A fail-closed AI receptionist never invents an answer. Questions covered by your rules get answered from your rules, exactly. Questions outside them get the response a good desk gives anyway: let us get someone out to look. And everything it says is logged word for word, so "what did it tell my customer" is never a mystery.
01What happens on a question outside the rules?
The design principle is called fail-closed, and it is worth understanding because it separates two very different kinds of systems that get sold under the same name.
A fail-open system treats every question as answerable. Asked something it has no grounding for, it produces something fluent, because producing fluent text is what the underlying technology does naturally. Fluent and correct are not the same property, and on your phone line the gap between them has your name on it.
A fail-closed system inverts the default. Its rules and your business information define what it may assert. A caller asks something inside that boundary, the answer comes back exact. A caller asks something outside it, and instead of reaching for a plausible guess, the system reaches for the two moves that are always safe: book a visit so a professional can look, or take a message and escalate it to you.
So the honest answer to "what if it makes something up" is a design question to put to any vendor: what happens, mechanically, when your system does not know? If the answer involves the machine doing its best, keep shopping.
02Why "we have to see it" is the correct answer sometimes
Run through the calls where a confident phone answer would be malpractice in your own shop.
A homeowner describes a fifteen-year-old furnace that short-cycles, but only on cold mornings. A slab leak that might be a slab leak or might be a bad valve under the sink. A ceiling stain that could be a roof issue, a bathroom above, or condensation on a duct. You have run these calls for years, and you know the truthful answer to every one of them: it depends, and it depends on things nobody can see through a phone.
Your best dispatcher already says "we have to get eyes on it" a dozen times a week. Nobody calls that a limitation. It is the difference between a professional operation and a guy guessing over the phone.
A well-built AI receptionist gives the same answer in the same spots, with one improvement: it never gets tired enough, rushed enough, or eager enough to guess anyway. The caller still gets what they actually wanted, which was never really a phone verdict. It was a plan: someone is coming Tuesday at 8, here is what the visit covers, here is what to shut off in the meantime.
03How price rules keep numbers honest
Quoting is where wrong answers cost the most, so quoting is where the boundary has to be hardest. The mechanism is plain: your prices live in written rules that you set, and the system quotes from the rules or it does not quote.
You decide the shape of your own rulebook. Flat prices for standard work: service call fee, capacitor swap, water heater flush. Ranges where honest ranges exist. Estimate-only categories for everything that varies too much to price blind: anything behind drywall, anything on a system past a certain age, anything the caller describes two different ways in one call.
Then the arithmetic of trust: a number that comes from a written rule cannot be an invented number. Change your prices and the rules change that minute, so the phone never quotes last season's rate. And a question with no rule behind it gets no number, it gets a booked estimate visit instead. The quoting page walks through the mechanics.
There is a real cost to running it this way, and you should hear it from us rather than discover it: a fail-closed system will sometimes book an estimate visit for a question a seasoned human might have priced from experience on the spot. That is an extra truck roll here and there, on purpose. We take that trade every time, because the alternative is improvised prices with your name on them.
04Every answer on the record, word for word
The final guarantee is a paper trail, not a promise about behavior. Every call it takes is logged word for word. Every quote is logged with exactly what was said.
That changes what a dispute looks like. When a customer says "the phone quoted me a different number," you do not weigh two memories. You read the transcript, in thirty seconds, and settle it with the actual sentence. Either the quote matched your rule, in which case the rule speaks for itself, or you have found something to fix, precisely, tonight.
It changes vendor accountability too. A vendor whose system logs everything, and hands you the log, is betting their contract on the machine behaving. Accept nothing less from anyone.
The log also feeds your crew. What the caller said, what was quoted, and what was booked flows into the job brief, so the tech walks in knowing the conversation instead of restarting it at the front door.
Wrong answers are not a risk you manage with hope. They are a risk you design out: rules, fail-closed defaults, and a word-for-word record, all running behind the receptionist itself. If you want to see what the boundary would look like drawn around your price book, get in touch. Twenty minutes. We look at your call volume and tell you straight whether this pays for itself.
QUESTIONSCommon questions
Can an AI receptionist give a wrong quote?
Not if it is built fail-closed: no rule behind it, no number on it. Anything outside your rules books an estimate visit instead of inventing a price.
How do I know what the AI told my customer?
Every call and chat should be logged word for word. If a number was quoted, you can read exactly what was said.
Twenty minutes. We look at your call volume and tell you straight whether this pays for itself. If the math does not work for your shop, we say so on the call.
Get in Touch