The most important thing an AI receptionist does is recognise the call it must not handle.
Escalation is not a safety net bolted on at the end. It is the first thing worth designing, and it is specific to the business.
An escalation rule says: when this happens, stop handling the call and get a human, immediately. The important design property is that it should not be a judgement the system makes fresh each time. It should be a trigger someone decided in advance, written down, and biased deliberately toward firing too often, because the cost of a false alarm is an unnecessary phone call, and the cost of a miss is the reason the business was nervous about automating its phone at all.
Why should escalation be a rule rather than a judgement?
Because judgement is probabilistic and escalation needs to be reliable. A model asked to assess whether something is an emergency will be right most of the time, and the failures will be concentrated in exactly the unusual phrasings that characterise real emergencies.
The asymmetry is what settles it. A rule that fires on a mention of chest pain will sometimes escalate someone describing a movie. That costs a receptionist ten seconds. A model that weighs context and decides the caller probably is not in danger will occasionally be wrong in the other direction, and that failure is not recoverable.
So the useful design is: deterministic triggers for anything with a serious downside, and model judgement for routing and prioritisation where being wrong is cheap. Not model judgement everywhere with a confidence threshold, which is how these systems are often assembled.
What does escalation look like in different businesses?
Completely different, which is the reason it cannot be a default. The hardest decision on a business phone varies by industry, and the escalation rule is that decision written down.
| Industry | Escalate immediately on | Why |
|---|---|---|
| Veterinary | Any indication of a possible emergency | The right answer is sometimes 'go elsewhere now', and it is time-critical |
| Med-spa / clinic | A complication after treatment | Clinical assessment, and never something to leave in a queue |
| Law firm | Anything with a limitation period or a court date | A missed deadline is unrecoverable |
| Property management | Flood, fire, gas, no heat | Legal obligations and property damage both run on a clock |
| Trades | Gas smell, water, no heat in winter | Safety, and in any freezing climate, pipes |
None of these is derivable from a generic template. Each comes from the business knowing what its worst call looks like.
How do you write one that works?
Name the trigger, name who it goes to, name what happens if that person does not answer, and name what the caller is told meanwhile. A rule missing the third element is the most common failure, and it is the one that turns an escalation into a dropped call.
The other rule worth having is a universal one: if the caller asks for a person, they get a person. It costs occasional efficiency and it removes the single strongest objection anyone can raise to the agent existing.
- The trigger. Specific words or conditions, not a category. "Mentions gas" is a rule; "seems urgent" is a hope.
- The destination. A named person or rota, not a general line, and a real one that is actually monitored.
- The fallback. What happens when nobody answers. This is where escalation designs usually fail.
- What the caller hears. A caller being escalated should know it, and should not be left in silence while the system works.
Escalation should be a rule someone wrote down rather than a judgement the system makes fresh each call, because a false alarm costs ten seconds of a receptionist's time and a miss costs the reason the business was nervous about automating the phone in the first place.
The first question we ask a new client is what their worst call looks like. The escalation rules are that answer, written down.
Start free pilotWhat this page does not publish
The examples are design patterns rather than any client's live configuration.
- No client rule sets. Real escalation configurations belong to the businesses that specified them and are not published here.
- Not clinical or legal triage guidance. What should count as an emergency in a regulated practice is a question for that practice and its professional obligations.
Questions this raises
- What is an escalation rule?
- A predefined trigger that stops the automated handling of a call and routes it to a person immediately. It is specified in advance rather than decided by the system during the call.
- Why not let the AI decide what is urgent?
- Because its errors cluster in unusual phrasings, which is exactly how real emergencies get described. A deterministic trigger is less clever and far more reliable where the downside is serious.
- Should escalation rules err toward firing too often?
- Yes, deliberately. A false alarm costs a short unnecessary call. A miss costs the thing the business was worried about, and cannot be undone.
- What is the most commonly missing piece?
- The fallback: what happens when the escalation destination does not answer. Without it, an escalation can drop a caller more thoroughly than no escalation would have.
- Should a caller always be able to reach a human?
- In our view yes, as a universal rule. It costs a little efficiency and removes the strongest objection to having an automated line at all.
Sources
- Recording of Customer Telephone Calls: guidance for organizationsOffice of the Privacy Commissioner of Canadaupdated 2018-04-18, consulted 2026-08-09
Want an AI receptionist built around these rules rather than despite them?
We build a working receptionist on your calls, your booking rules and your escalation rules, and you call it yourself before committing to anything. About 14 days, no call limit during the pilot.
Start free pilot