How we decide when the product should ask a human instead of guessing

Confidence is a product decision, not a model setting

Every automated system eventually meets a case it hasn’t seen. What it does next determines whether people trust it — and trust, once lost, does not come back on the next release.

We decided early that Lumora would rather interrupt you than be quietly wrong. That sounds obvious written down. In practice, it means accepting a flag rate that looks worse on a slide than full automation would.

Where we draw the line

An agent acts on its own when the situation is clear and the action is reversible. It asks a person when either of those stops being true: when the data is contradictory, or when a mistake would be hard to undo.

Roughly one run in twenty gets flagged. We have had customers ask us to lower that. We have never had one ask us after an incident.

The compounding part

Every answer a person gives to a flagged case becomes context for the next one. The exception rate falls on its own over the first few months, without anyone tuning anything — because the system is learning your judgment, not a general one.