What a good AI receptionist should refuse to do
A plumbing-shop owner texts the Avidra number at 3:14pm: "Can you order me 200 feet of half-inch PEX from Wolseley?" The AI responds politely that ordering supplies isn't something it can do. Not because it's broken. Because it shouldn't. The shop owner finds another way to place the order, which takes him 90 seconds.
That refusal is the kind of constraint that makes the rest of the product work. An AI receptionist that tries to do everything ends up doing nothing well. The honest version of "what your AI can do for you" is also a list of what it shouldn't and won't. Most of those refusals are features.
The general principle
There are tasks that an AI assistant CAN technically do. Most modern LLMs are flexible enough to handle a startling range of requests. There are also tasks an AI receptionist SHOULDN'T do, even if it can, because the failure mode is unacceptable for that specific task.
The distinction matters because cost-of-failure is asymmetric across task types. Misreporting how many leads landed today is annoying but recoverable. Misrepresenting a quote to a customer because the AI guessed at pricing is a different kind of mistake. The AI doesn't know which trade you're in, what your hourly rate is, or what your shop charges for an after-hours emergency. If it tries to answer those questions confidently, you have a brand problem that's worse than not having an AI at all.
The right shape is a small, sharp scope of "yes" and a long, explicit list of "no." A receptionist that knows what to refuse is a receptionist you can trust.
What a good AI receptionist should refuse on calls
Quoting prices it wasn't given. The AI should not invent numbers. If a homeowner asks "How much for a water heater swap?" and the shop hasn't configured a quote in the intake script, the right answer is "Our team gives quotes after a quick look at the situation. Want me to book a free estimate?" Not "Probably around $1,500." The shop's actual cost might be $800 or $3,000 depending on tank size and install complexity. Made-up numbers create downstream conflict.
Committing to time slots it can't actually deliver. The AI should book against a connected calendar, never against an assumed schedule. If the calendar isn't connected, the right answer is "I'll have someone from the team confirm a time. Expect a text back within the hour." Not "We can be there at 4pm." Without calendar integration, "4pm" is a wish, not a booking.
Diagnosing technical problems. The homeowner says "I think my P-trap is leaking, can you confirm it's the trap?" The AI should not become a remote diagnostician. Even if the LLM has plumbing knowledge in its weights, the cost of being wrong is real (the homeowner might tighten the wrong fitting and break it further). The right answer is "I'll capture the details and a tech will look at it. Want to send a photo so they can see what you're seeing?"
Handling emergencies as routine inbound. The AI should escalate, not handle, anything that sounds like a true emergency. A flooded basement is not a routine intake. The AI flags it, pages the on-call tech if configured, and tells the homeowner what to do in the meantime (shut off the main, move valuables off the floor). It doesn't try to book the call into a 2pm Thursday slot.