The winners of the AI bubble may not be who we think
The prevailing assumption in AI discourse is that the biggest winners will be the companies that integrate AI most deeply into their products. The more agents, the stronger the competitive advantage. This assumption may well prove wrong, if we look at the broader history of technology.
Text by Martti Asikainen, 13.7.2026 | Photo by Adobe Stock Photos
AI discourse is currently awash with claims that the winners in business will be whoever integrates AI fastest, most widely and most tightly into their operations. The more agents, the flashier the chatbot, and the bigger the “Powered by AI” banner, the better. This is a fairly logical assumption — and probably also a mistaken one.
The history of technology is surprisingly consistent on this point (David 1990; Bresnahan & Trajtenberg 1995). The greatest beneficiaries are usually not the companies that proudly showcase a new technology to their customers, but those that tear their own production processes apart and rebuild them on the technology’s terms. Sometimes this happens so quietly that no one pays it any attention at all.
Perhaps the classic example is the electric motor. The economist Paul David wrote an excellent piece in 1990 on the productivity paradoxes of the modern age. In it, he emphasises that factories did not become more productive the moment the electric motor replaced the steam engine. The productivity leap only arrived decades later, once entire production lines were redesigned around the logic of the motor — no longer built around one great central shaft driven by a steam engine, but organised as a distributed, flexible system instead (David 1990).
In other words, new technology alone was not enough. The old, stagnant ways of thinking had to be dismantled too. Later research on the productivity J-curve confirms the same phenomenon more broadly for general-purpose technologies (GPT) (Brynjolfsson, Rock & Syverson 2021).
As I see it, we have now reached just such a seam with AI. Although it is all but impossible at this stage to predict who will win the race, I can say with near certainty that it will not be the company that adopts AI first and most visibly, but rather the one bold enough to dismantle its old processes to make way for the new (see Brynjolfsson, Rock & Syverson 2021; Barney 1991).
What the data actually tells us
To sharpen my point, I want to borrow an analogy from construction, because it is almost embarrassingly simple. Nobody buys a building because the construction firm has an efficient crane. The customer buys the building, and the crane is simply what allows it to be finished faster and more cheaply. AI has become the software industry’s crane. And there is solid data to back this claim (Peng et al. 2023).
When GitHub, Microsoft Research and MIT studied the question in a controlled experiment, developers using Copilot completed a given task roughly 55% faster than those without it (Peng et al. 2023). That result is no fluke. When the same phenomenon was later examined in real working environments — not a lab task, but genuine production work at three companies (Microsoft, Accenture and a Fortune 100 firm) — data from nearly 4,900 developers showed that completed-task volumes rose by around 26% among developers using AI tools (Cui et al. 2026).
A similar effect has been observed in China. The rollout of Ant Group’s own language model boosted code output by more than 50%, though the benefit was most pronounced among newer, younger employees (Gambacorta et al. 2024). Not every study, however, paints such a straightforward picture. When the research organisation METR examined the question among experienced open-source developers working in repositories they already knew well, AI use actually slowed their work down on average.
This is not a contradiction but a confirmation. AI brings no benefit when the process around it is left unchanged. Experienced developers were already working in optimised, familiar repositories, so there was nothing left to “tear apart.” This is precisely why simply adopting the tool is not enough. At the same time, it illustrates the very unevenness at the heart of AI’s benefits: they are not distributed equally across all tasks or all workers, but depend on how well a given task falls within AI’s zone of competence (Dell’Acqua et al. 2026).
Yet no customer ever buys “AI-assisted development” — they buy software that works and is ready on time. The competitive advantage comes from the crane, but the money comes only from the building. The crane must be used precisely on those jobs where it genuinely speeds up construction. Elsewhere, it is just an expensive machine standing idle in the yard.
Why competence beats the tool
Finnish companies offer a telling illustration of this. According to a recent survey by the Finnish Institute of Occupational Health and Statistics Finland, more than half of Finnish companies already use AI, but only 17% have a written AI strategy. Interview data from the AI for Productivity project paints the same picture in sharper detail: three out of four companies already make use of AI, yet only 7% have managed to integrate it into their organisational processes. The rest are experimenting, but not building.
The gap between experimentation and integration also explains the apparent contradiction between the METR findings (Becker et al. 2025) and the “jagged frontier” research (Dell’Acqua 2026). The question is not whether an organisation has adopted AI, but whether it has the ability to identify where AI actually adds value — and the courage to rebuild its processes around that.
