Beyond the Transformer: Why Organizational Intelligence Outpaces Individual AI Performance

Every time I stand before a class of MBAs or seasoned executives to begin a session on artificial intelligence, I start with a single, deliberate slide. It features a cropped fragment of the first page of the 2017 research paper, Attention Is All You Need. I intentionally choose a fragment rather than the full document because I want to avoid subjecting my students to the dense, often impenetrable prose of academic literature. While I generally steer clear of such papers in executive education—preferring practical application over theoretical complexity—this specific document is the unavoidable exception. It is, quite simply, the most historic document in modern computing.

Authored by eight researchers then at Google, the paper defined the "Transformer" architecture. It was a watershed moment that birthed the current generative AI wave, providing the structural backbone for GPT and the subsequent explosion of large language models. The rest, as they say, is history. However, as we move further away from that 2017 milestone, it is becoming increasingly clear that while the world has mastered the art of building "smarter" artificial workers, we have yet to solve the much harder problem of creating intelligent organizations.

The Transformer as an Architectural Discontinuity

To understand why the Transformer was so transformative, one must look past its success as a model and view it as a radical architectural discontinuity. Before 2017, the path forward in AI was largely incremental—improving existing structures to squeeze out slightly better performance. The Transformer changed the game entirely. By reorganizing computation around the concept of "self-attention," the architecture eliminated the need for the sequential, recursive processing that had defined deep learning for years.

This breakthrough allowed for unprecedented levels of parallel training, effectively breaking the bottleneck that had previously limited the scale and speed of AI development. It did not simply enlarge what was already there; it fundamentally reinvented the architecture of language understanding.

Yet, in the immediate aftermath of the paper’s publication, the industry landscape shifted in ways that continue to offer lessons in corporate strategy. Google, the company that employed all eight authors of the paper, clearly understood the brilliance of the technology. They did not ignore it; in fact, they integrated the Transformer into core products like Google Translate and Search. However, the company famously underestimated the category-defining potential of the breakthrough.

While Google treated the Transformer as an optimization tool for existing infrastructure, others saw it as a new paradigm. OpenAI, which had been founded just two years prior to the publication of the paper, recognized that the Transformer was not merely an incremental improvement for search queries, but the foundation for an entirely new era of human-computer interaction. They built GPT, and eventually ChatGPT, on this foundation, effectively capturing the zeitgeist of the generative AI boom.

The subsequent diaspora of the paper’s original authors—all eight eventually left Google, with seven moving on to found their own companies and one joining OpenAI—serves as a case study in the "Transformer moment." In the context of business innovation, a "Transformer moment" represents an organizing principle whose eventual importance and utility vastly outstrip its initial application. It is the realization that a technological spark can ignite a fire that burns far beyond the laboratory where it was first struck.

Making an Artificial Worker More Capable Versus Making an Organization Intelligent

As we navigate the current landscape of AI adoption, I have consistently emphasized a distinction that is often lost in the hype: the difference between improving an individual’s capability and transforming an organization’s intelligence. An individual—whether human or artificial—is excellent at completing discrete tasks. They can summarize reports, draft emails, or analyze data sets. But a company is fundamentally different. An organization must preserve state, coordinate specialized capabilities across departments, enforce institutional authority, and, perhaps most importantly, remember outcomes and adapt its strategy across time.

The Transformer provided the architectural foundation for extraordinary advances in model-level and task-level capabilities. It gave us machines that can reason, write, and code at a level previously thought impossible. Yet, there remains a cavernous gap between having a powerful model and achieving true enterprise-level performance.

The data confirms this disconnect. In McKinsey’s 2026 survey, a staggering 80% of respondents reported improved personal productivity due to AI integration. This is the "easy" win—giving an employee a better tool. However, only 37% of those same respondents reported a positive contribution to enterprise EBIT, with that figure showing no significant improvement from the previous year. This discrepancy highlights the painful distance between individual augmentation and tangible organizational results.

The high performers identified in the study share a common trait: they do not simply insert AI into their existing, often inefficient workflows. Instead, they redesign those workflows entirely, acknowledging that adding an intelligent agent to a broken process simply results in a faster, more efficient broken process.

In practice, the industry’s current starting point is almost exclusively the "agent." We see this in the development roadmaps of the biggest players in the field. Anthropic, for instance, focuses on the "augmented LLM"—a model equipped with retrieval capabilities, external tools, and a form of memory. Microsoft, meanwhile, promotes various multi-agent patterns, ranging from sequential tasking and concurrent processing to complex, group-chat-style handoffs between models.

These are, without question, useful and powerful techniques. They represent a significant step forward in the sophistication of our tools. But they nearly all begin with the same premise: taking an artificial worker and asking how we can connect it to other workers. This leads us to a critical question: Is coordinating a swarm of artificial agents the same thing as designing an intelligent organization?

To my mind, the two are fundamentally different. Designing an intelligent organization requires a deep understanding of corporate culture, decision-making hierarchies, institutional memory, and the human elements of collaboration that no agent, no matter how sophisticated, can currently replicate. We have successfully built the "worker" of the future, but we are still in the early, experimental stages of building the organization that can actually harness that power to drive lasting, bottom-line impact. We have the architecture of the Transformer, but we are still looking for the architecture of the intelligent enterprise.