AI receptionists for service businesses: what works and what breaks.

Voice is the demo that sells, but it is rarely the first thing a service business should install. The cheapest and most reliable automation in the trades is a missed call text back: the call rings out, and seconds later the caller gets a message from your number asking what they need.

It works because it never has to understand anything. There is no transcription to get wrong and no conversation to mishandle. It just converts a dead end into an open thread, and it does that at a cost measured in cents. If you install one thing this year, install that, then decide whether voice is worth adding.

There are real jobs a modern voice agent does well, and they share a shape: predictable, low stakes, and better than the alternative, which is usually voicemail.

• After hours answering, where the honest comparison is not a human receptionist but silence.
• Overflow, when every line is already busy and the call would otherwise be lost.
• Screening, where the job is to find out what the caller needs and how urgent it is, then route accordingly.
• Answering the same five questions you answer forty times a week: hours, service area, whether you handle a given job, roughly how pricing works.
• Collecting details accurately: name, callback number, address, and what is wrong, handed to you as a readable summary.

None of these are hypothetical, and none are solved by a better prompt. They are properties of the situation your callers are actually in.

• Audio reality. A caller standing next to a running compressor, a thick accent, a bad cell connection, a toddler in the background. Transcription degrades, and the agent confidently acts on a misheard detail.
• Interruption and repair. Real people talk over each other, change their minds halfway through a sentence, and correct themselves. Scripted flows handle this poorly, and the caller can tell.
• Edge cases that matter. The unusual job, the angry customer, the insurance question, the one that needed judgment. These are exactly the calls you cannot afford to have handled averagely.
• Confident wrongness. The worst failure is not the agent saying it does not know. It is the agent booking a Thursday appointment at the wrong address and nobody finding out until a truck is parked on the wrong street.

Every decision that matters in an AI phone setup is a handoff decision. The agent should be built to give up early and cleanly rather than to push through a conversation it is losing. That means explicit triggers: anything about money beyond your published ranges, anything emergency flavored, any caller who repeats themselves twice, any caller who asks for a person.

And the whole thing needs a floor. If the service is down, the call must fall through to ringing your phone exactly as it did before you installed anything. An automation that fails loudly is worse than no automation, because it breaks the one thing that was already working.

The instinct is to make it sound as human as possible and say nothing. That is a bad trade. Customers are forgiving about talking to an assistant and unforgiving about being tricked, and a caller who figures it out mid sentence remembers that, not the convenience.

Saying plainly that an assistant is taking details and a person will call back costs almost nothing and buys you the right to use the tool at all. It also lowers the caller's expectations to something the technology can actually meet.

If any part of your setup sends automated text messages, that traffic has to be registered with the mobile carriers before it will deliver reliably. Registration takes paperwork and days, not minutes, and unregistered traffic gets filtered quietly, which looks exactly like the system not working.

This matters when you are choosing a vendor. Anyone who offers to switch your texting on this afternoon either already handled registration for you or is about to skip it. Ask which.

Run it as an experiment with a defined exit, not as a platform migration.

• Start with after hours only. The comparison is voicemail, so the downside is near zero.
• Use a dedicated tracking number so you can read exactly what happened rather than relying on anyone's dashboard.
• Read the transcripts for the first two weeks. All of them. That is where you learn what your callers actually ask.
• Decide in advance what failure looks like, for example any wrong booking, or more than one caller asking for a person and not getting one.
• Keep the fallback live the entire time, so switching it off costs you nothing but a phone call.

Text back first, voice second, and only for overflow and after hours. Design the handoff before the greeting, tell callers it is an assistant, register your messaging traffic, and pilot against voicemail rather than against a human. Judged that way, the technology is genuinely useful. Judged as a receptionist replacement, it will disappoint you and your customers at the same time.