The Metaphor Trap
For all the billions poured into enterprise AI, the technology remains stubbornly artisanal. Demos dazzle, but deployments disappoint. The usual suspects—weak models, short context windows, poor prompts, or corporate inertia—are often blamed. Yet these are symptoms, not the root cause. The deeper problem is that the industry is building from metaphors, not formal models. And metaphors do not industrialize.
Over the past two years, enterprise AI has become saturated with human analogies. We speak of memory, reflection, planning, delegation, feedback, and even sleep. Business Insider recently described Anthropic's 'dreaming' technique for AI agents—a telling example of how naturally the industry reaches for human metaphors when describing systems that are, in reality, computational architectures.
Metaphor vs. Model
Metaphors are useful. They make complex systems easier to understand. They help product teams explain what their systems do. They help executives believe they are buying something familiar. But there is a critical difference between a metaphor and a model: a metaphor describes something; a model formalizes it.
This distinction may explain why enterprise AI still feels trapped between astonishing demos and frustrating deployments. When we say an AI 'remembers' or 'plans,' we are using shorthand that obscures the underlying mechanics. A model, by contrast, provides a precise, testable representation of how a system works, enabling engineers to predict behavior, identify failure modes, and optimize performance at scale.
The Industrialization Gap
Industrialization requires repeatability, reliability, and measurability. Metaphors are inherently fuzzy; they invite interpretation and variation. A model, especially a formal mathematical or computational model, leaves no room for ambiguity. Until enterprise AI shifts from metaphor-driven design to model-driven engineering, it will remain a craft rather than an industry.
Consider the difference between a handcrafted piece of furniture and a factory-produced one. The handcrafted item is unique, often superior in quality, but impossible to reproduce at scale. The factory item is consistent, measurable, and improvable through systematic feedback. Enterprise AI today is largely handcrafted, with bespoke prompts, custom fine-tuning, and ad hoc evaluations. This is not a path to widespread adoption.
What Needs to Change
To move beyond the artisanal phase, the AI industry must develop formal models for the behaviors it currently describes with metaphors. Instead of saying an agent 'reflects,' we need a model of iterative self-correction with defined parameters. Instead of 'memory,' we need a formal specification of how information is stored, retrieved, and forgotten. Instead of 'planning,' we need a decomposition of tasks into verifiable subtasks with known dependencies.
This shift will not be easy. It requires a deeper collaboration between AI researchers, software engineers, and domain experts. It demands that we stop being satisfied with 'it works' and start asking 'how does it work?' and 'how can we guarantee it works at scale?'
The Path Forward
Some companies are already moving in this direction. They are building evaluation frameworks, stress-testing their systems, and developing internal benchmarks that go beyond surface-level accuracy. They are treating AI as an engineering discipline, not a magical art. But these efforts are still the exception, not the norm.
Enterprise AI will remain stuck until the industry collectively decides to trade metaphors for models. The tools, techniques, and talent exist. What is missing is the recognition that the deepest problem is not technical but conceptual. We have been building with the wrong blueprints. It is time to draw new ones.
This article is based on reporting by Fast Company. Read the original article.
Originally published on fastcompany.com
