When not to use a language model

Language models are good at judgment over messy language. They are expensive, non-deterministic, and overconfident at everything else. The most useful thing an AI studio can do is refuse a model when a constraint, a queue, or a form will do.

Ask: is the input well-typed? Is the output a closed set? Is the cost of a wrong answer higher than the cost of a human? If yes, yes, and yes — write software. Use a model for the residue: PDFs, emails, conversations, documents that never agreed on a schema.

Extraction with a schema, classification with an abstain path, drafts that a human sends, retrieval that cites. Always: traces, evaluation, and a tool allowlist. An agent that can 'do anything' is a security incident with a progress spinner.

A trigger, fetched context, a typed response, policy, side effects that are idempotent, and a review UI for anything that writes to a system of record. That list is the product. The model is a component.