How Does an AI Receptionist Work? From Ring to Booked Consult

First, what an AI receptionist is — and what it isn't

Strip the marketing away and an AI receptionist is four pieces of software wired to a phone line: speech recognition that turns the caller's voice into text, a language model that works out what the caller wants, a knowledge base that constrains what the agent is allowed to say, and integrations that let it actually do things — book an appointment, transfer a call, take a message — instead of just talking about them.

That last part matters, because "AI receptionist" gets used loosely. It is not:

• An IVR phone tree. "Press 1 for appointments" is a menu, not a conversation. An AI receptionist holds open-ended dialogue — the caller says "I had Dysport a few months ago and want to come back in," and the agent understands that's a returning-patient booking, not a pricing question.
• Voicemail with transcription. Voicemail collects a message after you've already lost the moment. An AI receptionist resolves the call while the caller is still on the line.
• A human call center with software. Virtual receptionist services like Ruby put people on your line — a different model with different economics — while hybrids like Smith.ai mix AI with live agents on the same system. We compare the categories, Smith.ai included, in best answering service for med spas.

For a med spa, two more layers sit on top: a knowledge base that actually understands aesthetic medicine — services, pricing ranges, downtime questions, when a question needs a licensed provider — and a compliance wrapper, because patient phone calls can be protected health information. (Whether HIPAA applies to your med spa at all is its own question — we wrote a full guide on that.)

How does an AI receptionist work? The five-step call flow

Here's what happens on a real call, from ring to resolution:

1. It answers. Instantly, on the number you already have — 3 AM, Saturday, or the fourth simultaneous call during a lunch rush. There's no hold queue, because software doesn't have a line. Cordiva's agent also detects the caller's language automatically and runs the whole call in English or Spanish.
2. It understands. Speech recognition converts the audio to text, and a language model interprets intent — new consult, reschedule, pricing question, confirmation, wrong number. This is the step that separates it from a phone tree: the caller speaks normally, interrupts, changes their mind mid-sentence, and the agent follows.
3. It responds — or books. For questions, the agent answers only from the knowledge base built during onboarding: your services, pricing ranges, providers, promotions, policies, hours. For bookings, it writes directly into your scheduling software — Cordiva integrates with Boulevard, Mangomint, Aesthetix CRM, Jane App, Zenoti, and GoHighLevel — booking, rescheduling, and canceling in real time. This is the difference between an agent that books appointments and one that takes booking requests; ask any vendor which one they are.
4. It transfers. Some calls should reach a human, and you define which: VIP clients, complications, anyone who asks. During business hours, the agent hands the call to your team with context — who's calling and why — instead of a cold ring.
5. It escalates. When it can't resolve the call and can't transfer — a complex situation at 9 PM, a question outside its knowledge base — it doesn't improvise. It takes a structured message, flags the urgency, and queues the callback for your team. The failure mode is a clean handoff, not a made-up answer.

You also decide how much of your line the agent owns: every call, after-hours only, or overflow when your team is already on the phone. Most med spas start with overflow plus after-hours and expand once they've listened to the recordings.

What it can do — and what it honestly can't

The "can" list is genuinely long. A well-built AI receptionist can answer every call 24/7, hold multiple calls at once, answer the questions your front desk repeats forty times a week, book and reschedule appointments, switch between English and Spanish mid-call, take structured messages, follow your transfer rules, and log every conversation so you can audit exactly what it said.

The "can't" list is what vendors mumble. Here it is out loud:

• It never gives medical advice. Not "rarely" — never, by design. "Can I get filler while I'm pregnant?" is a question for a licensed provider, and the agent's job is to recognize that and route it: book a consult or take a message for your clinical team. An AI that answers clinical questions isn't impressive, it's a liability.
• It doesn't replace your in-person front desk. Nobody's AI greets a walk-in, hands a nervous first-timer a water, reads the room when a client is unhappy, or closes an upsell face to face. The AI covers the phone; your coordinator covers the humans in the building.
• It only knows what you've given it. The agent answers from your knowledge base — it won't invent a price for a service you never listed, which also means it can't answer questions about things you forgot to include. Good vendors catch those gaps in call reviews and add them.
• It's not for emergencies. A caller describing an urgent medical situation gets directed to emergency services and flagged to your team immediately — not triaged by software.

If a vendor's pitch has no "can't" list, that's the first thing to be suspicious about.

What happens when an AI receptionist can't answer?

This is the question that separates serious products from demos — because every AI receptionist eventually hits a call it can't handle. What matters is what happens next.

A properly built agent has a defined path for each failure mode:

• Question outside its knowledge base: it says so plainly, then offers the exits — transfer to your team, take a message, or book a consult where the question can be answered properly. What it never does is guess.
• Clinical question: redirected to a licensed provider, every time.
• Caller wants a human: immediate transfer during business hours; after hours, a structured message with a committed callback — not a dead end.
• Audio it can't parse: the agent asks the caller to repeat once or twice, then falls back to the human path rather than looping the caller through "sorry, I didn't get that" purgatory.
• Caller is upset: de-escalation isn't software's job. The protocol is a fast, warm handoff to a person.

