Somebody is going to call your business and ask a question nobody ever wrote down. Not a trick question. A real one. Does that price include hauling the old unit away? Can you come Sunday? Is your tech licensed in Kenton County, because the last guy wasn't. Your AI receptionist will hit that wall inside the first twenty calls, and what it does in the following four seconds tells you more about the product than any demo ever will.

Every demo you'll watch, ours included, runs the happy path. Caller wants an appointment, the appointment exists, the calendar opens, everyone claps. That call proves the plumbing works. Worth something. Tells you nothing about how the thing behaves when a caller wanders off the map.

Bluffing is the failure that costs money

A voice AI has three options when it doesn't know. One of them is acceptable. All three look identical on a pricing page.

Option one is the bluff. A caller asks what full ductwork replacement runs on a 1,900 square foot ranch. Your price list stops at "system install, from 6,500." Instead of stopping where your facts stop, the AI produces a confident number. Say 4,200. Now somebody is holding a recorded phone call in which your business quoted 4,200 for nine grand of work. You eat the difference or you have the worst conversation of your month.

Then there's the stall. "I can definitely help with that! Now, what can I help you with today?" Round and round. Nothing false ever gets said, which is why owners assume it's the safe failure. The caller hangs up somewhere around ninety seconds and dials the next name on the map.

The third one is what you're paying for. "I don't have that in front of me, let me get you to Dave." Or, when Dave is under a house at 8:40 at night: "I can't answer that one and I don't want to guess at it. I've got you at 859-555-0148, is that the best number? He'll call you before ten tomorrow."

Look at what that call gave up and what it kept. It surrendered the question and held onto the customer. That trade is the one most voice AI pitches get backwards, because conceding anything sounds terrible in a sales call.

A model has to be forced to say I don't know

Language models want to be helpful. Left alone, one will generate something plausible for any question you throw at it, because producing plausible text is the whole job. Getting it to stop at the edge of what it actually knows means fencing it to a specific set of facts: your hours, your services, your prices, your rules about deposits and travel fees and Saturday rates. Inside the fence it answers. Outside it says so and moves the caller to a person.

Humans do this too, which nobody wants to admit. The nineteen year old covering your front desk in July guesses at prices constantly, and nobody calls that a hallucination, they call it Tuesday. Volume is the difference. A summer temp guesses forty times before school starts. An AI answering every call at every hour guesses forever, the same wrong way each time, and it compounds. Anyway.

Half the time nobody picks up the transfer

Handing the call to a live person is the clean outcome, and it works maybe half the time, because the other half your plumber is in somebody's crawlspace with his phone sitting in the truck. So what matters is what the fallback message contains. A useful one carries:

  • the caller's number, read back and confirmed out loud on the call
  • what they asked, in their words, not flattened into "customer had a pricing question"
  • what the AI already told them, including anything it committed you to
  • when they called and what they were trying to book

That third item is the one everybody skips and the one that bites. If your receptionist told a caller someone would confirm by five, you need to know that promise exists before five o'clock. In Rocketship those details land as a record you can open, in the same place a lead from your website form lands. Which mostly matters because it's one list to work instead of a voicemail box, an inbox, and a notepad by the register.

The bar is low, which isn't permission to be sloppy

A 411 Locals study of 85 businesses found that 62 percent of calls to small businesses go unanswered. The Numa Small Business Phone Report found that 85 percent of callers who reach voicemail never call back. So an AI that picks up, admits it doesn't know one thing, and takes a proper message is already ahead of what most owners do today, which is nothing.

Clearing a low bar isn't the goal. A bluffed price does more damage than a missed call ever did. The missed call costs you one customer. The invented answer costs you the customer, the refund, and a one star review sitting on your Google listing for three years telling everyone you quote one number and charge another.

Three calls to make before you pay anyone

Start a free trial, ours or anybody's, and phone your own number three times.

First call: ask about something next door to your service list but not on it. A two truck window cleaning outfit gets asked about gutters constantly. Listen for whether it invents a gutter price.

Then call back and state a wrong premise with total confidence. "Your site says you do same day emergency work, right?" when you don't. Software that agrees with you to be pleasant will agree with your customers to be pleasant.

Last one, and do it at 9:40 at night. Demand a human, refuse to explain why, see where you end up. Then check in the morning whether the message waiting for you actually has your caller's number in it.

If it bluffs once across those three, it'll bluff on a paying customer, and you won't be in the room to hear it happen.