An AI agent for emergency service calls: auto shops, plumbers and electricians at 10 p.m.

On an emergency call, the customer is looking for just two answers: when someone will get there and roughly how much it will cost, and every minute without them is a minute they spend dialing the next number. An AI agent for emergency service calls is built around three actions: classifying how serious the problem is in two or three questions, giving an arrival window drawn from a real source of availability rather than a polite phrase, and handing off to a human when there's danger or a decision only the business owner is authorized to make. This guide also covers what breaks in practice: a wrong classification, a window that wasn't kept, and when this kind of answering doesn't pay off.

Why an emergency call behaves differently from any other inquiry

A sales inquiry leaves breathing room: the customer compares and comes back tomorrow. An emergency call leaves none. Water is running across the floor, a car won't start in the parking lot at work, the electrical panel trips at 10 p.m. The customer tapped the first number on Google, and when no one answered, tapped the next one.

The customer has already decided; they just want to know who's available

There's no persuading left to do, only logistics: are you coming, when, and what's the ballpark cost. Built for service calls, an AI agent that answers calls and WhatsApp behaves the opposite of a sales agent: it keeps things short, confirms and books. Questions with no immediate use are dropped, and the caller's name and phone number are collected along the way.

How does an AI agent prioritize an emergency call?

It sorts calls into four levels set in advance with the business owner: immediate danger, a complete shutdown, a limiting problem, and an inquiry that isn't a service call. It doesn't prioritize by the word "urgent," because almost everyone who calls at 8 p.m. says it.

• Immediate danger: the smell of gas, smoke, water in the outlets, a car stuck in a traffic lane. A fixed safety instruction and an immediate handoff to a person.
• Complete shutdown: a leak that won't stop, no power in the house, a car that won't start. Booked for tonight or first thing in the morning, according to your policy.
• Limiting but bearable: a water heater that won't heat, a dead outlet, a shaking steering wheel. Scheduled for a workday with a window of a few hours.
• Not a service call at all: the price of an annual vehicle inspection, opening hours, a specific model. An immediate answer, without taking up a slot in the calendar.

For an auto repair shop, the key distinction is between a car that can be driven and a car that needs a tow; for a plumber, between water that's still flowing and water that's been shut off at the valve; and for an electrician, between a local outage and a burning smell.

When in doubt, raise the level

A customer says it's a drip, and on site it turns out a pipe has burst behind a wall. That's why the classification is a routing decision, not a diagnosis, with a hard rule on top of it: when the answers are contradictory or partial, you move the level up, never down. Red-flag words such as water near an outlet, a smell or smoke raise the level automatically. A mistaken upgrade costs an unnecessary phone call; a mistaken downgrade costs a flooded apartment. The summary that reaches you always includes the customer's original words, not just the label.

Where the arrival window comes from, even when you have no calendar

The arrival window is the real product of the call, so it has to have a source. In a business with a digital calendar, the agent reads the bookings from it. For a tradesperson whose schedule lives in their head and a notebook, and there are plenty of those, you build a written capacity rule that replaces a calendar: how many calls a day, how long a call takes, which areas on which days, and after what time you stop heading out. With these four, the agent says "tomorrow between 8 and 11 in your area" without a calendar.

A proposed window versus a confirmed window

Without a synced calendar, you separate the offer from the confirmation. The agent gives a window as a proposal, says it will confirm within a set time, and sends you a message with two options: confirm, or an alternative window. If no reply comes in time, it tells the customer the window hasn't been confirmed yet and escalates the call to the wake-up chain.

The same window for two callers, and a window that wasn't kept

A window that's been offered is locked until the customer responds and is then either released or booked, so two callers in the same minute don't get the same slot. If the window passes and no one has arrived, the system catches the slip before the customer does: a delay message goes out, you get an alert, and the new window is given only after you set it. For the same reason, a range that's kept, between 8 and 10 with an update on the way, beats a "within an hour" that turns into an hour and a half and a bad review. And when there's no one who can come tonight, the agent says so and offers the first available window.

When does the agent have to wake someone up?

