The failure mode confidence thresholds exist to prevent
An AI agent that always answers, regardless of how sure it actually is, will eventually answer a billing dispute, a safety question, or an account-specific edge case with confident, plausible-sounding text that happens to be wrong. That single bad answer costs more trust than a dozen good ones build — customers don't average their experience, they remember the time the "chat" told them something false.
A confidence threshold is the mechanism that prevents that: the AI agent estimates how well the visitor's question matches what it actually has grounded knowledge of — your documents, your product data, your policies — and only answers directly when that match clears the bar the workspace has set.
What happens instead of a bad guess
Below the threshold, the AI doesn't stall or repeat itself — it hands off to a human agent, and the handoff carries the full conversation summary with it: what the visitor asked, what the AI already tried, and why it stopped. The human isn't starting cold; they're picking up a conversation that's already been triaged.
This is also exactly what the `handoff.requested` webhook event exists for — a business that wants its own escalation logic (paging a specific person, opening a ticket in another system) can hook into the exact moment the AI decided it needed a human, not a fixed timeout or a keyword match.
Where the threshold should sit, in practice
There's no universal right number — a workspace answering general product questions can afford a lower bar than one handling account-specific billing issues, where a wrong answer has real financial consequences. The right instinct is to start conservative (more handoffs than strictly necessary) and loosen the threshold only as you build confidence in what the AI's knowledge base actually covers well, rather than the other way around.
Feeding the AI better, more current knowledge sources through RAG is usually a more effective lever than tuning the threshold itself — an AI agent with genuinely good source material naturally clears the confidence bar on more real questions, without you having to loosen the safety margin to get there.
The metric that actually matters
Resolution rate isn't a vanity number — on iNTELIGENCIA VIVA it's literally what you're billed on, since a resolution is a conversation the AI closed without a human stepping in. That alignment is deliberate: it means the incentive is for the AI to be honestly good at the questions it handles, not to maximize how often it answers regardless of accuracy.