Not everything needs an LLM
Summary
The article argues that not every task requires an LLM and advocates using decision models (like Jev) to gate, route, and simplify AI tasks. It covers practical patterns such as gating expensive LLM calls, routing to cheaper models, and using decision models for face-value checks, with real-world results from PostHog showing significant cost and latency reductions. It also discusses when to rely on LLMs for open-ended work and how to calibrate thresholds and workflows across models.