Why enterprise AI Remains Artisanal: The Missing Formal Layer

Enterprise AI’s Stubborn Artisanal Phase

Despite breakthroughs in model size, enterprise AI still feels like a handcrafted project. The obvious culprits—short context windows, prompt‑tuning, or corporate inertia—are merely surface symptoms. The deeper flaw is cultural: the sector builds on human‑centric metaphors instead of rigorous abstractions.

Metaphors versus Models

Terms like “memory”, “reflection” or “sleep” make the technology digestible, but they stop short of a formal data model. A metaphor describes; a model defines identity, state, permissions and invariant rules. Without that, deployments flop after flashy demos.

“We keep talking about AI agents as if they were people, and then wonder why they won’t scale,” a senior engineer told Reuters.

Consider “memory” in current platforms. Azure’s Assistants API stores thread history; Anthropic warns about context drift across windows. Useful, yet it merely records events. A true model would encode a customer record, an approval chain, a compliance rule—each with constraints and lifecycle transitions.

Because enterprises still have to hand‑craft those constraints, vendors ship engineers to every client. The result is a consultancy‑heavy delivery model rather than a plug‑and‑play platform. When a ↑ 3x surge in AI spend is observed, the ROI remains uneven.

Industrialization Demands Formal Abstractions

History repeats itself. Relational databases rose after Codd’s relational model; the web exploded once URLs, HTTP methods and status codes were standardized; ERP systems dominated by formalizing transactions, master data and process flows. Each leap required a grammar that guaranteed predictable behavior.

Enterprise AI lacks that grammar. Companies juggle unique vocabularies, legacy systems and political realities. Without a shared invariant layer, every implementation becomes a bespoke translation exercise. pandemic‑era digital upgrades showed similar fragmentation.

McKinsey’s latest State of AI report notes that firms with shallow integration see negligible gains, while those that redesign workflows achieve measurable impact. The lesson is clear: intelligence alone is insufficient; it must be embedded within a structured representation of work.

The next frontier will not be a prettier copilot or longer prompt. It will be a formal layer that captures identity, state, permissions, provenance and business semantics in a machine‑readable yet human‑auditable form. When such invariants exist, ecosystems of extensions, marketplaces and standards can emerge—turning enterprise AI from artisanal to industrial.

Reported by Julian Reed (Consumer Electronics Expert).