Behind those paths there should be a human feedback loop. Cordiva is a managed service: our team reviews calls, catches the questions the agent couldn't answer, and adds them to the knowledge base — so "couldn't answer" becomes "answers now" within days, not quarters.

When you evaluate vendors, ask for the escalation protocol in writing: what triggers a transfer, what happens after hours, and who reviews the calls that fell through. If the answer is a shrug, it's a shrug at 9 PM on a Saturday too.

The three objections every med spa owner raises — taken seriously

"My patients will hate talking to a robot." The honest answer: some callers do prefer humans, and any vendor waving a "callers love our AI" satisfaction stat should be asked where it came from. But the comparison isn't AI versus your best receptionist — it's AI versus whatever answers when she's in a treatment room. In a 411 Locals study of 85 small businesses (2016), only 37.8% of inbound calls were answered live; the rest hit voicemail or got nothing. And 45% of appointments booked through Zocdoc happen after hours (Zocdoc platform data) — when the alternative isn't a warm human, it's a beep. There's also a group of callers for whom the AI is simply the better experience: 35% of Hispanic adults prefer a Spanish-speaking provider for routine care — 81% among Spanish-dominant Hispanics (Pew Research Center, 2022) — and 22.6% of Floridians age 5+ speak Spanish at home (U.S. Census Bureau, 2020–2024 ACS). A bilingual agent serves them in their language on every call (why that matters in Florida); an English-only front desk doesn't. Design matters too: the agent identifies itself, and a human is always one request away — no trapping people in bot jail. The full missed-call math lives in our missed-call statistics breakdown.

"It'll say something wrong and embarrass us — or worse." An unconstrained chatbot on a phone line absolutely would. That's why serious agents are scoped, not clever: a closed knowledge base, hard refusal on clinical questions, defined transfer paths, and a test battery before launch. With the average med spa visit worth $527 (AmSpa 2024 Medical Spa State of the Industry Report), both failure directions are expensive — the wrong answer that damages trust, and the unanswered call that books elsewhere. The fix for this objection isn't trusting the AI; it's verifying the vendor's testing — more on that in the next section.

"Is it even compliant to let AI answer patient calls?" It can be, with paperwork many vendors don't offer. If your med spa is a HIPAA covered entity, a vendor answering your calls is a business associate and must sign a BAA — and whether you're covered at all turns on how you bill, not what you inject; here's the full breakdown. Even outside HIPAA, Florida law already treats your clients' treatment information as protected: the Florida Information Protection Act (Fla. Stat. § 501.171) covers medical-history data held by any Florida business and runs a 30-day breach-notification clock, and Florida's practitioner-records law (§ 456.057) makes patient records confidential because your clinicians are licensed — regardless of how you bill. Cordiva signs a BAA on every plan, including the $299 base; our compliance statement is at /hipaa, and the deep dive is in HIPAA-compliant AI receptionists, explained.

How accurate are AI phone agents?

Here's the uncomfortable truth about this question: there is no standardized public benchmark for AI phone agents. When a vendor says their agent is "99% accurate," that could mean transcription quality, task completion, or answer correctness — three different things — measured on tests you'll never see.

So instead of a number, here's what actually determines whether an AI phone agent is accurate in practice:

• Scope. An agent that handles one vertical — med spa calls, nothing else — has a smaller universe of things to get wrong than a general-purpose bot with a med spa page bolted on.
• The knowledge base. The agent is exactly as accurate as what it was given. Setup quality is accuracy — which is why onboarding should happen with your team, not through a self-serve form.
• Refusal behavior. Counterintuitively, the most accurate agent is the one most willing to say "let me get you a human." An agent that always produces an answer is less trustworthy, not more.
• Testing, before and after launch. Scripted test calls before going live, and human call review afterward — because accuracy isn't a launch-day property, it's an ongoing practice.

Our position: any vendor can claim their AI behaves. Ask them to show the methodology. We publish ours — what a Cordiva agent must pass before it takes a single live call — at how we test. If a vendor won't show you theirs, you've just learned something about their accuracy claims too.

The fastest way to understand it: hear one

You can read explanations of how an AI receptionist works all day, or you can spend two minutes listening to one handle a med spa call. The second option settles more.

Book a 15-minute discovery call and we'll play you a live Cordiva agent built for a practice like yours: hear it answer, book into a real calendar, switch to Spanish mid-call, and refuse a medical question. If it fits, you're live in 7 business days, from $299/month with the BAA included — see what's in each plan or the full AI receptionist for med spas page.

Frequently asked questions