When Should AI Escalate to a Human?
An AI support agent should escalate to a human whenever it is below its confidence threshold, when the knowledge base does not cover the question, when the ticket is angry or emotional, when the account is high-value, or when the requested action is on a forbidden list such as refunds, fraud, or legal commitments. Escalation is a rewarded outcome, not a failure. The safe ceiling on automation is set by the hardest ticket you let the AI close.
The escalation triggers
- Low confidence: the calibrated composite score falls below the threshold.
- No grounding: retrieval finds no solid source, so answering would mean guessing.
- Negative sentiment: an angry or distressed customer should not get a bot auto-close.
- High-value account: enterprise or sensitive accounts held to a higher bar.
- Forbidden action: refunds over a threshold, chargebacks, legal, or ownership changes.
Why a zero escalation rate is a red flag
A bot that never escalates is not a triumph, it is a bot answering questions it should have passed up. Neural networks are miscalibrated toward overconfidence, so forcing the front line to handle everything produces wrong answers and reopens. The right escalation rate is the one where the machine handles what it should and hands up what it should not.
How Resolve escalates cleanly
When Reva hits any of these triggers, it escalates with a decision card rather than a bare handoff, carrying the issue summary, the relevant knowledge, a suggested action, and the guard that tripped. Intent guardrails route bugs, outages, and refunds to a human without confirming a fix or moving money, and the sentiment guardrail keeps angry tickets out of auto-close.