It is 9:40 on a January night. A homeowner calls your line and says, "There's a weird smell by the water heater, kind of like rotten eggs. Can someone come look tomorrow?" Your AI receptionist does exactly what it was built to do. It confirms the address, checks the service area, and offers an 8-to-10 a.m. window. The call is polite and efficient, and it is completely wrong. That call is why AI receptionist escalation to a human is the first setting I check in any setup. I look at it before the greeting, before the booking logic, and before the integrations.
The caller asked for a tomorrow appointment, so the AI gave them one. Nothing in "can someone come look tomorrow" sounds like an emergency. The emergency is in the words "rotten eggs," and the agent won't treat it that way unless someone told it that phrase outranks everything else the caller says. AI voice agents are good at following the conversation the caller leads. Escalation rules tell the agent when to stop following and start protecting the caller.
Below are the five triggers I'd make mandatory for any HVAC, plumbing, electrical, or general home-service shop running an AI front line. For each one: how to configure it, and the caveats I learned from reading transcripts where the rule existed on paper but never fired.
What does escalation actually mean on an AI receptionist?
An escalation rule has four parts: a detection condition, a script, a destination, and a fallback. Most AI receptionist platforms let you define a transfer action and describe in plain language when to use it. Many also support hard keyword rules. Those two methods behave differently:
- Plain-language conditions ("transfer if the caller describes an emergency") are interpreted by the model. They catch phrasing you didn't anticipate, but they can be talked out of firing by a calm caller.
- Keyword rules ("if the transcript contains 'gas,' 'rotten egg,' or 'sulfur'") are deterministic. They fire every time, including on false positives like "gas water heater tune-up."
For life-safety triggers, use both and accept the false positives. A tune-up caller who gets transferred to a person has lost 30 seconds. A gas-leak caller who gets booked for tomorrow could lose far more. For the business triggers (repeat callers, commercial accounts), plain-language conditions plus a data lookup work better, because keywords alone over-fire.
One ordering rule applies to all five: evaluate escalation triggers before anything else in the call flow, including the service-area check. In the scenario above, the agent's first question should never have been a zip code.
Wondering which AI tool pays for itself first? Request a free AI audit and find out.
Trigger 1: Should an AI receptionist transfer a gas smell call, or end it?
Both, in that order. The instinct is to wire gas calls straight to the on-call tech. That's the wrong first move. Natural gas has no odor of its own; utilities add mercaptan so people notice leaks. Gas utility safety guidance is consistent: leave the building, don't operate switches or phones inside, and call the utility or 911 from outside. PHMSA, the federal pipeline safety regulator, and local utilities publish the same basic instructions. Your AI should say them before it does anything else.
Carbon monoxide belongs in the same bucket. According to the CDC, more than 400 people in the U.S. die each year from unintentional CO poisoning not linked to fires, and more than 100,000 visit the emergency department. A furnace call that mentions a beeping detector is a life-safety call.
Rule card:
- Detect: gas, rotten egg, sulfur, "smells like a match," hissing near the meter or range, CO alarm, carbon monoxide detector, "the detector keeps going off," plus headaches or dizziness mentioned alongside the furnace or water heater.
- Say: "For your safety, please leave the home now. Don't flip any light switches or start a car in an attached garage. Once you're outside, call [your gas utility's emergency number] or 911. I'm alerting our on-call technician right now."
- Route: Immediately text the on-call tech the caller's number and address, then attempt a warm transfer if the caller confirms they are outside.
- Fallback: If no human answers, the AI repeats the utility/911 instruction and ends the call. It never books an appointment.
Two caveats from real transcripts. First, callers downplay. "It's probably nothing, but there's kind of a smell" is the most common phrasing, and model-only rules sometimes rank that below the booking request, which is why the keyword backstop matters. Second, don't let the AI diagnose. CO alarms use a four-beep pattern and a low-battery chirp is a single beep, but the AI is not the one who decides which the caller is hearing. Any mention of a CO alarm fires the rule. A human makes the judgment call from there.
Trigger 2: When should water actively running hand off to a human?
Only when the water is moving right now, and the AI has to ask that directly. Most water calls are not emergencies. The EPA's WaterSense program estimates household leaks waste nearly 1 trillion gallons a year nationwide, and about 10 percent of homes have leaks wasting 90 gallons or more a day. Nearly all of those are slow drips, running toilets, and weeping fittings, and your AI should book them like any other job.
The call that can't wait is active flow. Residential systems commonly run 40 to 80 psi, and both major plumbing codes require a pressure-reducing valve above 80. A split supply line or a failed water heater at that pressure is soaking drywall while the caller talks. So the rule hinges on tense, not on the word "leak." "We had a leak under the sink" books normally. "Water is coming through the kitchen ceiling" escalates.
