How to Set Up Emergency Call Escalation That Actually Wakes Someone

Every service business eventually writes the same sentence into its after-hours greeting: if this is an emergency, we'll get someone to you. Almost nobody writes down what happens next. The gap between that promise and a tech actually standing in a flooded basement is a set of unglamorous mechanics - triggers, timers, acknowledgements, a rotation, and a 911 boundary - and when an escalation fails at 2 a.m. it is almost never because the software broke. We build AI receptionists that run these ladders, so this is written from the failures: here is how to specify one that works, including the parts that look fine on paper and fall over in the dark.

The short answer

An escalation policy is five things written down: what qualifies as an emergency, who gets it, how fast they have to acknowledge it, what happens when they don't, and what the caller is told while all of that runs. Your answering service or AI receptionist executes it; you own it. If you only want the conceptual version - can an AI do this at all - that is answered here. This page is the specification you hand to whoever configures it.

The five parts of an escalation policy

Writing triggers that don't over-fire

The instinct is to hand over a list of words - emergency, urgent, flooding, ASAP. It fails in both directions, reliably:

• False positives."This is an emergency, my garbage disposal is jammed" is a sincere sentence and a three-word trigger. Wake a tech twice for that and they start ignoring the pager - the pager-fatigue dynamic documented at length in Google's SRE book chapter on being on-call, which is about servers but describes human beings.
• False negatives.A calm caller saying "there's water coming through the ceiling and I can hear it running" uses none of the trigger words and is the worst call of the night.

Write conditions instead. Three lists, on one page:

1. Always escalate. Described as situations, in the language a caller would actually use: water running that cannot be stopped; no heat with outside temperatures below a number you name; no power to the whole property; sewage backing up; a smell of gas; anyone hurt; a security failure like a door that will not lock.
2. Never escalate.Equally explicit, and equally important: cosmetic damage, one appliance, one outlet, noise complaints, billing questions, scheduling, "can someone come look at it this week." This list is what keeps your on-call person willing to stay on call.
3. Ask one question, then decide.The middle band, and the part that separates a good script from a checklist. "No hot water" is a Monday problem in a house and an emergency in a twelve-unit building; "the AC is out" is a Monday problem in April and an emergency in a heat advisory with an elderly resident. Write the disambiguating question rather than guessing at the answer.

This is prompt work, and the general craft of it - being specific, bounding the model, writing what not to do - is in our guide to AI receptionist prompts.

The ladder and the timers

A ladder is a sequence of targets with a timer between each. The shape that works for most small operations:

Three rules about timers that are learned the hard way. Work backwards from the promise: if the caller was told fifteen minutes, the whole ladder has to finish inside fifteen minutes with room for the callback. The last rung must be a human, not a voicemail box - a ladder that terminates in a mailbox is a ladder that terminates in nothing. And keep escalating in parallel, not instead: when the ladder moves on, the previous rung should still be able to pick it up; someone waking up ninety seconds late is a good outcome, not a race that has already been lost.

Escalate on silence, not on "no answer"

The fix is to define delivery as an explicit acknowledgement, not a connected call: press 1 to accept, reply YES, tap the notification. Anything else - answered and silent, declined, voicemail, no answer - is treated identically as a failure and the timer keeps running. This one change fixes more broken escalation setups than any other, and it costs one line of configuration.

Its close cousin is the phone itself. iOS Focus modes and Android Do Not Disturb will silence your paging number by default, and every on-call person has their phone in one of them at 2 a.m. - add the paging number as an emergency-bypass or favorite contact on every on-call phoneand confirm it with a real test call while the phone is silenced. It is the single most common cause of "the system didn't page me."

An on-call rotation people will actually keep

• One owner per window."Whoever's around" means nobody. A window has exactly one name and exactly one backup.
• A published calendar, not a group chat.If the schedule lives in someone's head or in a message from three weeks ago, it is wrong by now.
• An explicit handoff. The outgoing person says what is still open. Rotations fail at the seam more than in the middle.
• Protect the load.The SRE literature's core finding transfers cleanly: on-call people who are woken too often stop responding well, and the fix is fewer false pages, not more discipline. Every false escalation you eliminate makes the real ones faster.
• Pay for it, visibly. A stipend for the window turns on-call from an imposition into a shift. It also makes you honest about how much of it you are asking for.

