$0.02 trillion in AI spending hasn't industrialised a single enterprise workflow. The bottleneck isn't model capability—it's that we're still using metaphors where we need formal models.
Memory Isn't a Data Model, It's a Conversation Logger
Every platform now touts "memory" for agents. Microsoft's Azure OpenAI Assistants API stores message history in persistent threads. Anthropic's engineering team describes long-running agents preserving context across sessions. Useful? Yes. Industrial? No.
A proper data model defines identity, state, relationships, permissions, constraints, and valid transitions. It creates invariants—properties the system guarantees regardless of who uses it. Memory alone retrieves context; it doesn't represent a customer, a contract, an approval chain, or a compliance rule. Companies operate on structures, not recollections.
The Artisanal Trap: Intelligence Without Structure
OpenAI, Anthropic, and Google now send engineers to every customer to map workflows and translate organisational reality into something the model can touch. If intelligence were truly a utility, you wouldn't need a specialist to make the faucet work. The persistence of this model confirms the missing layer is supplied manually.
McKinsey's latest State of AI report confirms the pattern: AI usage is widespread, but most companies haven't embedded it deeply enough into workflows to produce material benefits. The outliers aren't using more AI—they're redesigning workflows. Intelligence alone isn't enough. It needs structure.
Every Major Software Revolution Started With a Formal Abstraction
Relational databases didn't emerge from a better filing cabinet; Edgar F. Codd's relational model defined operations, redundancy, and data independence first. The web didn't scale because browsers got prettier; it scaled because resources got URIs, HTTP formalised stateless requests, and HTML provided a shared grammar. SAP succeeded by formalising the enterprise around processes, transactions, and master data—not by writing prettier screens than consultants.
Enterprise AI is repeating the same mistake Michael Hammer warned about in 1990: automating outdated processes instead of redesigning work. The question isn't "how do we add AI to existing processes?" It's "what formal representation of work would let AI operate safely, repeatably, and accountably?"
What the Industrial Layer Looks Like
The next stage won't be defined by who gives the best name to memory, reflection, or delegation. It will be defined by who formalises them. The winning architecture will preserve state, enforce constraints, encode business semantics, govern permissions, track provenance, and connect actions to outcomes. It will create invariants that partners, extensions, and marketplaces can rely on.
Until then, every deployment is a custom translation exercise. And custom translation is not a platform—it's consulting. The industrial era of enterprise AI begins not when models become more humanlike, but when intelligence becomes more structured.
Source: The real reason enterprise AI is stuck
Domain: fastcompany.com
Comments load interactively on the live page.