Setting Up an AI Answering Service That Never Misses an Emergency

Write down exactly what counts as an emergency, give the AI trigger words and screening questions for each one, and route every match by live transfer to the on-call phone. Back that with a text alert, a second person if nobody answers within a set time, and a forwarding rule for outages. Then prove it works with staged test calls every month.

"Never misses" is a design goal, not a feature you can buy. In practice the AI recognising an emergency is rarely the weak link. What fails is everything after it: the engineer's phone on Do Not Disturb, a transfer that rings out, a text nobody reads, or the AI platform itself being down at 2 a.m. So most of the work here is building a chain where each missed step triggers the next one.

Follow me on Instagram@sagnikteaches

Write the emergency list before you open any software

Every AI answering service asks you for "escalation rules", and most firms type in the word "emergency" and move on. Callers rarely use that word. They say "there's a funny smell" or "Dad's house is freezing". Start with a table of real situations, the question that confirms each one, and what the AI must do. Here is an illustrative version for a six-engineer heating-and-cooling firm:

Connect on LinkedInSagnik Bhattacharya

The fixed lines matter. You don't want an AI improvising safety advice, so give it exact wording you have approved and tell it to read those lines verbatim and add nothing technical.

Subscribe on YouTube@codingliquids

Other trades need their own list. An electrician's might include a burning smell from a socket or consumer board, sparking, exposed wiring after damage, and total power loss where someone relies on medical equipment. A roofing contractor's covers storm damage with water coming in, and loose tiles or slates above a path or doorway. A plumbing firm's centres on burst pipes and water that can't be stopped; triaging emergency plumbing calls out of hours goes into that list in detail.

The screening script the AI follows word for word

Once the list exists, turn it into instructions. Whatever service you use, the escalation section of its set-up should read something like this:

EMERGENCY SCREENING (follow in order, every call)
1. Ask what the problem is. Listen for: gas, smell, eggy, fumes,
   alarm, CO, carbon monoxide, dizzy, headache, leak, pouring,
   ceiling, sparks, burning, no heat, freezing, elderly, baby, unwell.
2. If ANY smell is mentioned, ask: "Is it a smell of gas, or a burning
   or hot smell?" If the caller is unsure, treat it as gas.
3. For gas or CO: read the fixed safety line exactly, then transfer.
4. For no heat: ask about vulnerable people. If yes, transfer.
5. If you are unsure whether it is an emergency, treat it as one.
6. Never give repair advice. Never promise an arrival time.
7. Before transferring, say: "I'm putting you through to our on-call
   engineer now. If we get cut off, they will call you back on this
   number."

Rule 5 is the one to argue about internally and then keep. False alarms cost an engineer a phone call; a missed gas leak costs far more. You can tune the false-alarm rate down later using transcripts.

An illustrative transcript shows why rule 2 exists:

Caller: The boiler's making a banging noise and there's a bit of a smell.
AI: Is it a smell of gas, or a burning or hot smell?
Caller: Not sure. Sort of eggy?
AI: Please leave the property now, don't use light switches or flames, and call the gas emergency number from outside. I'm putting you through to our on-call engineer now.

Without the follow-up question, "a bit of a smell" alongside a noisy boiler is exactly the kind of call that gets booked for Tuesday. Writing call scripts for the rest of the line is covered in call scripts and escalation rules for an AI receptionist.

The escalation chain after "this is an emergency"

This is the part that decides whether the line truly never misses. Each step has a timer, and a missed step triggers the next. The timings below are illustrative; pick yours and write them down.

1. Warm transfer to the primary on-call engineer's mobile, ringing for about 25 seconds.
2. If no answer: the AI tells the caller "I'm alerting our engineer now; you'll get a call back within 10 minutes", texts the engineer a summary (name, address, problem, callback number) and rings once more.
3. After 10 minutes with no acknowledgement: call and text the backup engineer.
4. After 20 minutes: call the owner or manager.
5. Caller confirmation: the caller receives a text with a reference and the promise of a callback, so they aren't left wondering whether anyone heard.
6. Log everything: every attempt, time and outcome, for the morning review.

