You will hear the same thing if you walk into any boardroom in Dubai, London, or Frankfurt right now. Everybody is making investments in AI. Everyone believes they must. Furthermore, very few people can accurately predict when it will pay off.
This conflict between urgency and uncertainty is likely what defines modern business anxiety. In a comprehensive 2025 survey of nearly 1,900 executives in Europe and the Middle East, Deloitte found that 85% of organizations had increased their AI spending in the previous year. Ninety-one percent more intend to repeat the action. The money is coming in. Whether it’s working is the question that no one fully responds to aloud.
The truth is that, for the most part, not yet. It usually takes two to four years for an AI investment to yield a significant return. That makes sense, but keep in mind that the majority of technology investments are supposed to pay for themselves in seven to twelve months. Although the water hasn’t started flowing in most places, AI is being treated like infrastructure spending, such as installing pipes. The percentage of respondents who reported returns in less than a year was just 6%. Just 13% of projects, even the more successful ones, reached that milestone in less than a year. On a quarterly earnings call, these figures don’t look good.
The fact that AI seldom appears on its own contributes to the challenge. It appears concurrently with process redesign, team reorganization, and data quality overhauls, all of which are intertwined. To put it simply, one executive at an energy company stated that it is basically guesswork to separate the benefits of AI from the benefits of everything else going on at the same time. “We only managed to get a ballpark estimate,” they replied. Press releases don’t contain that kind of sincere admission.
The particular tension surrounding agentic AI, which promises to manage entire workflows autonomously rather than merely help a human perform a task, is difficult to ignore. There is a great deal of ambition. It seems that the complexity is equally high. Just 10% of organizations that currently use agentic systems report substantial returns. Within three years, half anticipate something significant. Three to five years is what a third are considering. That’s a long way off for something that is being hailed as the next significant increase in corporate productivity in investor decks and conferences.
However, investment continues to grow. It’s nearly impossible to refute the logic’s underlying fear: if you don’t move, the rival across the street will. One financial services executive stated, “You’re going to be left behind if you don’t invest,” and it sums up something true. Adoption of AI has started to feel more like a survival instinct than a strategic decision. One telecom executive acknowledged with unusual candor that the hype is so intense that businesses anticipate AI will “just magically solve everything.” However, the discrepancy between reality and expectations is subtly growing.
In the meantime, tension is developing on a different level, completely outside of the boardroom. The competitive landscape of the AI sector is getting truly bizarre. A flurry of startups appeared with the promise of upending everything, only to discover that only a small number of tech behemoths could truly afford the processing power needed to train the most powerful models. The companies in charge of the infrastructure are Microsoft, Google, and Amazon. In addition to competing with their products, startups are becoming more and more dependent on them to survive. The true impact of these co-opetition agreements on the market is currently being examined by competition authorities in France, the UK, the US, and the EU. As recently as mid-2024, the US Department of Justice demanded “urgent” examination of big tech’s control over AI inputs, including chips, data, and processing power. There’s a feeling that no one really understands how to control a market where the largest participants are also, in many respects, landlords.
Training data is a pain in and of itself. Tighter enforcement of copyright may limit the amount of data that can be used to train the next generation of models, increasing expenses and further restricting the field. One major obstacle is still the cost of specialized AI processor chips. The economics of large-scale AI training may remain concentrated among a small number of players virtually indefinitely, and some of these constraints may never fully resolve. One European policy analysis points out that policymakers have few practical tools to alter this. They are able to examine contracts. They can reduce blatantly detrimental clauses. However, dissecting the economics of AI infrastructure is a completely different issue.
All of this paints a picture that resembles a challenging infrastructure transition rather than a clear-cut gold rush. It is purposefully compared in Deloitte’s research to the transition from steam to electricity in early industrial manufacturing. The benefits of rewiring factories for electricity were gradual. Workflows needed to be completely redesigned. Employees required retraining. Years passed before the change was evident in productivity figures. The same reasoning appears to hold true in this case, but the companies undergoing the change are also vying with one another to achieve an outcome that has not yet been precisely defined. It’s an odd race to be competing in. Additionally, the majority of participants appear to be aware of it.