Rule card:
- Detect: any water keyword (leak, burst, flooding, pouring, dripping through, water heater) triggers one required question: "Is water still coming out right now?" A yes, or phrases like "won't stop" or "can't find the shutoff," fires the rule.
- Say: shutoff coaching first. The main valve is usually where the line enters the house, near the water heater, in the garage, or at the meter box by the curb (which may need a meter key). A water heater has its own cold-supply valve on top.
- Route: warm transfer to the on-call plumber, with the AI staying on the line through the shutoff coaching if the transfer is ringing.
- Fallback: second on-call, then the owner. Voicemail is not an acceptable endpoint for this trigger.
Trade-specific edge cases matter here. HVAC shops should include "water dripping from the ceiling under the attic unit." That's usually a condensate overflow, and the script should tell the caller to turn the system off at the thermostat. Condo and apartment callers often share a shutoff with neighbors, so the script should also tell them to notify building management. The plumbing-specific setup we run carries more of these branches than any other trade, because plumbing has the most urgent water scenarios.
Trigger 3: How should an AI receptionist handle an angry repeat caller?
By not trying to solve it. The caller who says "your guy was just here Tuesday and it's still not cooling" is a callback, the unbilled return visit that eats margin (see how callbacks show up as a margin leak). That caller is also a likely one-star review. An AI that answers with "I can get you on the schedule Thursday" confirms the caller's worst assumption about your company.
Sentiment detection alone is a weak trigger. Some callers are loud and perfectly happy, and some are quiet and ready to leave. The strong signal is data: the caller ID matches a job completed in the last 30 days, or the same number has called two or more times in 48 hours. That requires your AI to read from your field service software. If you're not sure yours can, this breakdown of which AI tools connect to Jobber, Housecall Pro, and ServiceTitan covers the lookup question.
Rule card:
- Detect: a recent-job or repeat-call match plus any of: "again," "still," "third time," "just here," "cancel," "refund," "review," "manager." An explicit request for a manager fires on its own.
- Say: "I see we were out there recently, and I'm sorry it's not fixed. I'm getting our service manager on the line now." No troubleshooting and no rebooking.
- Route: the service manager or owner, not the on-call tech. The on-call tech may be the one who did the original job.
- Fallback: a specific callback promise ("within 30 minutes") pushed to the manager by text with the job number, and a task logged in the FSM.
The caveat: the fallback promise has to be kept. An upset customer who is told "a manager will call you in 30 minutes" and doesn't get a call in 30 minutes is angrier than before they called. If you can't staff that promise after hours, have the AI promise a morning callback time you can meet.
Trigger 4: Why should commercial accounts bypass the AI receptionist queue?
Commercial work follows different rules than residential work. A property manager or facilities contact is usually operating under a service agreement with a contractual response window, a not-to-exceed amount, and a PO number that has to be on the invoice. Work orders often arrive through platforms like ServiceChannel or Corrigo. If the AI drops that call into a residential two-hour arrival window, you can breach the agreement even if the tech shows up and does good work. If you run agreements, how AI fits into maintenance plans and service agreements covers the recurring side. This trigger covers the reactive call.
Rule card:
- Detect: a caller ID match to a commercial customer record (the Residential/Commercial customer type in ServiceTitan, or the tag or custom field most Housecall Pro and Jobber shops use), or phrases like "work order," "PO," "store number," "property manager," "our location," or "facility."
- Say: capture the site address, store or unit number, PO or work order number, and the problem, then transfer.
- Route: the commercial dispatcher or account manager during hours. After hours, the on-call tech, with the account name read out in the transfer whisper.
- Fallback: an immediate text to the account manager with the captured fields.
Two refinements. Unknown commercial numbers need judgment. A property manager with 40 doors is a different caller than a homeowner who mentions their LLC, so have the AI ask "Is this for a business or a rental property you manage?" when a commercial phrase shows up from a number you don't recognize. And if you do commercial refrigeration, a restaurant with a walk-in cooler down is on a food-loss clock, so it goes at the top of this list alongside gas and water.
Trigger 5: What counts as the AI asking twice?
This is the quietest trigger and the one that saves the most hang-ups. The rule: if the agent has to ask for the same piece of information a second time and still doesn't have it, it stops asking and hands off. The same goes for failing to answer the same caller question twice.
Loops happen for ordinary reasons in this business: a caller on speakerphone in a truck, a compressor running in the background, a spotty rural cell signal, a strong accent, a street name the speech-to-text engine has never heard. The third "could you repeat the address?" is where callers hang up and call the next company on the list.
