The quiet rise of small, boring, useful AI

The AI that actually changed my week doesn't impress anyone in a demo. It takes unstructured text and turns it into structure. That's it. And it's worth more to me than any flashy generation demo I've built.

Examples from my own stack: extracting structured data from messy form submissions so they land in a database instead of an inbox. Classifying incoming issues into buckets so I see what matters. Summarising long documents into the five facts I needed. Routing messages to the right part of a workflow.

Why the boring stuff wins

Generation is probabilistic, which means errors are visible and embarrassing. Classification and extraction are probabilistic too, but you can wrap them in validation: check the output against a schema, fall back to a human when confidence is low, log everything. A 95%-accurate extractor with a 5%-to-human fallback is a fully reliable system. A 95%-accurate writer is just... sometimes wrong.

The pattern worth copying

• small scope: one input type, one output schema
• cheap model: the task doesn't need the flagship
• hard validation: reject anything the schema doesn't accept
• human path: every failure lands somewhere a person can catch it

The future, at least the part that pays, is a thousand small systems exactly like this.