What an AI Agent Actually Does on a Plumbing Call: An Annotated Transcript
Everyone says AI can answer the phone. Almost nobody shows you the turns. Here is a real emergency intake, line by line, with margin notes on why each question exists — and where a human takes over.
“AI can answer your phones” is an easy sentence to say and a hard one to trust. What does it actually do when a panicked homeowner calls at 2 a.m. with water coming through the ceiling? Not the marketing version — the actual turns. So here is one: a real-shaped emergency plumbing intake, transcribed line by line, with a margin note on every turn explaining why it exists and what it is quietly deciding.
This isn’t a demo script written to look good. It’s the anatomy of a good intake — the same job your best dispatcher does on their best day, run consistently on the worst night of the year. Read it as an operator, not a spectator: at every step, ask whether your current after-hours setup would have done the same thing.
Why this one call is worth dissecting
Plumbing is the purest emergency-services case in home services. The customer is standing in water, they will call three companies in the next ten minutes, and the winner is simply whoever picks up and sounds like they have a plan. Everything downstream — ticket size, the review they leave, whether they ever call you again — gets decided in the first sixty seconds. And the stakes of missing it are not abstract.
The industry benchmark most often cited is that roughly a quarter of inbound calls to home-services businesses go unanswered (Invoca, 2024, vendor-published), and the overwhelming majority of callers who hit voicemail never leave one and never call back (CallRail, 2023, vendor-published). Treat the exact figures as ranges, not gospel — but the direction is not in dispute. A missed emergency call is almost always a permanently lost customer.
The transcript, annotated
The setup: it’s 2:14 a.m. A homeowner calls the shop’s main line. Nobody is awake. The agent answers on the first ring. Customer lines are marked C; the agent is A; the italic lines are the margin notes.
A:“Thanks for calling Rivera Plumbing, this is the after-hours line. I can help right now — are you dealing with an active leak or flooding this second?”
Why it exists: the very first turn does two jobs. It discloses that this is the after-hours line (honest, sets expectations) and it opens with the single highest-value triage question — is this actively getting worse right now? Severity gets established before anything else, because everything after this branches on the answer.
C:“Yes — there’s water coming down through the kitchen ceiling from upstairs, it’s pouring.”
A:“Okay, we’ll get someone to you. First, so we limit the damage — do you know where your main water shutoff is? It’s usually near the water heater, in the basement, or where the line enters the house.”
Why it exists: this is the turn that separates an intake agent from an answering service. Before booking anything, it delivers the safety instruction that protects the customer’s home. Every minute of water is money and goodwill. A voicemail can’t do this; a script-reading call center usually won’t.
C:“I think it’s in the garage. Hold on… okay, I turned the valve, it’s slowing down.”
A:“Perfect, that’s the right move. Now — is anyone in danger, and is there any water near electrical outlets, the panel, or light fixtures?”
Why it exists: the safety escalation check. Water plus electricity changes this from a plumbing job into a life-safety situation, and the agent needs to know before it decides how fast a human has to be involved. This is a hard branch, not a nice-to-have.
C:“No, nobody’s hurt. There’s a ceiling light it’s dripping near, but I shut the breaker off for that room.”
A:“Good. Keep that breaker off. Can I get the property address and a good callback number in case we get disconnected?”
Why it exists: capture the two pieces of data that make the job recoverable no matter what happens next. Callback number first, so a dropped call doesn’t become a lost customer. Notice the order: safety was handled before data collection, not after.
C:“It’s 4820 Cedar Lane, and my cell is 555-0148.”
A:“Got it — 4820 Cedar Lane, 555-0148. This qualifies as an emergency call-out, so there’s an after-hours trip fee before any repair work; the technician will confirm the full price on site before starting. Is that okay to proceed?”
Why it exists: honest price disclosure, up front. The emergency fee is stated before dispatch, not sprung at the door. This is a guardrail your shop sets — the agent quotes the trip fee rule you configured and explicitly does not improvise a repair price it can’t stand behind.
C:“Yes, that’s fine, just get someone here.”
A:“You got it. I’m flagging this as an active-leak emergency and getting our on-call tech on the line now so they can head your way. Stay on with me one moment — I’m bringing them in.”
Why it exists: the escalation trigger. A confirmed active leak with a customer who’s agreed to the call-out is exactly the case that should reach a human fast — not after three menu loops. The agent has already done the expensive part (triage, safety, address, consent) so the tech starts warm, with the full story attached.
The takeaway
The more common branch: it isn’t an emergency
Most after-hours calls aren’t geysers. They’re a water heater that quit, a running toilet, a dripping faucet that can wait until morning. The same agent handles those on a completely different path — and this is where the economics live, because these are the calls a human answering service takes a message on and your front desk chases the next afternoon, if ever.
- Same triage opener.“Active leak this second?” The no answer routes the whole conversation into the non-emergency lane instead of scrambling a truck.
- Book, don’t take a message.The agent offers the next real open slot against your actual calendar and confirms it — no “someone will call you back.”
- Capture the job details anyway. Brand and age of the water heater, symptom, address — so the tech arrives prepared and the morning dispatcher inherits a complete ticket, not a sticky note.
- Confirm by text. The customer gets an SMS with the appointment and a reply thread, so the whole thing lives in one place they can reach.
Where these turns come from (they aren’t guesses)
Every turn above maps to a decision your shop already makes — the agent just runs those decisions the same way every time, at 2 a.m., without getting tired or improvising. The design questions that produce a script like this are the ones worth arguing about internally:
| The turn | The decision it encodes | Who owns the rule |
|---|---|---|
| Severity opener | What counts as a true emergency vs. next-day work | Owner / ops |
| Shutoff instruction | Which safety steps the agent gives before dispatch | Lead tech |
| Electrical / life-safety check | The hard branch that forces a human, fast | Owner / ops |
| Fee disclosure | Exactly what price gets quoted and what never does | Owner |
| Escalation trigger | Which cases warm-transfer vs. book vs. message | Dispatch |
A good AI intake isn’t the agent being clever. It’s your best dispatcher’s judgment, written down once and run every single time.
That’s the honest reframe. The intelligence isn’t magic — it does not know your fee schedule or your escalation rules until you tell it. What it changes is consistency and coverage: the difference between the version of your shop that answers a 2 a.m. flood perfectly and the version that lets it ring out to voicemail and lose the customer for good. Given that roughly a quarter of these calls go unanswered today, that difference is most of the argument.
Sources
Keep reading