Only in two situations: danger to people or property, and a financial or operational decision it isn't authorized to make. In every other case, it closes the call on its own and leaves a summary for the morning. The handoff is a chain, not a button: a call to the main phone, and if there's no answer within a set number of seconds, a call to the second number, and at the same time a three-line message: what happened, where, and how urgent it is. If no one answers, the agent goes back to the customer and tells them what the next step is. A night policy is worth having too: after midnight, only danger rings a phone. That's a business decision, and it's yours to make; when an AI agent for business is built, it's settled in writing in advance.

In the first week there are almost always false alarms, because the threshold was set before any real calls had been seen. That's why the first week is a calibration week: you go over every handoff, mark which ones were justified, and refine the red-flag word that triggered them. A business owner who gets up three times for nothing stops answering the fourth time.

Where agents like this break: spoken Hebrew, noise and numbers

A caller with an emergency doesn't speak textbook Hebrew. They say הפוזה קפצה ("the fuse blew") or המים עולים מהריצפה ("the water's coming up through the floor"), on speakerphone while driving or over the noise of an auto shop, sometimes in Hebrew mixed with Arabic or Russian. Recognizing the problem is worthless without a mechanism behind it, so three are set up.

• Repeating back and confirming every critical detail: "Got it, a leak under the kitchen sink, right?" Nothing moves forward without confirming the type of problem, the address and the callback number.
• A confidence threshold and a fallback to a person: when the speech recognition engine returns a low-confidence result, the agent tries a different wording once, and if that fails too, it escalates.
• Closed questions in noisy settings: in a noisy environment, open questions are swapped for yes-or-no questions, and the agent offers to continue in writing.

The real danger lies in the exact details. Street names are a common failure point in Hebrew speech recognition, so the address is read back in full, and in the written channel the agent asks the customer to share their location. A phone number isn't transcribed: the source is the number the call came from, and only a different number the caller dictates is read back digit by digit. An agent that wrote down the wrong address has created a bigger problem than an agent that didn't answer.

What the agent says about itself

It opens with a short introduction that makes it clear this is the business's automated answering service, and it doesn't pose as a receptionist. That's not just fair; it also protects your reputation, because someone who finds out after the fact that they were talking to a machine tends to vent about it in a review.

How do Israelis really contact a tradesperson in an emergency?

They call first, and if no one answers, they immediately send a WhatsApp message to the same number and move on to the next result on Google. That's why one channel isn't enough: a phone without WhatsApp misses the message sent a second after the phone stopped ringing, and a bot for tradespeople that sits only in chat misses the person whose hands are wet. The load peaks on Saturday nights after Shabbat ends, on holiday eves and on the first rainy days of the season, and a business that's closed on Shabbat and holidays needs its answering service to say so and offer an appointment.

Two numbers, one thread

Many service businesses advertise a landline, and WhatsApp wasn't built for that kind of line. You can verify a landline number with a voice verification call, but if it sits behind a switchboard or is already connected to a personal WhatsApp account, that falls apart. The practical solution is a dedicated mobile or virtual number for WhatsApp, while the existing line keeps working and is forwarded to the answering service. That creates an identity problem, because the call comes in on one number and the message arrives on another, so you merge by the caller's number: the call and the message go into the same service ticket, and the agent doesn't ask again what happened and where.

The 24-hour window that shapes the follow-up conversation

On the WhatsApp Business Platform, you can reply freely only within 24 hours of the customer's last message, and after that a message you initiate is allowed only as a template approved in advance. This is an operational constraint, not a technical detail: the next-day check-in has to be a template with variables (the type of problem and the window that was given) that's written and approved during setup.

On the regulatory side, two things. Call recording falls under Israel's Wiretapping Law, 1979, in which the rule for someone who's a party to the conversation differs from the rule for recording other people's conversations, and the accepted practice is to notify callers at the start and store the recordings properly. An advertising message you initiate is subject to Amendment 40 to the Communications Law (Israel's anti-spam law), which requires prior consent and an opt-out option, whereas replying to an inquiry the customer initiated is a service. This isn't legal advice, and the distinction is written into the settings before going live.

The message that goes out after the call, and the next-day check-in

A large share of service calls don't close in the first conversation. The customer hears a two-hour window, says they'll check and hangs up, which means they're going to talk to two more providers. This mechanism doesn't promise the job stays with you; it makes sure you're the only organized record the customer has when they compare. Three actions: a message right at the end, with the business name, the window and what the service call includes; a proactive offer of an earlier window if one opens up; and a check-in the next day: was the problem taken care of, and if not, an offer of an appointment. The check-in reopens some of the inquiries that slipped away, and it builds a record of which hours and which problems slip through the cracks at your business.

