The Invisible Fruits of AI
— Structural Transformation of Generative AI through a Three-Layer Lens —
Introduction
A recent MIT report claimed that only 5 percent of companies adopting generative AI have experienced rapid revenue growth. On the surface, this figure fuels the narrative that “AI is overhyped.” Yet, I believe this interpretation is fundamentally flawed.
The impact of generative AI cannot be fully captured through the narrow lens of corporate accounting. Its real effects are diffused, hidden, and yet increasingly pervasive.
1. Why Accounting Fails to Capture AI’s Impact
Corporate financial statements do not directly reflect the productivity gains from AI. Even when employees adopt ChatGPT or Copilot to streamline daily tasks—research, drafting, translation, or design—such benefits do not immediately translate into higher revenues or profit margins.
Instead, they appear as subtle reductions in labor hours, outsourcing fees, and operational costs. In this sense, AI’s impact resembles the “Solow Paradox”: as Robert Solow once noted, “You can see the computer age everywhere but in the productivity statistics.” New technologies often fail to appear in accounting data until, suddenly, they surface dramatically in aggregate productivity metrics.
Note: The “Solow Paradox” refers to this paradoxical phenomenon—new technologies initially fail to show up in accounting data, but eventually emerge forcefully in productivity statistics.
2. The Invisible Productivity Revolution
On the ground, AI’s fruits are already being harvested—though often invisibly. Employees individually leverage generative AI tools, and companies unknowingly absorb the gains through leaner workflows.
Consider NVIDIA’s Japan office: reportedly operating with only 200 staff while managing an enormous market. This illustrates organizational flattening—a structure where small teams create disproportionate value. Such structural changes, invisible to external analysts, represent AI’s deeper impact.
Similarly, Oracle Japan’s leadership has emphasized the extraordinary productivity improvements observed in practice. These transformations are real, even if they remain obscured in conventional accounting reports.
3. The Myth of “No Learning, No Memory”
MIT’s report criticized AI for lacking “learning, memory, and integration.” While accurate as a present limitation, this critique misses the trajectory.
Within the three-layer structure of AI provisioning I propose, Layer II (“Local Inference Entities”) is rapidly advancing. These entities specialize in fine-tuning or distilling foundation models into task-specific engines—whether for manufacturing defect detection or anti-money-laundering analysis. In such domains, AI systems already learn, retain context, and integrate with workflows.
Thus, what seems absent today is in fact emerging unevenly, across layers and sectors.
4. Beyond the Bubble Critique
Sam Altman himself has likened today’s AI enthusiasm to the dot-com bubble. Certainly, the hype is excessive. But like the Internet, AI is built upon a real substrate. After the bubble bursts, what remains will be massive infrastructure investment and flatter, more agile organizations. That is when productivity statistics will finally reflect what is already reshaping workplaces.
Conclusion
To say “AI is not delivering” is only partially true. If one looks solely at financial statements, results appear elusive. Yet beneath the surface, the way people work and organizations are structured is already undergoing a silent revolution.
The invisible fruits of AI, scattered across countless micro-interactions, are accumulating. By the time we recognize them in macro statistics, they will have already redrawn the contours of our economic and organizational landscape.
