The Architect and the Workers: Most Daily AI Tasks Don't Need a Frontier Model

Here is the number the entire centralized-AI buildout rarely mentions. Across published enterprise routing research, only 25–35% of real-world queries require frontier-model capability.

In one production financial-services deployment documented in the academic literature, over 70% of queries were routine tasks — entity extraction, FAQ matching, structured formatting — that smaller models handle at equivalent quality. NVIDIA researchers argued in a 2025 position paper ("Small Language Models are the Future of Agentic AI") that 40–70% of enterprise AI tasks can run on sub-10-billion-parameter models.

The routing literature quantifies it precisely. The RouteLLM framework (LMSYS, 2024) demonstrated that intelligently routing between a strong and weak model achieves 95% of GPT-4's benchmark performance while invoking GPT-4 for only 14% of queries. Follow-on systems report 40–85% cost reductions while retaining 96–100% quality on structured tasks. Stanford's FrugalGPT work estimated 50–90% of inference spend is wasted on over-provisioned capability.

The difficulty distribution of real work is heavy-tailed: a small number of genuinely hard tasks, and a long tail of routine ones currently being sent, at frontier prices and frontier wattage, through gigawatt facilities.

The architecture this implies

We call it the architect-worker pattern, and we believe it is the correct end-state for personal and organizational AI:

Frontier models as master architects

The rare, hard 15–30%: novel reasoning, complex planning, cross-domain synthesis, designing the workflows and reviewing the edge cases. This work justifies datacenter economics and datacenter energy.

Local models as the workforce

The routine 70–85%: summarization, extraction, drafting, classification, retrieval-augmented answers over your own documents, transformation of your own data. This work runs on the device where the data already lives — at desk-lamp wattage, with nothing transmitted anywhere.

The unsolved research problems live at the seam, and they are ours: routing that runs locally (the decision "is this task hard?" must itself not leak the task to the cloud — see published work on zero-shot confidence estimation approaches, where a small model's own token log-probabilities can rival trained routers), quality assurance on the local tier (bounded-hallucination and attestation work), and honest accounting (our EQAT benchmark, measuring whether the energy savings survive contact with real workloads and real quality requirements).

If the routing literature's numbers hold for consumer and professional workloads — a question our planned adoption study addresses directly — the grid implication is enormous: the majority of interactive AI demand growth could, in principle, be absorbed by hardware already purchased, already powered, and already on people's desks.