Rule card:
- Detect: any required field (address, callback number, problem description) requested twice without a confident capture. The caller saying "I already told you." Two consecutive questions the AI can't answer from its knowledge base ("Do you work on boilers?" followed by "What about Navien?").
- Say: "Let me get you to someone on our team so you don't have to repeat yourself."
- Route: the office during hours. After hours, a text to the on-call tech, or a text to the caller with a short intake form link if it isn't urgent.
- Fallback: a logged callback task with the partial transcript attached.
A related rule I'd treat as non-negotiable: if the caller asks for a person, a human, or a representative, honor it on the first ask. Making someone argue with a bot to reach a human is a common complaint about AI phone systems, and it's avoidable. Also review every loop in your weekly transcripts. Loops are the most useful list you have for improving the knowledge base.
Where does the escalation go when the on-call tech doesn't answer?
A trigger that fires into a phone nobody picks up is only half-built. Before a trigger goes live, set up the destination:
- Build a ladder, not a number. Primary on-call, then backup, then owner, ringing each for about 20 seconds. Most platforms transfer to any phone number, so a ring group or a sequential forward works.
- Use a whisper. Before the human is connected, the AI plays a five-second summary: "Gas smell, caller is outside, 1412 Oak Street." The tech walks in knowing the call type, and the caller doesn't have to repeat anything.
- Send the same summary by text. If the transfer drops, which happens on cell calls, the tech still has the number and the details.
- Match the fallback to the trigger. Gas and water never end in voicemail. Repeat callers, commercial accounts, and loops can end in a callback promise with a specific time.
Some shops use a live answering service as the escalation destination and keep the AI as the front line. That hybrid is a reasonable setup, and the AI receptionist vs. live answering service comparison walks through when it makes sense. Two compliance notes also apply here. If you record calls and pass them along, California Penal Code 632 requires all-party consent, so disclose recording in the greeting if you take calls from California. And if your fallback has the AI place an outbound callback, the FCC's February 2024 ruling classifies AI-generated voices as artificial under the TCPA, so consent rules apply.
Can you over-escalate an AI receptionist?
Yes, and it's the second most common failure after under-escalating. If "any caller who sounds frustrated" goes to the on-call phone, your techs start ignoring it by week three. At that point the AI is an expensive phone tree in front of the same guy's cell. Keep the triggers narrow and let the AI finish everything else: scheduling, service-area questions, membership questions, reschedules, and "do you work on my brand." The whole point of the setup is that the five triggers go to people and the other 90-plus percent of calls don't.
How do you test escalation rules before a real emergency does?
Call your own line. Once a month, and after any change to the agent's instructions, run a short script of test calls:
- A tune-up request that mentions a rotten-egg smell in the last sentence.
- "Probably nothing, but my CO detector beeped a few times."
- "We had a leak last week" (should book) followed by "it's dripping through the ceiling now" (should escalate).
- A call from a number with a job completed last week, saying "still not working."
- An unknown number saying "I've got a work order for store 214."
- A mumbled address on speakerphone with a fan running.
- "Can I just talk to a person?"
For each call, check three things: did the rule fire, how many seconds passed between the trigger phrase and a live human, and what the human heard in the whisper. For gas and water I'd aim for under a minute from trigger to a person. If you want the wider checklist for the rest of the call flow, the 12-point AI answering audit covers it. At Turnkey AI, escalation is where we spend the most setup time on any AI receptionist. Booking logic is easy to fix after launch. A missed gas call can't be fixed after the fact.
Frequently asked questions
Will an AI receptionist tell a caller with a gas smell to leave the house?
Only if you write that into its instructions. Out of the box, most agents try to help with the stated request, which is usually an appointment. Put the utility and 911 safety script at the top of the escalation logic, back it with keywords, and test it with a caller who buries the smell at the end of the call.
Can the AI transfer calls to my cell phone after hours?
Yes. Most AI receptionists can transfer to any number. Use a sequential ring list (primary on-call, backup, owner) rather than one cell, and add a text summary so the details survive a dropped transfer.
What if a caller just asks for a real person?
Transfer on the first request. Making callers repeat the request or argue with the agent costs you the call and often a review. The AI can ask one quick question about the reason first, but it shouldn't refuse.
How does the AI know a caller is a commercial account?
The most reliable method is a caller ID lookup against your field service software, using the Residential/Commercial customer type in ServiceTitan or a tag in Housecall Pro or Jobber. Phrases like 'work order,' 'PO,' or 'store number' catch commercial callers from numbers you don't have on file.
Should escalation rules change after hours?
The triggers stay the same; the destinations change. Gas and active water go to the on-call ladder around the clock. Repeat callers and loops can fall back to a guaranteed morning callback after hours, as long as someone actually makes that call.