The 911 boundary

There is one hard line in every escalation policy, and it is worth being precise about why.

An automated system must never attempt to place a 911 call on a caller's behalf. Emergency calls are routed to a dispatch center based on the calling device's validated location - the FCC calls this dispatchable location, the street address plus the suite or apartment needed to actually find the caller. A call placed by your phone system carries your phone system's registered location, which means it reaches the wrong dispatch center and reports the wrong address. The correct script, in every trade, is the same three beats: tell the caller to hang up and dial 911 now, say nothing else that keeps them on the line, and log the call. Same for a gas smell, with the addition of "get outside first, and call from outside."

Related and easy to miss if you are rebuilding your phone system anyway: under Kari's Law, a multi-line telephone system in the US must let anyone dial 911 directly, without a 9 prefix or any other code, and must notify a front desk or equivalent when someone does. The FCC's MLTS requirements page is the authority. It is a compliance item, not an optional configuration, and the moment you are most likely to break it is the moment you are rewiring routing for after-hours coverage.

What the trigger list looks like by trade

The mechanics above are universal; the "always escalate" list is not. Broad strokes, with the detail in each trade's own guide:

Closing the loop with the caller

Half of what people describe as "the emergency was handled badly" is actually "nobody told me anything." The caller is standing in the dark with no idea whether their call went anywhere. Three cheap things fix it:

• Say the time out loud."Our on-call technician is being paged now and will call you within fifteen minutes" - then keep it. A concrete promise you hit beats a warm promise you miss.
• Text the confirmation. What was reported, who is coming, what to do in the meantime - shut the valve, kill the breaker, stay out of the room. The caller has something to hold onto.
• Give them the escape hatch. Anyone who asks for a person gets a transfer, and the script says so before they have to ask twice. What happens when the assistant is out of its depth is worth deciding explicitly too - here is how that should behave.

Testing it (including the 2 a.m. drill)

1. Call your own line and describe a real emergency, in the words a caller would use, without saying "emergency." It should escalate anyway.
2. Call and describe something trivial while insisting it is urgent. It should not escalate.
3. Let rung one ignore the page deliberately. Time the ladder. Confirm rung two fires and that rung one can still pick up.
4. Answer the page and say nothing. The ladder must keep going. If it stops, you are escalating on connection rather than acknowledgement - fix it before anything else.
5. Run one drill at 2 a.m. With phones in their real overnight state - Focus on, chargers across the room, everyone asleep. This is the only test that tells you the truth, and it is the one nobody runs until after the first miss.
6. Test the 911 script.Say "I smell gas." The correct response is short, immediate, and ends the call.

Tuning it with two numbers

Review weekly for the first month, then monthly. Two numbers matter and only one of them will find you on its own:

• False escalations.Techs woken for something that could have waited. You will hear about these. Each one is a missing entry on the "never escalate" list or a missing disambiguating question.
• Missed escalations. Nobody reports the emergency that quietly became a Monday message - so audit a sample of non-escalated overnight calls deliberately, every week. A trigger list producing zero false escalations is almost certainly missing real ones; the target is a small, visible false-positive rate, not zero.

Also track acknowledgement time. If rung one is routinely acknowledging at three and a half minutes on a four-minute timer, the timer is holding the whole thing together by luck.

Where this fits

Escalation is the sharp end of after-hours coverage, not the whole of it - the shape of the rest is in the after-hours answering service guide and the 24 hour guide. If you are setting this up for the first time, the practical sequence is: write the three trigger lists, wire the ladder with acknowledgement, point your line at it with conditional call forwarding, and run the 2 a.m. drill before you rely on it. Our plans are on the pricing page and run month-to-month, which means you can test a full ladder on real calls for two weeks before committing to anything.