Shipping AI without the theatre

The industry is full of demonstrations. A box that answers questions about a PDF is not an operating system for knowledge. The moment you care about the wrong answer, you need retrieval boundaries, a test set, and a person who can override.

We start with the job. Draft this. Classify that. Route the document. If the job cannot be named, the model will be asked to be a colleague, and it will fail in ways that look like personality.

Evaluation is not optional. A golden set of fifty examples, rerun when the prompt or the model changes, is more valuable than another demo. If quality has no number, it will only have a story.

Permissions belong in the retrieval layer, not in the prompt. A model should not see what the user cannot see. Logs should not become a second copy of the file.

The unfashionable conclusion: many problems still want a rule, a classifier, or a better form. Generative models are one instrument. Use them when they are the cheapest honest method.