NextFin News - The gap between what artificial intelligence can theoretically achieve and what corporations are actually deploying has widened into a structural chasm, even as the technology begins a process of "autophagy"—consuming the very software ecosystem it once helped expand. While large language models (LLMs) now demonstrate the capacity to handle nearly 94% of coding tasks and a vast majority of administrative functions, real-world adoption in these sectors remains stalled at roughly 30%, according to data released by Anthropic in its March 2026 Economic Index.
This "execution gap" is the central finding of Peter McCrory, Anthropic’s top economist, whose research introduces the concept of "observed exposure." McCrory, a former Federal Reserve economist known for his data-driven approach to labor markets, argues that the "theoretical exposure" often cited by Silicon Valley evangelists fails to account for the friction of the real world. His findings suggest that while AI’s capabilities are evolving exponentially, corporate integration is moving at a linear, often sluggish, pace. McCrory’s stance is increasingly influential among institutional researchers, though some critics argue his data—drawn largely from Claude’s user base—may overstate the readiness of AI for high-stakes, regulated industries.
The friction points are largely operational rather than technological. Companies are currently grappling with a "corporate FOMO" that drives them to launch dozens of pilot programs, yet few of these initiatives scale beyond the experimental phase. Integration with legacy systems, the necessity for human-in-the-loop validation, and shifting regulatory landscapes have turned what was expected to be a sprint into a grueling marathon. In fields like finance and law, where the theoretical potential for automation is highest, the actual "observed exposure" is frequently 50 to 60 percentage points lower than the technical benchmarks would suggest.
Parallel to this implementation lag is a shift in the software landscape itself. The era where AI served as a "plugin" for existing SaaS tools is ending, replaced by a cycle of autophagy where advanced models like Claude and GPT-5 are absorbing the functionalities of specialized third-party applications. Design platforms and specialized coding assistants are seeing their core value propositions swallowed by the foundational models they once relied upon. This consolidation suggests that the competitive advantage is shifting away from the tools themselves and toward the proprietary data and workflow integration of the firms using them.
Market sentiment remains divided on whether this consolidation will eventually force a breakthrough in implementation. While some analysts at major investment banks view the current lag as a temporary "digestive period" before a massive productivity surge, others remain skeptical. The cautious view, held by several risk-averse institutional desks, is that the complexity of organizational change is being underestimated. They point to the fact that even as spot gold prices reached $4,724.20 per ounce this week—reflecting a broader market hedge against various systemic uncertainties—the promised "AI dividend" in corporate earnings has yet to materialize for the majority of the S&P 500.
The risk for enterprises is no longer being "left behind" by the technology, which is now largely commoditized and accessible. Instead, the danger lies in the accumulation of "innovation debt"—a graveyard of proof-of-concepts that fail to reorganize the business around the new capabilities. As AI models continue to eat into the software stack, the winners will likely be those who stop treating AI as a layer of magic and start treating it as a fundamental, if difficult, restructuring of the corporate machine.
Explore more exclusive insights at nextfin.ai.
