Start with decisions AI should not own
An escalation rule should name the trigger, the person to call, and what happens if no one answers. “Use judgment” is not a rule. “If the caller reports smoke, follow the safety script and try the on-call person” is specific enough to test.
Keep a person responsible for safety decisions, clinical or legal advice, payment disputes, angry callers, exceptions to policy, and any request the script cannot support.
Four useful rule types
- Immediate safety script — Give only the approved direct instruction, such as contacting emergency services or a utility, then use the urgent path your team chose.
- Time-sensitive follow-up — Send court dates, same-day access, active property damage, or another defined deadline to a named team.
- Caller-requested transfer — If policy allows, try the chosen person rather than forcing more intake.
- Out-of-scope fallback — Capture a short message and callback window instead of guessing.
Define the unanswered path
A transfer attempt is not a successful transfer. Decide what happens when the target is busy, unavailable, or rejects the call:
- try one approved backup person or number;
- capture a callback request with the context already collected;
- set an honest expectation without inventing an ETA;
- record the attempted outcome for owner follow-up.
Avoid loops. The caller should not bounce between the same transfer target and the AI.
Test the awkward calls
Test silence, a caller who changes topics, an unsupported language, and a false emergency keyword. Then test a real safety concern, an angry customer, and an unanswered transfer. Check both the caller experience and the resulting record.
See the AnswerCheap call-transfer controls and the setup and human-transfer limits before routing real callers.