What an AI receptionist actually does in 2026

That last sentence is the part vendors skip. The demo makes it look like the system understands your business. It does not. It follows the decision tree you gave it, and it will follow that tree at two in the morning exactly as written, including the parts you wrote badly.

I ran multi-unit restaurants for sixteen years. The phone was never a technology problem. It was a staffing problem that showed up as a technology problem, and the fix was always the same: decide in advance what the answer is, then make sure the answer happens when nobody senior is standing there.

What does an AI receptionist do on a real call?

Four things, in order.

It answers. A person hears a greeting within a ring or two, at any hour, with no hold music and no queue.

It qualifies. Name, callback number, what they need, where they are, when they want it. The same questions your best employee asks, asked the same way every time.

It acts. Depending on what you allowed, that means booking into a calendar, taking a message, texting the caller a link, or transferring to a human.

It records the outcome. A transcript, a summary, and a structured record you can actually count at the end of the month. The written record is the part that changes how you run the business, more than the answering.

Underneath, it is four separate systems in a chain: speech recognition, a language model, text to speech, and telephony. Each one adds delay and each one bills in a different unit. That matters later, and it is why two products that sound the same in a demo behave differently on a real line.

How much of your actual call volume can it cover?

Work it out from your call mix rather than from a feature list. Here is the arithmetic with published numbers, and you can rerun it with your own.

Maple published an intent taxonomy from 1.2 million calls across more than 1,000 US restaurant and local business locations. Reservations were 38% of calls. Hours, location and directions were 22%. Menu questions 15%. Takeout and delivery orders 12%. Event and large-party inquiries 7%. Modifications to existing reservations 4%. Everything else 2%.

Now sort those into three buckets. Fully scriptable: hours, location, directions, reservations, modifications. That is 64%. Scriptable with real work: menu questions at 15% and orders at 12%, because menus change and orders carry allergen risk. Human only: event and large-party inquiries at 7%, because those are negotiations.

Apply it to 300 calls a month. Roughly 192 calls are handled outright, 81 are handled if you invest in the knowledge base and the escalation rules, and 21 belong to a person. The honest pitch is 192 calls, not 300, and any vendor quoting you a containment figure in the nineties is either counting differently or has not seen your call mix.

Then check the timing, because the volume number hides the real problem. In the same dataset, 68% of calls arrive during the lunch and dinner rushes, 28% at lunch and 40% at dinner. The phone peaks exactly when the floor is least able to answer it. That is the case for automation stated properly. It is not that your team is bad at the phone. It is that the phone rings hardest during the ninety minutes they cannot get to it.

Health care shows the same shape from a different angle. Patient Prism scored 11,552,668 dental calls across 8,280 locations in calendar 2025 and found that of every 100 calls, 69 connected with the front desk and 31 hung up before reaching an agent. Only 34 were real booking opportunities. Another 35 were existing patients, vendors and wrong numbers. A third of the calls were never going to become anything, and a third never reached a human at all.

What can it not do?

It cannot exercise professional judgment, and in several industries it is legally barred from trying.

A dental AI is unlicensed personnel. It can state whether you are in network and collect a plan name and member ID. It must never tell a caller a procedure is covered or quote an expected out of pocket amount. That line is not a style preference, and a practice should have counsel confirm the wording before anything goes live.

A restaurant AI must never answer whether a dish is safe for someone's allergy. The FDA Food Code assigns allergen knowledge to a designated person in charge and satisfies its notification duty by written notice, which a voice agent is not.

It also cannot hear as well as the demo suggested. Deepgram's own explainer on word error rate walks through a phone call transcript with a 25% error rate, which it describes as about average for off the shelf speech recognition, and makes the point that matters: the transcript was perfectly usable except that it turned "declined" into "designed," so a call about a declined card got silently misrouted. The metric to ask about is not word error rate. It is entity error rate, meaning how often it gets the name, the number, the address and the appointment type wrong. Those are the errors that cost money.

And it cannot rescue a bad offer, fix a business that is genuinely too busy to serve the work, or make a caller want something they did not want when they dialed.

