Automation that never talks to your customer

The most reliable AI workflow I’ve shipped sends zero AI-written messages to customers. It classifies feedback, summarizes it, logs it, and — when sentiment turns negative — alerts a human within seconds. The human replies. That’s the whole design.

This sounds unambitious until you count the failure modes it deletes. No hallucinated apology promising a refund policy that doesn’t exist. No tone-deaf reply to an angry customer. No audit problem, because every AI output lands in a spreadsheet a human reviews, not an inbox a customer reads.

The design rule I now apply to every automation: put the AI on the reading side, keep the human on the writing side — until the eval suite proves the writing side is safe. Reading-side AI (classify, extract, summarize, route) fails cheap: a mislabeled row. Writing-side AI fails expensive: a customer, a contract, a reputation.

When a client asks for a "fully automated" agent, the real question is where the approval gate belongs: before every send (early days), before novel situations only (once evals are green), or after-the-fact review (mature, measured systems). Moving down that ladder is earned with evidence — never assumed.