Enterprise AI Is Stuck — And the Cause Might Surprise You
Ask most technology leaders why their AI initiatives haven't scaled, and you'll hear a familiar list of complaints. The models aren't powerful enough yet. The context windows are too short. The agents need better prompts. The organization is resisting change. These are real problems, and they are visible ones — which is precisely why they attract so much attention, investment, and energy.
But according to a growing number of analysts and practitioners watching enterprise AI closely, none of these are the deepest problem. The deepest problem is something far more structural, and far less discussed: the entire industry is still building artificial intelligence on top of human metaphors. And metaphors, as history has shown time and again, do not industrialize.
Why Enterprise AI Remains Stubbornly Artisanal
Despite billions of dollars in investment and years of development, enterprise AI deployments remain largely hand-crafted. Each implementation is custom-built, carefully tuned, and heavily dependent on the specific people who assembled it. This is what it means to be artisanal in a technological context — and it is the opposite of what enterprises need to achieve genuine transformation at scale.
This artisanal quality is not a side effect of immaturity or limited compute. It is a structural outcome of the conceptual framework the industry has chosen to work within. When the building blocks of your technology are borrowed from human experience rather than engineered from first principles, you inherit all of the ambiguity and fragility that comes with human analogy.
Over the past two years, the language surrounding enterprise AI has become saturated with anthropomorphic framing. We talk about AI systems having memory, performing reflection, engaging in planning, delegating tasks, processing feedback, and even — in a recent description of Anthropic's training techniques — dreaming. These are evocative, intuitive concepts. They make AI feel approachable and legible to business audiences. But they are also deeply imprecise, and precision is what industrial-grade technology requires.
The Problem with Building from Metaphors
Metaphors are extraordinarily useful tools for communication and early-stage exploration. When a technology is new and its mechanisms are poorly understood, reaching for familiar human concepts helps people grasp what is happening and why it might matter. That is how metaphors serve us well in the early days of a technological wave.
The problem arises when those metaphors stop being communication aids and start being design blueprints. When engineers build systems around concepts like "memory" or "planning" borrowed from human cognition without a rigorous technical definition underneath, those systems inherit the vagueness of the metaphor. You end up with components that seem to work in demos and controlled environments, but that behave unpredictably under the variable, high-stakes conditions of real enterprise operations.
Industries that successfully industrialized previous generations of technology — from electrical systems to software development to cloud infrastructure — did so by replacing intuitive metaphors with precise, measurable abstractions. They built standards, protocols, and interfaces that left no room for interpretation. Enterprise AI has not yet done this work, and until it does, scaling will remain elusive.
What Genuine AI Industrialization Would Look Like
To move past the artisanal phase, enterprise AI needs a different kind of foundation — one built on engineering rigour rather than linguistic intuition. This means several things in practice.
- Standardized interfaces and abstractions: Just as TCP/IP created a common language for networked systems regardless of the underlying hardware, enterprise AI needs interoperability standards that allow components built by different teams and vendors to work together reliably without custom integration work for every deployment.
- Measurable, auditable behavior: The shift from "this agent seems to be reasoning well" to "this agent's decision outputs fall within defined parameters under these specified conditions" is the difference between a craft product and an industrial one. Enterprises need the latter before they can trust AI systems in high-stakes environments.
- Separation of concerns: In mature software engineering, you separate data, logic, and presentation because mixing them creates brittle systems. Enterprise AI architectures have yet to achieve equivalent clarity about what belongs where — and that confusion is a direct contributor to the fragility organizations experience when they try to scale deployments.
- Organizational readiness structures: Technology infrastructure alone is not sufficient. Industrialized AI deployment requires that the human systems around it — governance, accountability, training, and change management — are built to the same standard of reliability and repeatability as the technical components.
The Cost of Staying Artisanal
It would be tempting to view the current artisanal phase as simply a passing stage — an awkward adolescence that enterprise AI will naturally grow out of as models improve and tooling matures. But there is a real cost to waiting for that maturity to arrive on its own.
Every month that organizations build custom, metaphor-driven AI implementations is a month of technical debt accumulating. The hand-crafted integrations of today become the legacy systems of tomorrow. Companies that invest heavily now in architectures built on imprecise abstractions may find themselves facing expensive rewrites precisely at the moment when AI technology finally reaches the maturity needed for genuine scale.
The organizations best positioned for the next phase of enterprise AI will be the ones that recognized this structural problem early and invested in building more rigorous foundations — even when the pressure to ship fast and demonstrate quick wins was enormous.
The Path Forward Requires Honesty About the Problem
None of this means that current enterprise AI efforts are wasted or that the technology is overhyped in any fundamental sense. The capabilities of modern AI systems are genuine and impressive, and they are growing rapidly. The problem is not the technology itself — it is the conceptual scaffolding the industry has built around it.
Acknowledging that enterprise AI is stuck because of metaphors, not models, is the first step toward addressing the right problem. It shifts the conversation from "how do we make the AI smarter" to "how do we build infrastructure worthy of deploying it at scale." That is a harder conversation, and a less exciting one. But it is the conversation that will ultimately determine which organizations manage to turn AI's genuine potential into durable, industrial-grade transformation — and which ones remain perpetually in pilot mode.