Every action NUA AI can take automatically is gated by a confidence score, a specific mechanism, not a vague notion of "the AI decides." Here's what actually feeds it.
What raises a confidence score
- Strong historical precedent: this exact pattern has occurred and resolved predictably many times before
- Low downside if wrong: a small reorder adjustment carries less risk than a large one
- Consistency with recent, similar automated actions that produced good outcomes
What lowers it
- Limited historical data: a new venue, or a genuinely novel situation
- High potential downside: a large financial commitment or a guest-facing change
- Conflicting signals across modules that don't point to one clear answer
Why it errs toward asking, not acting
The scoring is deliberately conservative. When signals are mixed, the system defaults to flagging for a human rather than guessing. A false "ask me" costs a few seconds of an operator's attention. A false "just do it" on a genuinely uncertain call can cost real money or a guest's trust.
It's built to be right when it acts alone, and honest about uncertainty when it isn't, not to maximise how much it automates for its own sake.
This is what makes automation earn trust
Operators generally don't distrust automation because it acts. They distrust it because it's opaque about when it's guessing. A visible, reasoned confidence score is the difference.