A surprising number of “AI projects” should have been a cron job and a SQL query. The reverse is also true — some “automation” projects fail because rules can’t keep up with the variability of real input.
A quick decision rule
- Deterministic + low variance → script it.
- High variance, judgment required, tolerable error rate → AI-assisted.
- Compliance-critical, zero-error requirement → AI as suggestion, human as decision.
Pattern we like
Pair an LLM with a structured output schema and a deterministic validator. The LLM handles ambiguity; the validator enforces correctness. It’s boring. It works.