AI Receptionist That Answers and Books
How we gave a wellness franchise a receptionist that picks up the calls the front desk can't, knows members by name, and books, moves, or cancels sessions in the live schedule while the caller is still on the line.
54 days
From rebuild plan to production calls
151
Real guest conversations in the first 5 weeks
16
Booking and texting tools on every call
CLIENT OVERVIEW
A wellness franchise with more than 40 open locations, each with its own front desk and its own schedule of private treatment rooms. When the desk couldn't pick up, the call forwarded to an answering service that could only send a text. The brand had already built its own AI receptionist in-house on a large cloud contact-center platform, with a 500-question knowledge base drawn from years of real call transcripts. It was close on logic but stuck on voice quality, a 2 to 3 second delay on every reply, and a permissions block that kept it from booking. They asked us to get a receptionist on the phone that could actually book.
THE PROBLEM
What was broken
Most calls never reached a person
Front desk staff run the treatment rooms, check guests in, and check them out. Phones ring while they're busy. Unanswered calls went to a service that replied by text, so a guest who called to talk got a message instead, and replies to it sat in a separate inbox nobody had time to watch.
Callers don't want to book online
The text fallback sent people to an online booking page that asked them to log in. Members who called specifically to book over the phone were back to hunting for a password. Every handoff like that is a booking that might not happen.
The in-house AI build had stalled
The first version ran through several layers of contact-center plumbing. Replies lagged 2 to 3 seconds, the voice didn't sound right, and a blocked cloud permission meant the booking tools couldn't be connected at all. The hard work, the business rules and the booking integration, was done. The platform around it was the problem.
THE GOALS
What success looks like
Answer every forwarded call with a natural voice that replies without a multi-second lag
Recognize returning members from caller ID and book them on the call against the live scheduling system
Never state a price, promotion, or booking the system didn't confirm
Give each spa one inbox for calls, texts, and callback requests, and make adding a spa a half-day setup
OUR APPROACH & SOLUTION
Keeping the hard-won rules and replacing the plumbing
We kept what the brand had already earned: its booking rules, its 500-question knowledge base, and its booking system integration. We replaced the contact-center stack with a realtime voice model behind a small gateway we control, then shipped in phases and tested every change against recordings of real calls before it reached a guest.
Phase 1
Voice Gateway and Caller Recognition
The phone line streams audio through our gateway to a speech-to-speech voice model, so there's no transcribe, think, and speak relay adding delay. Before the greeting plays, the gateway looks the caller up across every location, so a member hears their name. It was running on a staging server 9 days after the plan was written.
Phase 2
Real Booking on the Call
We gave the receptionist 16 tools for live availability, booking, rescheduling, cancelling, and texting. A member with a card on file books on the call, with the card held as the booking guarantee and payment taken at the front desk after the visit. Anyone else gets that spa's booking link by text, but only after a real open time has been offered. The receptionist can't say a guest is booked unless the scheduling system confirmed it. The same rules run a two-way text channel on the same number.
Phase 3
Production Launch and the Front Desk Inbox
We stood up a production stack where a deploy waits for live calls to finish, and the first spas started forwarding calls 54 days after the rebuild began. Every call and text lands in the dashboard staff already use, with callback requests flagged. When a staff member replies, the AI steps back for 30 minutes.
Phase 4
Hardening From Real Calls
We reviewed all 151 real guest conversations from the first 5 weeks, traced every loop, wrong answer, and silence back to its cause, and shipped 25 fixes. Each fix carries a replay test built from the actual call it broke on, and a CI gate runs the simulators and booking-system contract tests before every release.
Phase 5
Location Rollout
Each spa has its own details, opening line, hours, and room schedule in settings, and staff can test-call the receptionist from the browser. A written setup guide brings another spa online in about half a day of hands-on work. Twelve locations were synced and onboarding by October 9, 2026, and 6 of them were taking live calls.
WHAT WE BUILT
Features and capabilities
01
Answers Forwarded Calls
When the front desk doesn't pick up after about 3 rings, the spa's landline forwards the call to the receptionist. Audio streams to a speech-to-speech voice model, and callers can interrupt it mid-sentence the way they would a person. Every call opens by telling the guest they're talking to an AI.
02
Knows Which Spa Was Called
Each spa has its own number, and the receptionist identifies the spa whether the call was forwarded from the landline or dialed directly. It answers with that spa's name, address, hours, and notes, so a guest in one city never hears another location's details.
03
Greets Members by Name
Before the greeting plays, the gateway matches the caller ID against client records across every location. A returning member hears their name in the first line of the call, and the lookup works no matter which location the member usually visits.
04
Books, Moves, and Cancels Sessions
The receptionist checks live availability and books, reschedules, or cancels a session in the scheduling system while the guest is on the line. A guest without a card on file gets that spa's booking link by text after a real open time has been offered. The booking tool returns one of 18 typed results, the receptionist tells a guest they're booked only on a confirmed result, and the test suite fails if any result has no test.
05
Two-Way Texting on One Number
Guests can text the same number they called. A text agent with the same tools and rules handles the thread and remembers earlier calls, and it can text booking, gift card, and package links and spa details on request.
06
Front Desk Inbox and Callbacks
Every call and text lands in the dashboard staff already use, and a guest who was promised a callback stays flagged until someone resolves it. When a staff member replies in a thread, the AI pauses for 30 minutes. The client's in-house developer built the inbox screens and call-summary emails, and we built the API behind them.
07
Per-Spa Settings and Test Calls
Each spa has its own display name, opening line, and location details in settings, beside a read-only view of what the receptionist knows. The brand's prompt is versioned, so a saved change reaches every spa on the next call or text without a deploy. Staff can test-call the receptionist from a browser microphone as any client.
08
Replay Tests and a Release Gate
Offline simulators play a caller against the receptionist and replay recorded production calls. Before every release, a CI gate runs type checks, both simulators twice, unit tests, and recorded contract tests against the scheduling system. Production runs on 2 tasks, and a deploy waits for live calls to finish.
KEY TAKEAWAYS & IMPACT
What changed for this business
54 days
From stalled project to production calls
The rebuild plan was written in early July and real guests were talking to the receptionist on September 1. It reused the business rules and booking integration the brand had already built, so nothing they'd invested in was thrown away.
151
Real guest conversations reviewed in 5 weeks
The first two spas logged 151 real guest conversations in the first 5 weeks, and every one was reviewed. Problems became replay tests, so a failure a guest hit once can't quietly come back in a later release.
16
Tools that book without the front desk
Members with a card on file book on the call, and everyone else gets a booking link for the right spa by text. The 16 tools sit on 20 calls to the live scheduling system, and on live calls they return in under 1.2 seconds at the 95th percentile.
THE OUTCOME
The wins
-
A call the front desk can't pick up reaches a voice that answers questions and books, instead of an automated text.
-
Calls, texts, and callback requests land in the dashboard staff already use, so guest replies no longer sit in a separate inbox.
-
Replies come from a speech-to-speech model with no relay in between, and booking-system calls return in under 1.2 seconds at the 95th percentile on live calls.
-
The receptionist decides whether a caller said yes in under a second at the median, with no false yes across two runs of a 70-case test set.
-
A prompt change reaches every spa on the next call or text without a deploy, and staff can test it from the browser before guests hear it.
-
Rollout is opt-in per franchisee, and each spa that opts in takes about half a day of hands-on setup.