After a few weeks, the record adds up to a map of the areas and problems that actually generate urgent demand, and that's the material that decides which pages and which keywords get priority in your work on SEO for auto repair shops or SEO for tradespeople, instead of guessing.

When it doesn't pay off, and what you'll need to provide

An emergency call agent adds little for a business that already answers almost every call, for a business whose owner won't allow even a price range to be given, and for a business with no policy at all: no hours, no areas, no ballpark cost. Even with excellent answering, some callers will go with whoever answered three minutes before you, and that's not a glitch; it's the nature of the market.

We won't cite other clients' numbers here, because a figure you can't verify is worth zero. You can work out your own number yourself: pull an itemized call log for the past two months, count how many incoming calls went unanswered outside business hours, subtract repeat calls from the same number, and multiply by the average job value and a conservative closing rate. After going live, there are four things you can see with your own eyes in the very first week.

• How many calls and messages came in outside business hours, and what happened with each one according to the summaries.
• How many times the agent woke a person up, and how many of the handoffs were justified.
• Whether the arrival window that was given was actually kept, because that's the metric that generates reviews.
• How many calls that didn't close in the first conversation came back after the check-in, and how many turned into a job.

On your end, you'll need five things: a mobile or virtual number for WhatsApp with forwarding from the existing line, a single source of availability, a pricing policy that says what can be shared before a diagnosis, an escalation list with the call order and hours, and a closed list of scenarios. You fill in the list on a short intro call, and the scope is set in writing in advance: a product with boundaries, not an open-ended project. The ongoing cost is made up of a fixed component and a component that grows with usage volume, and the breakdown is on our page about building AI agents for business.

What to launch first

If you have to choose, you almost always start with text. An agent on WhatsApp deals with clean text, you can read every conversation afterward and correct it, and mistakes get caught before they turn into a wrong window. Live voice answering in spoken Hebrew, with background noise, is a heavier project, more expensive and more sensitive to errors, because the customer has no screen to correct things on.

Frequently asked questions

What does the agent say when a customer asks outright whether they're talking to a person?

It answers that it's the business's automated answering service and doesn't deny it, because a denial that comes out later turns into a negative review. Right after answering, it moves on to the practical question (what happened and when someone can come), because a person under pressure wants a solution, not a discussion about technology. An explicit request for a person triggers the escalation chain, and in a dangerous situation the handoff happens without anyone asking.

Can the agent tell a customer to shut off the main water valve or switch off a breaker?

Yes, but only according to fixed safety instructions worded in advance and approved by the tradesperson, with no improvising. These are set sentences spoken in defined situations, such as an active leak or a burning smell, and they always come with an immediate handoff to a person. The agent doesn't diagnose over the phone and doesn't walk the customer through any action that calls for professional judgment on site.

What happens when two urgent customers call in the same minute?

Both get a full answer at the same time, with no queue and no busy signal. The difference is in what happens next: the classification decides which call is shown first, and the slot given to the first caller is locked, so the second caller gets the next available window, not a double promise. This prevents the less urgent call from being handled first just because it came in a second earlier.

Can the answering service run only at night and on weekends?

That's one of the most common setups, especially at the start. During working hours, calls keep going to the tradesperson's phone, and the agent steps in only outside the defined schedule, or when the line is busy and a call goes unanswered. That way the business keeps the human conversation and covers the empty hours.

What happens if the call drops midway or the reception is bad?

Everything collected up to the moment the call dropped is saved, including the number the call came from. The agent detects a call that was cut off before it ended and acts according to what was set up, usually a short message to collect the rest of the details, or one callback attempt. This is critical, because people calling from the road or from a parking lot get cut off a lot, and without this mechanism the call disappears.

The bottom line

An emergency call doesn't get closed by a polite conversation. It gets closed by the right classification, an arrival window with a source behind it, and a clear rule for when to wake someone up. An agent built around these three, with a plan set in advance for what to do when it's wrong, turns the hours when no one is available into hours when the business still takes in calls.