Guide · 10 min read · June 23, 2026
Designing escalation rules for an AI receptionist
How to decide when an AI receptionist hands a call to a person, where that call goes, and what happens when nobody is available — with rules you can copy.
Escalation rules tell an AI receptionist when to stop handling a call and hand it to a person. Good rules cover four things: the triggers (caller asks for a human, urgent keywords, repeated misunderstanding, requests outside the script), the destination (a named person, ring group or queue, varying by time of day), the fallback when nobody answers (message with transcript, callback, on-call cell), and what context travels with the call.
Why escalation is the part worth designing carefully
An AI receptionist is judged on its worst call, not its average one. Callers forgive a slightly stiff greeting; they do not forgive being trapped in a loop when they needed a person. Escalation rules are the mechanism that prevents the trap, and they deserve more design time than the greeting or the booking script.
The temptation is to write one rule — “if in doubt, transfer to reception” — and stop. That works until reception is busy, closed or on the other line, at which point the caller is transferred into voicemail and the receptionist has done worse than a plain answering machine. Real escalation rules answer three questions in sequence: when does the call leave the receptionist, where does it go, and what happens if that destination cannot take it.
Triggers: when the receptionist stops handling the call
Triggers are the conditions that end the receptionist’s attempt and start a hand-off. Write them as a list, in priority order, and keep the list short enough that you can remember it.
- Explicit request: the caller asks for a person, an operator, or someone by name. Honour it on the first ask.
- Urgency language: words your business defines as urgent — “emergency”, “no heat”, “flooding”, “chest pain”, “locked out” — trigger an immediate transfer with no further questions.
- Repeated misunderstanding: two failed attempts to identify the request, or the caller repeating themselves with rising frustration.
- Out-of-scope request: anything the script has no answer for, such as a complaint, a billing dispute, a request for a quote on non-standard work, or a media enquiry.
- Identity-sensitive requests: changes to an account, cancellations with penalties, anything that needs the business to verify who is calling.
- Silence or noise: no usable speech after two prompts, which usually means a bad line or a pocket dial, and should end in a polite close rather than a transfer.
Destinations: where the call goes, and when
A destination is not one number. It is a small table that changes with the time of day and the day of the week, because the person who should take an escalated call at 10 a.m. Tuesday is rarely the person who should take it at 9 p.m. Saturday.
For business hours, the usual destination is a ring group or a queue rather than an individual extension. A group of three people who can all handle an escalated call is far more likely to answer than one named person who may be on another line. For after hours, the destination is typically an on-call cell phone for urgent triggers and a message with transcript for everything else.
Some businesses want a middle tier: an escalated call first tries a specific person (the practice manager, the dispatcher), and only if they do not answer within a set number of rings does it move to the group. That is reasonable, provided the timeout is short — fifteen to twenty seconds — so the caller does not sit through a long unanswered ring.
Whatever the destination, the receptionist should announce the transfer to the caller in plain words: “I’m connecting you with our service team now.” Silence followed by ringing feels like being dropped.
Fallbacks: what happens when nobody picks up
Every destination needs a fallback, and the fallback must not be a dead end. If the ring group does not answer within the timeout, the receptionist should come back on the line, say so honestly, and offer a choice: leave a message that will be transcribed and sent to the team, or request a callback at a specific number.
For urgent triggers after hours, the fallback is usually a second on-call number, and then a message flagged as urgent that is delivered to more than one person. The point is that an urgent call never terminates in an ordinary voicemail box that will be checked at 8:30 the next morning.
Write down the fallback chain for each destination as a sequence: try A for twenty seconds, then B for twenty seconds, then offer transcript or callback. Test it by calling the number yourself at 7 p.m. on a weekday and at noon on a Sunday, and following the path a caller would follow.
Context: what travels with the call
An escalation is only useful if the person receiving it knows why the receptionist handed the call over. At minimum, the transfer should carry the caller’s name and number, what the receptionist understood the request to be, and which trigger fired. A whispered announcement — “transfer from the receptionist, caller Marie, says no hot water” — before the call connects lets the person answer with the right first sentence.
For messages and callbacks, the transcript should include the same fields plus the time, the number called, and any partial data the receptionist collected. A structured message — name, number, reason, urgency — is far easier to act on than a paragraph of transcript, so configure the receptionist to summarize, not just transcribe.
Where the receptionist is connected to a calendar or a customer record, the escalated call can also carry the account or the next appointment. That should be confirmed during planning, since availability depends on the integration.
A worked example
Consider a plumbing company with a dispatcher, three technicians and an owner. Their rules, written in a single page, look like this. During business hours: urgency words transfer immediately to the dispatcher, then to the owner after twenty seconds, then take an urgent message delivered to both by text. Any request for a person goes to the dispatcher, then to a message. Bookings and quotes for standard jobs are handled by the receptionist; quotes for non-standard jobs are captured as messages for the owner.
After hours: urgency words transfer to the on-call technician’s cell, then the owner, then an urgent message to both. Everything else is captured as a message and the caller is told the office opens at 7:30 a.m. and someone will call back before 9.
None of that is complicated. What makes it work is that every path ends somewhere a person will actually see, and that the company tested it before the first real caller did.
Questions that come up during setup
These two decisions are the ones most teams debate longest.
Should the AI receptionist try to resolve the call before escalating, or escalate first?
Let it try once. A caller who says “I need to talk to someone about my invoice” has already told the receptionist enough to route the call to accounting with context, and a single clarifying question — “is this about a payment or a question on the amount?” — makes the transfer more useful. Two failed attempts is the limit. Beyond that, insisting on resolution frustrates callers more than a fast hand-off would.
What should happen to an escalated call at night when the on-call person does not answer?
It should never end in a standard voicemail box. The receptionist should return to the caller, explain that the on-call person is not available, and take an urgent message delivered immediately by text and email to at least two people. Some businesses add a second on-call number before the message step. Test the chain by calling it yourself after hours and timing how long a caller waits.
Want to see your own call flow?
Three minutes of questions, a recommended setup.
Build My Phone System