As a professor of innovation, there’s a slide in my deck I use every time I start a session on artificial intelligence: a fragment of the first page from the “Attention Is All You Need” paper, from 2017. I use just a fragment because I don’t want the students to read it… I usually refrain from using a lot of academic papers when I teach MBAs and executives: they tend to be boring, and this one is no exception and is not particularly easy to read. However, it is historic. Written by eight Google researchers, it is the paper that defines the Transformer architecture, which among other things, became the architecture behind GPT and much of the subsequent generative-AI wave. The rest, as I said, is history.
The Transformer was not only a successful model architecture: it was also a huge architectural discontinuity. The 2017 breakthrough was able to completely reorganize computation around self-attention, to eliminate recurrence and to make parallel training much easier. And in doing so, it did not simply enlarge what was already there: it completely changed the architecture.
In 2017, Google understood the technology, but underestimated the category it could create. It did not ignore the Transformer: it incorporated the architecture into products such as Translate and Search, but it failed to turn its own breakthrough into the product that defined the next era. OpenAI, founded two years before the paper appeared, recognized its broader potential and built GPT, and ultimately ChatGPT, on top of it. Meanwhile, all eight authors eventually left Google, seven to become founders and one to join OpenAI. For present purposes, “Transformer moment” means an organizing principle whose eventual importance greatly exceeds its initial application.
Making an artificial worker more capable versus making an organization intelligent
I have been insisting on this for a long while now: an individual, human or artificial, can complete tasks, but a company, on top of that, must preserve state, coordinate specialized capabilities, enforce authority, remember outcomes and adapt across time. In that sense, transformers provided the architectural foundation for extraordinary advances in model-level and task-level capabilities. But from there, there is still a significant gap until you can reach enterprise performance.
In McKinsey’s 2026 survey, 80% of respondents reported improved personal productivity, but only 37% reported a positive contribution to enterprise EBIT, with no significant differences from the previous year. This reveals the distance between individual augmentation and organizational results. McKinsey also notes that the high performers are typically those companies that redesign workflows, rather than simply inserting AI into existing ones.
In practice, the industry’s current starting point seems to be the agent. Anthropic explicitly begins with the “augmented LLM”, a model supplied with retrieval, tools and memory, while Microsoft presents sequential, concurrent, group-chat, handoff and other multi-agent patterns. These are useful techniques, but nearly all begin with an artificial worker and ask how to connect it to other workers. Is coordinating agents the same thing as designing an intelligent organization? To me, at least, it looks completely different.
This proliferation of patterns is the evidence of an unsettled foundation
The problem is not just variety: what fails is that every implementation, each agentic project, repeatedly reconstructs the same underlying capabilities, from identity, to permissions, state, policy, evidence, recovery and learning, because no common foundation exists.
If we take a look at Gartner’s forecasts, they predict that more than 40% of agentic projects will be cancelled by the end of 2027 due to cost, unclear value or inadequate risk controls. This supports the idea that model capability alone is insufficient. Gartner also explicitly recommends focusing on enterprise productivity as a whole, rather than individual task augmentation.
What’s the real problem, then? Simply put, that the agent should not be the organizing unit. The center should instead be a persistent, goal-directed organizational system. Models and agents can contribute reasoning or execution, but the company-level system owns the representation of the business, its operational state, objectives, constraints, authority and accumulated experience.
In order to work properly for an organization, such an architecture should have persistent organizational state; formal objects, relationships and permissible actions; objectives and trade-offs; deterministic authorization outside the model; ephemeral agents and credentials; escalation mechanisms; and an append-only record connecting intent, action and outcome. Recent NIST guidance strongly supports architectural separation between reasoning and policy, deterministic access decisions and immutable evidence. Microsoft’s reference architecture similarly separates the agent runtime, policy layer, tools, systems of record and audit pipeline.
What should be the test for a genuine architectural breakthrough?
The evidence to prove a breakthrough should therefore not be whether a system outperforms an agent on some benchmark. A genuine breakthrough should make several apparently separate problems — coordination, governance, memory, replaceability, accountability and continuous learning — much simpler to manage, because they are handled by the same underlying design.
Google invented the Transformer and successfully deployed descendants such as BERT, but it failed to capture the defining consumer product moment in the way OpenAI did, perhaps because they were uncertain about the effects of the technology on their business model, like what Clayton Christensen exposed in “The Innovator’s Dilemma”: Large Language Models were much more expensive to run for web searches, provided a complete answer in such a way that most users wouldn’t bother to click on an ad, and they sometimes hallucinated, endangering the confidence in their results that Google had spent decades building.
Later, Alphabet’s hurried Bard response culminated in errors and roughly $100 billion in lost market value in February 2023. As we can see, research leadership does not automatically translate into product or category leadership. Google has improved a lot since then and has outperformed other competitors on several occasions, but at the time, they lost an opportunity to exploit what could have been a clear pioneering advantage.
When Larry Page and Sergey Brin presented PageRank, they tried to license their technology to several companies, the internet behemoths of the moment. Yahoo! declined even to meet, Excite doubted PageRank’s business value and declined to license it. Those incumbents evaluated an unfamiliar technological advantage as if it were another search technology whose value they doubted, and continued with what they knew worked for them… until it stopped working. Again, the rest is history.
Will today’s AI leaders recognize organizational intelligence as a new category? Or dismiss it because its first implementations appear narrower than the eventual architecture? By insisting on putting the LLM in the center of everything, they are making the same mistake that these internet behemoths committed with Google, and a similar one, misreading the category, that Google committed with the Transformer. Placing the LLM or the agent at the center risks optimizing the previous paradigm instead of discovering the next one.
The current agent-centric paradigm leaves a clearly identifiable architectural gap. Whoever supplies the organizing principle for intelligence at organizational scale will have the opportunity to completely reshape the industry. And the missing architecture will not necessarily be built by today’s model leader.
The only question is whether today’s AI leaders will recognize that moment, or provide another company with the opportunity to define it for them.