A written rota keeps the chain honest:

The "phone check" column exists because of one very common failure: the engineer's phone silences calls at night. Most phones let you allow calls from chosen contacts through Do Not Disturb or Focus modes. Save the answering service's transfer and text numbers as contacts on every on-call phone and add them to the allowed list. Check it at every rota handover.

Choosing a service that can run the chain

Vendors all say they handle emergencies. Ask each one these questions, and ask to see the settings rather than a slide:

• Can I define trigger words and screening questions, not just the word "emergency"?
• Can it both transfer the call and send a text summary, and in which order?
• Can I set a second and third person to try if the first doesn't answer, with timers?
• Does the on-call person have to acknowledge, and what happens if they don't?
• Can rules differ by time of day and day of week?
• Do I get a full transcript and recording of every call?
• What happens to calls if your platform is down, and how will I know?

Three examples of what's available, as of September 2026. Jobber's Receptionist lets you add trigger terms such as "emergency" or "urgent" to an escalation policy and choose whether it texts you or transfers the call; it's a $29-a-month add-on including 30 conversations ($0.79 per conversation after that) and included on Jobber's Plus plan. Smith.ai answers with AI first and can pass a call to a live human receptionist at any hour; its AI receptionist plans run from a free tier of 25 calls a month to a Pro plan from $150 a month. Avoca is built for heating, plumbing and electrical firms and says it routes emergency calls to the on-call technician, with integration into ServiceTitan. None of these confirms, on its public pages, the full multi-step chain above, so test the fallback behaviour in a trial rather than assuming it.

If you're weighing an AI line against a staffed service for this job, AI versus human answering services for heating and cooling firms sets out the trade-offs. Whichever you choose, tell callers at the start that they're speaking to an automated assistant; if you serve customers in the EU, the EU AI Act's transparency rules require it.

A Saturday night call, step by step

Follow one illustrative call through the heating-and-cooling firm's chain on a cold weekend night.

• 22:14 A caller says her father's boiler has stopped and the house is cold. The AI asks about gas smell and CO alarms (no to both), then about vulnerable people: her father is in his eighties with a heart condition. That matches the list.
• 22:15 Warm transfer to the primary on-call engineer. It rings for 25 seconds with no answer.
• 22:15 The AI tells the caller to expect a call back within 10 minutes, texts the engineer a summary and sends the caller a text with a reference.
• 22:25 No acknowledgement. The AI calls the backup engineer, who answers and accepts.
• 22:27 The backup engineer rings the caller. He is on site by 23:10.
• Monday The transcript review finds the primary engineer's phone had a sleep Focus mode on that silenced the transfer number. It's added to the allowed contacts, and the rota's phone-check column gets filled in properly.

Without the chain, that call ends in a voicemail at 22:15. With it, the only cost was ten extra minutes, and the review turned a near miss into a fix.

The morning summary for everything that wasn't an emergency

An emergency line that works well also takes a pile of ordinary overnight calls, and they need to reach the office by 8 a.m. in a form someone can act on. Ask the service, or an AI reading its transcripts, for one summary sorted by what needs doing. An illustrative one:

Overnight, 9 calls. 1 emergency (handled, see log). 8 others:
CALL BACK TODAY
- No heating, no vulnerable people, caller asked for first slot
- Boiler pressure dropping for a week, wants advice
BOOK
- 3 annual service requests (plan members: 2)
- 1 quote request: heat pump, wants a survey
NO ACTION
- 1 wrong number; 1 sales call
CHECK
- Caller hung up after the first question; number withheld

The "check" line matters. A call that ended after one question could be a wrong number or someone who gave up, and it's worth a look at the transcript. In this illustration, the "no heating, first slot" caller got a 7:55 phone call and a 9:30 visit, which is the kind of service people remember when their plan renewal arrives.

