I’ve spent the last several months writing about what I believe is the central problem in enterprise AI. Not the models. Not the prompts. Not the context windows. Not even the agents.
The architecture.
Across a series of articles, I argued that large language models were never designed to run companies. Companies operate through memory, context, state, constraints, permissions, incentives, workflows, and feedback loops. Language models operate by predicting the next token.
That mismatch explains a surprising amount of what we see today. It explains why enterprise AI adoption is widespread, but business transformation remains elusive. It explains why organizations report productivity gains while struggling to produce meaningful operational impact. It explains why so many deployments still depend on consultants, systems integrators, and increasingly, forward deployed engineers embedded inside customer organizations. It explains why enterprise AI often feels simultaneously revolutionary and incomplete.
Taken separately, each of those observations is, I believe, interesting. But taken together, they point to something bigger: They suggest that enterprise AI is approaching a discontinuity. And not eventually—soon.
My prediction is simple: Before the end of this year, someone will launch a product that fundamentally changes how companies think about AI. Not a better chatbot. Not a more capable copilot. Not an agent with a longer context window—a new layer.
And once it appears, much of today’s enterprise AI landscape will start to look transitional.
The internet worked before the web
One reason technological transitions are difficult to recognize is that the underlying technology often arrives years before the architecture that makes it useful.
In 1991, the internet already worked. TCP/IP moved packets. Email connected institutions. FTP transferred files. Universities and technically sophisticated organizations could use the network effectively. But the internet was not yet the web.
The breakthrough was not more networking. It was the emergence of a layer that made networking understandable, usable, and buildable by ordinary organizations. URLs, HTTP, HTML, browsers, and servers transformed infrastructure into a platform.
The same pattern appears repeatedly throughout the history of enterprise software. Relational databases became transformative when Edgar F. Codd formalized the relational model. ERP (Enterprise Resource Planning) emerged when enterprise operations acquired a common representation. CRM (Customer Relationship Management) emerged when customer relationships became a manageable system rather than a collection of disconnected interactions. The underlying technologies mattered enormously. But the category-defining breakthrough happened when someone discovered the abstraction that organized them.
I increasingly believe enterprise AI is approaching the same moment.
The models are not the product
This does not mean models stop mattering. Quite the opposite: The frontier models being developed by OpenAI, Anthropic, Google, Meta, xAI, and others are improving at an extraordinary pace. Every new generation expands the amount of intelligence available to organizations.
But that observation leads to a different conclusion from the one most people reach: When a technology improves faster than organizations can absorb it, attention eventually shifts from the technology itself to the architecture that organizes it.
Companies do not buy ERP because they are fascinated by databases. They do not buy Salesforce because they admire SQL (Structured Query Language). They do not choose cloud platforms because they enjoy thinking about virtualization.
The underlying technology remains essential, but it gradually becomes infrastructure.
The business value moves upward
I suspect the same thing is beginning to happen with AI. The most important question is becoming less about which model is smartest and more about how intelligence is organized, deployed, governed, measured, and continuously improved inside the enterprise.
The breakthrough will be simpler than people expect.
The interesting thing about major abstractions is that they often seem obvious after someone discovers them. Think “everything is a file,” “the web is a collection of resources identified by URLs,” or “business operations can be represented as processes and transactions.” These ideas were not simple to invent, but they were simple to explain. And once explained, they felt inevitable.
My suspicion is that the next major enterprise AI breakthrough will have the same characteristic. Not simple to build. Simple to understand.
The strongest innovations rarely arrive as collections of features. They arrive as abstractions that suddenly make a fragmented landscape coherent.
The reaction will not be, “How extraordinary.” The reaction will be more like, “Of course. How else could it have worked?”
What changes next
Over the last two years, the enterprise AI conversation has been dominated by prompts, copilots, agents, context windows, orchestration frameworks, memory architectures, and model benchmarks.
Those discussions are important, but they increasingly feel like discussions about components rather than systems. The next layer will have a different focus:
- persistent state instead of sessions.
- formal representations instead of metaphors.
- governance instead of improvisation.
- optimization instead of generation.
- outcomes instead of outputs.
The organizations that eventually win with AI will not necessarily be the ones with access to the smartest models. They will be the ones that learn how to organize intelligence most effectively.
That shift is already visible in research from McKinsey, Deloitte, MIT, Gartner, Microsoft, and others. Across different vocabularies and industries, the same pattern keeps appearing: Isolated productivity gains are relatively easy. Transforming enterprise performance is much harder. The difference almost always comes down to workflows, systems, measurement, feedback, and organizational architecture.
The conversation is moving away from intelligence itself and toward the structures that make intelligence useful.
That is why I believe the next breakthrough will not arrive from making AI more human-like. It will arrive from making enterprise intelligence more structured.
The prediction
Predictions are dangerous, particularly in technology. Most deserve to age badly. But after spending months studying the evolution of enterprise AI and trying to explain it in this series of eight articles, I am increasingly convinced of one thing: Before the end of this year, a new layer will emerge. A layer that sits above the models rather than competing with them. A layer that benefits from every improvement in underlying intelligence. A layer that makes intelligence part of the operating fabric of the company rather than a separate tool employees occasionally consult.
When that happens, many of today’s architectures will not look wrong. They will simply look incomplete. And we will probably realize that the most important question in enterprise AI was never how to make models smarter.
It was how to make intelligence itself behave like software.
