-financial areas such as patent filings and employee satisfaction," the analysts wrote.

Nevertheless, with all reservations, the MIT study is useful already because it has sparked a lively discussion about the challenges of AI implementation and contains interesting insights in itself.

In particular, the researchers noted that more than 90% of respondents already use general-purpose AI such as ChatGPT and Claude in their work, often on their own initiative, without the knowledge of their companies' IT departments (only 40% of companies surveyed have official subscriptions to these AIs). And even in cases where the company has already purchased a special customized AI tool.

"We bought a specialized contract analysis tool for $50k, but I still use ChatGPT for drafts. Our tool gives rigid, inflexible summaries. With ChatGPT, I can have a dialog and bring the text to the right quality. Although both systems supposedly use the same technology, the difference is obvious," the study quotes a lawyer from a medium-sized company as saying.

On the other hand, for complex and mission-critical tasks, 90% of respondents prefer to work with humans. General-purpose chatbots and custom AI assistants suffer from a lack of memory and learnability: they forget context and don't learn from feedback. "AI has already won the 'battle of the easy job' but is losing in tasks where memory, adaptability, and learning are needed," the MIT researchers wrote.

In other words, we can hardly speak of a general failure of AI implementation. Rather, the problem lies elsewhere - it is being implemented in the wrong place and in the wrong place.

The majority of AI budgets are spent on attempts to implement in sales and marketing departments, because there are clear KPIs and the idea is easier to "sell" to the CEO, while the benefits of automating back-office work (for such internal processes as document processing, resource allocation, legal monitoring) are less obvious, but can potentially bring a higher return on investment, according to the authors of the report.

From the point of view of Nathan Ferr and Andrei Shipilov, professors at INSEAD Business School, in all the hype around AI, it is important not to forget the main thing: the company works to best solve customers' problems, they write in Harvard Business Review.

What can really pay off is a careful analysis of internal processes and the external customer journey to identify the points where you can quickly create real value with AI. Then you can embark on a series of experiments with AI tools, measuring their usefulness and assessing how they can scale quickly. The best tools and practices should be scaled by dedicated teams of corporate "ninjas" with "air support" from the company's top management, as is done, for example, at Amazon, Shipilov and Ferre advise.

Ram Menon, co-founder of healthcare AI startup Avaamo, complements that companies need to focus on the tedious and complex tasks of integrating AI into existing processes rather than the hype. "This MIT study that everyone is panicking about? It doesn't prove that AI is a bubble. It proves that we are exactly where we should be on the first day of a decade-long transformation," he writes.

Gartner, which developed the Technology Hype Cycle chart, estimates that generative AI in 2025 has just begun its descent from the "pinnacle of inflated expectations" into the "trough of disillusionment." As they say, you're here now.

It is too early to despair: just at this stage, cases will be forged that will lead the industry further - to the upward "slope of enlightenment", followed by the "plateau of productivity", where companies will finally understand how to use the new technology profitably. Not everyone will get there, but those who have started doing boring "homework" right now have a better chance.

This article was AI-translated and verified by a human editor