No signs of AI payoff outside the tech sector

Discussion overview

Commenters mostly treated the missing economy-wide payoff as a timing and deployment problem, not a clean verdict on AI’s usefulness. Software is the easiest first target: code has rich training traces, established tooling, automated tests, and engineers who can spot when a model goes off the rails. Physical industries, by contrast, need capital investment, workflow redesign, safety controls, and institutions willing to absorb risk. Several readers compared this lag to earlier general-purpose technologies whose gains appeared only after organizations rebuilt around them.

But the thread was not uniformly patient. Skeptics argued that widespread instant access to LLMs weakens the infrastructure-lag analogy, and pointed to customer support, insurance administration, warehouses, and education as domains where polished demonstrations have not become reliable production automation. A recurring explanation was validation: coding agents can be fenced in by tests, while many real-world decisions have slow, ambiguous, or costly feedback. Some also worried that AI is accelerating low-value corporate output—more messages, documents, and dashboards—rather than outcomes anyone measures.

There were concrete signs of value outside pure software, including clinical transcription, internal business automation, and falling costs in video and 3D asset creation. Yet anecdotal workflow failures, especially document extraction and unreliable chatbots, kept the mood cautious. The broad conclusion was that useful capabilities exist, but dependable, economy-moving adoption remains unproven.

Points of disagreement

The sharp dispute was whether AI’s weak non-tech productivity signal is a temporary diffusion lag or evidence that current systems cannot clear real-world reliability thresholds. Optimists argued that firms need years to redesign processes around a general-purpose technology, and that better models are only now becoming usable enough for broad automation. Skeptics countered that LLMs are already universally accessible, so the lack of production-grade customer service, operations, and education gains points to hallucinations, edge cases, and expensive human verification rather than mere institutional inertia.