Is this the phone tree I already hate?

No, and the difference is worth being precise about.

A phone tree makes the caller do the routing. Press one, press two, listen to five options, press zero, get sent back to the main menu. The caller does labor on your behalf and resents it.

A conversational agent takes the routing back. The caller says what they want in their own words and the system decides where it goes. There is no menu to memorize and no wrong button.

The failure mode is different too. A phone tree fails by boring people into hanging up. An AI agent fails by confidently saying something wrong, which is worse, and which is why the escalation rules matter more than the greeting. Good implementations get the caller to a human before the caller has to ask, and how escalation is designed separates the working systems from the demos.

The deployment failure rate is real. Coval, a vendor selling evaluation tooling for voice agents, reports that roughly 95% of voice agents work in a demo and about 62% survive their first week live. It is a vendor figure and should be labelled as one, but the gap it describes is the correct thing to plan around.

Does it record calls, and does that matter in California?

It matters a great deal. California is a two party consent state, and consent has to come before the recording starts, not after.

Most AI phone platforms record and transcribe by default because transcription is how they work. That default is a compliance decision you are making whether or not you noticed making it. The specifics of California consent and AI phone systems deserve their own read, and the sentence at the end of that read is: confirm your setup with your own counsel before it takes live calls.

What laws actually apply, and which ones are being miscited?

Fewer than the pitch decks claim, and the miscitations are consistent enough to be diagnostic.

There is no California statute that literally requires a general commercial phone call to announce that the caller is speaking with AI. I still think it should disclose, for reasons that have nothing to do with law and everything to do with how people react when they find out later. But that is an operating choice, not a legal one.

Here is what gets cited at it anyway, and why each one is wrong:

The California B.O.T. Act, Business and Professions Code sections 17940 to 17943, is the bot disclosure law vendors reach for first. It does not reach voice.

SB 1120, from 2024, governs utilization review by health plans and payers. It has nothing to say about a dental office's front desk.

SB 243, from 2025, covers companion chatbots. A receptionist is not a companion chatbot.

AB 2013 requires generative AI training data transparency, and the obligation falls on the model developer, not on you as the deployer. It is still useful, but as a vendor diligence question rather than a compliance burden: ask each model and voice vendor for their AB 2013 disclosure URL.

The statutes that do apply to a health care adjacent deployment are AB 3030, which requires an AI disclaimer and human contact instructions on generative patient communications including verbal ones, and AB 489, which bars any persona implying a licensed professional. Those are narrow, real, and worth putting in front of counsel. Everything else on the list above is a feature being sold with a citation that does not support it.

How does it fit with everything else that generates leads?

The phone is one channel. Treating it as the whole problem is how businesses end up with a perfect receptionist answering a phone nobody is calling.

If you are a contractor, the images on your work pages do a lot of the qualifying before anyone dials, and what your photos do to quote requests is a cheaper thing to fix than your phone. If you are running paid traffic, the decision about whether that traffic lands on your homepage or a dedicated page usually moves more volume than call handling does.

Fix the leak that is actually leaking. For a lot of businesses that is the phone. For plenty of others it is not.

What to do this week

Call your own main number three times: once at 11am on a weekday, once at 7pm, once on Sunday morning. Use a number your team does not recognize.

Write down what happened each time. Rings, hold, voicemail, a human, how long, and whether you got a callback.

Then count last month's inbound calls against last month's booked jobs. If you cannot produce that number in ten minutes, that is your first project, not the AI.

Then sort a week of calls into the three buckets above: fully scriptable, scriptable with work, human only. The first number is what you are actually buying.

Only after those should you write the script. The script is the product. The voice is packaging.

Questions about whether an AI receptionist fits your phone?

Email me at eric@seod.com with your main business number and permission to call it. I will call three times at three different hours over a few days and send you back exactly what a stranger experiences, including how long each ring took and whether anyone followed up.

No pitch, no sequence. If the answer is that your phone is handled and your problem is somewhere else, I will say so.

More on the phone side sits in the AI phone and lead response library.