A recent ILO analysis even gives the phenomenon a name: the AI aggregation paradox, whereby substantial task-level productivity gains — typically in the range of 10–70% — fail to translate into measurable benefit at the organisational level without broad diffusion and complementary investment (Chan & Shedania 2026).
Deloitte’s Nordic survey, published this year, makes this concrete: 79% of organisations report efficiency gains from AI, but only 18% report revenue growth, and only one in five has even appointed someone responsible for AI value creation (Deloitte 2026).
McKinsey’s 2025 global survey — covering nearly 2,000 companies — confirms the same logic: of 25 organisational factors, it is the fundamental redesign of workflows that most strongly predicts whether a company sees financial benefit from AI, more strongly than any single choice of model. Yet only around 6% of companies achieved genuine, measurable benefit in this sense (McKinsey 2025).
Build, don’t rent
The surveys cited above amount to exactly the same finding David made about the history of the electric motor back in 1990. Technology is not enough unless the work around it is redesigned. The only difference is that the redesign now happens on a timescale of months, not decades.
If competence is ultimately what decides the outcome, the next question is unavoidable: where does that competence come from? Not by buying the best model, since your competitor can buy the very same thing on the very same day. A sustainable competitive advantage traditionally requires a resource that is valuable, rare, hard to imitate, and non-substitutable (Barney 1991).
Intelligence bought from an external model satisfies, at best, only the first of these four conditions — it is valuable, but neither rare nor hard to imitate, since any paying customer gets the same access at the same time. Capability built from understanding one’s own data, one’s own processes and one’s own organisation, by contrast, satisfies all four conditions, because it cannot be bought ready-made. It can only be built.
This does not mean external models should be rejected, however. That would be as naïve as refusing to use electricity because nobody owns the grid. The question is what gets built, and where. The model itself can remain rented, so long as the understanding built around it — what data feeds into it, where in the process AI works and where it doesn’t, who in the organisation is able to critically judge the result — stays in-house.
It is precisely this difference that explains how two companies can use exactly the same model at exactly the same price and still end up with entirely different outcomes: one owns the learning, the other just owns the bill.
Author
References
Barney, J. (1991). Firm Resources and Sustained Competitive Advantage. Journal of Management.
Becker, J., Rush, N., Barnes, B., & Rein, D. (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. arXiv:2507.09089.
Bresnahan, T., & Trajtenberg, M. (1995). General Purpose Technologies: Engines of Growth? Journal of Econometrics.
Brynjolfsson, E., Rock, D., & Syverson, C. (2021). The Productivity J-Curve: How Intangibles Complement General Purpose Technologies. American Economic Journal: Macroeconomics.
Chan, C.Y.C., & Shedania, K. (2026). The Aggregation Paradox of AI: Why Do Micro-Economic Productivity Gains from AI Disappear at Scale. ILO Research Brief. International Labour Organization.
Cui, K. Z., Demirer, M., Jaffe, S., Musolff, L., Peng, S., & Salz, T. (2026). The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers. Management Science.
David, P. A. (1990). The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox. American Economic Review, 80(2), 355–361.
Dell’Acqua, F., McFowland III, E., Mollick, E., Lifshitz, H., Kellogg, K. C., Rajendran, S., Krayer, L., Candelon, F. & Lakhani, K. R. (2026). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. Organization Science, 37(2), 403–423.
Deloitte. (2026). State of AI in the Nordics 2026. Deloitte.
Gambacorta, L., Qiu, H., Shan, S., & Rees, D. M. (2024). Generative AI and Labour Productivity: A Field Experiment on Coding. BIS Working Papers No. 1208. Bank for International Settlements.
Keränen, P., & Nygård, E. (2025). Current State of AI Adoption in Companies: AI for Productivity Project Report (Reports and Surveys, 84). Centria University of Applied Sciences.
McKinsey & Company. (2025). The State of AI: How Organizations Are Rewiring to Capture Value. QuantumBlack, McKinsey & Company.
Peng, S., Kalliamvakou, E., Cihon, P., & Demirer, M. (2023). The Impact of AI on Developer Productivity: Evidence from GitHub Copilot. arXiv.
Finnish Institute of Occupational Health. (2025). AI Adoption in Companies 2025: Results from the Digital-Green Transition and Work Company Survey. ISBN 978-952-391-236-6.