What false alarms cost, and when to tighten the screening

Rule 5, "if unsure, treat it as an emergency", creates false alarms by design. Put a number on them before deciding whether to tighten anything. A transfer that turns out to be a dripping valve costs the on-call engineer about five minutes and a broken night's sleep. A false alarm that sends an engineer out costs far more: say two hours of paid call-out time and 30 miles of fuel, easily $150 or more at typical engineer rates, plus a tired engineer the next day.

So tighten the script only where the transcripts show the same false alarm repeating, and only with a question that separates it cleanly. In the illustrative week above, one false alarm was a thermostat with flat batteries. Adding "Does the thermostat's screen show anything at all?" to the no-heat route, with a fixed line suggesting the batteries are checked when the screen is blank, removed that kind of call without making any real emergency harder to catch. Never tighten the gas or carbon monoxide routes to reduce false alarms.

When the AI line itself goes down

Any cloud service has outages. Plan for the answering platform failing, not just the engineer:

• Forwarding rules at the phone provider. Most business phone systems can forward calls when the destination is busy, doesn't answer or is unreachable. Point that fallback at the on-call mobile or a staffed overflow service.
• A morning test call. Someone rings the line at the start of each day and checks it answers. Two minutes, and you find a broken number before a customer does.
• Status alerts. Subscribe to the vendor's status or incident notifications with an email that someone actually reads.

The monthly fire drill: staged calls that prove it works

Once a month, someone outside the on-call rota makes six test calls from a phone the system doesn't recognise, and scores each one. The calls should be awkward on purpose, because real emergencies are.

An illustrative failed drill: the caller said "the boiler's leaking a bit onto the floor", and the AI booked a Thursday slot. In the scenario, the water was spreading towards a socket. The fix was adding the "near sockets or through a ceiling?" question to the leak row, which is why it now appears in the table above. Each failed drill should produce exactly one change to the script, re-tested next month.

What the weekly transcript review should catch

Spend twenty minutes a week reading the escalated calls and a sample of the non-escalated ones. An illustrative week for the six-engineer firm: 41 after-hours calls, 6 escalated, 2 of those false alarms (a dripping radiator valve and a thermostat with flat batteries), none missed, and a typical time to engineer pickup of under a minute. Look for three things: any call that should have escalated and didn't, false alarms that a better screening question would have filtered, and any delay between steps in the chain. Reviewing AI call transcripts for quality has a fuller routine.

It also helps to know what a miss would have cost. What missed calls cost a heating and cooling business gives you a way to put a number on it, which makes the case for keeping the drill going when the line has been quiet for months.

Further reads

• AI Receptionist for Tradespeople: What It Costs and When It Pays — What an AI receptionist costs a trades firm and when it pays for itself.
• How to Set Up an AI Receptionist Without Losing Callers — The general set-up steps for the non-emergency side of the line.
• Can AI Answer My Business Phone? What It Can and Can't Handle — What phone AI handles well and where it still struggles.
• How to Set Up an AI Phone Agent for After-Hours Calls — Handling routine after-hours calls that don't need an engineer tonight.
• When Should an AI Chatbot Hand Over to a Human? — Handover rules that also apply to web chat and messaging.
• What Is the Cheapest Way to Handle Enquiries Out of Hours? — Cheaper options if emergency call-outs are rare in your trade.
• AI or Human Answering Service: Which Suits a Property Manager? — AI and human answering services priced for a 300-unit portfolio, with emergency routing rules and ten test calls to make before you sign.
• AI Service Plan Reminders for Heating and Cooling Customers — Clean the due-date list, time reminders to the heating and cooling seasons, write messages about each customer's actual system and let AI sort replies.
• Seven AI Mistakes That Cost Tradespeople Jobs — The seven ways AI quietly loses jobs for plumbers, electricians, roofers and decorators, how each one shows up, and the fix for each.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Jobber help article on Receptionist powered by Jobber AI; Smith.ai AI receptionist pricing page; Avoca AI page for HVAC firms. Checked September 2026.