The failure mode handoff exists 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.
The tempting approach is to ask the model how sure it is and hand off below a number. We don't do it that way: research (Xiong et al., ICLR 2024) shows that the confidence a model reports about itself is systematically inflated, and a single instruction hidden in a message is enough to make it report total certainty. At iNTELIGENCIA VIVA, handoff is decided by rules the code checks, not by a number the AI writes about itself.
What triggers a handoff
The conversation passes to a person on your team when the visitor asks for one, when there are signs of crisis, when the visitor shows frustration, when the answer isn't backed by the knowledge you gave it, or when the AI couldn't produce a useful answer. The handoff carries the full conversation summary: what the visitor asked, what the AI already tried, and why it stopped. The human doesn't start cold.
If nobody on your team can attend at that moment — outside business hours, nobody online, or a chat with no people behind it — the AI doesn't leave the visitor waiting: it says so clearly and offers the business's other channels, or keeps helping with what it does know.
That's what the `handoff.requested` webhook event is for: a business that wants its own escalation logic (alerting a specific person, opening a ticket in another system) can hook in at the exact moment the conversation passed to a person, not on a fixed timeout.
What you do control, in practice
There's no universal bar — a business answering general product questions can tolerate more than one handling account-specific billing cases, where a wrong answer has real consequences. What defines the behavior is in your hands: the assistant's instructions (what it must never answer and which cases always go to a person), your business hours, and whether the chat has people behind it.
Feeding the AI better, more current knowledge sources is the most effective lever: an assistant with genuinely good source material can rely on it for more real questions, and hands off less without you having to loosen any safety margin.
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.