Skip to main content
Essay
AIFebruary 22, 20264 min read

When not to use AI

Five situations where a regular form, filter, or rule engine beats an LLM call.

Not every problem needs an LLM. Here are five situations where we'd push back on an "AI feature" request.

1. Deterministic lookups

If the answer is in your database, query the database. Don't embed documents and hope RAG finds the right row.

2. Simple classification with clear rules

"If subject contains X, route to team Y" doesn't need a model. Regex and rules are faster, cheaper, and debuggable.

3. Format conversion

JSON → CSV, markdown → HTML — use a library, not GPT.

4. Features users don't want

A chat sidebar nobody asked for isn't fixed by a better model. Kill the feature.

5. When latency budget is under 200ms

LLM calls are seconds. If the UX requires instant feedback, AI isn't the tool.