AI Voice Agents for Medspas — A Field Test of Vapi, Bland, and Synthflow in 2026

I spent six weeks in 2026 running the same 100 inbound medspa calls through Vapi, Bland, and Synthflow back to back. Same scripts. Same booking system. Same handoff rules. What follows is the field-test report. Which platform booked more appointments, where each one broke, what they cost, and the one I now install by default on every medspa engagement I take.

Why I ran this test in the first place

Every medspa I audit has the same hole in their funnel. Roughly 60% of inbound calls land outside business hours, on weekends, or during back-to-back appointments when nobody can pick up. Those calls go to voicemail. Voicemail gets returned the next day on average. By then the patient has already booked with the competitor down the street who answered first. The data on this is brutal: 78% of patients book with whoever responds fastest, and the medspa that takes 12 to 24 hours to call back loses most of those leads on contact, not on price.

I started looking at AI voice agents in early 2025 because every alternative I tested (24/7 human answering services, after-hours auto-attendants with shallow booking links, voicemail-to-text fallbacks) failed on the same axis. They captured the lead. They did not book the appointment. The whole point of an AI voice agent is that it does the second thing, not just the first. By spring 2026 there were three platforms credible enough to evaluate seriously: Vapi, Bland, and Synthflow. I built production-grade agents on all three, pointed them at the same medspa, and recorded every call.

The test setup, in detail

The medspa is a single-location practice in the Southwest US, four injectors, 380 booked appointments a month, roughly $96K average monthly revenue. They get about 520 inbound phone calls a month according to their Twilio dashboard. Their existing front desk picks up roughly 41% of those calls live; the rest go to voicemail, get returned within 6 to 18 hours, and convert at about 22%.

I forked their main Twilio number into three test routes, one per platform, and load-balanced 100 inbound calls each across a four-week window. Every agent had:

• The same knowledge base: 80-page document covering services, pricing ranges, provider bios, parking, prep instructions, and the FAQ list the front desk reads from
• The same handoff rules: transfer to live human if caller asks, if the caller mentions a medical concern, or if the agent confidence drops below 70% on next response
• The same booking integration: function calls into Aesthetic Record via GoHighLevel webhooks, with a 10-minute hold-and-confirm pattern to prevent double-bookings
• The same voice: ElevenLabs Flash v2.5, warm female voice, 1.0x speed, no emotional dial above neutral-warm
• The same opening line: “Hi, this is the booking assistant for [practice name]. I can answer questions about treatments and book your appointment. How can I help you today?”

I scored every call on six dimensions: booked-appointment outcome, qualified-lead outcome, handoff quality, first-token latency, total call duration, and patient-recognition rate (whether the caller realized or asked if they were talking to AI).

Vapi — what I found

Vapi felt the most production-ready of the three. First-token latency averaged 680ms across the test calls, which is the threshold below which most callers do not perceive any AI lag. The function-calling architecture is the cleanest of the three for medspa workflows; I had the agent calling Aesthetic Record availability, holding slots, collecting deposits via Stripe, and firing the confirmation SMS through Twilio inside one conversational turn without the caller hearing a pause longer than two seconds.

The booking results: 18 confirmed appointments out of 100 calls. 31 additional qualified handoffs (callers who needed a human for a consultation question, complex pricing on combination treatments, or a clinical concern that fell outside the agent’s scope). 12 callers asked some version of “is this a real person?” and continued with the AI after the disclosure. 4 callers hung up on the disclosure and called back during business hours.

The remaining 35 calls broke down into: 14 wrong-number or sales calls (filtered correctly), 11 callers asking generic questions and not booking, and 10 calls where the agent’s voice cut out mid-sentence due to Twilio connection issues that were not Vapi’s fault but were Vapi’s experience to fix. I had the agent script a recovery line (“I lost you for a moment, can you repeat that?”) and the cut-out rate dropped to 3 in the second half of the test.

Pricing landed at $0.07 per minute of voice usage on Vapi’s pay-as-you-go tier, plus $99 a month for the platform fee at the volume we were running. Total monthly cost on the production deployment runs $180 to $240 depending on call duration. Per booked appointment, that worked out to about $11, against an average ticket value of $200 at this practice. Net economics are not even close.

The one place Vapi lost me early was setup complexity. Out of the three, it requires the most technical scaffolding. There is no drag-and-drop builder. You write the assistant config in JSON or via the dashboard, you wire your own function endpoints, and you have to think carefully about the prompt structure to get function-calling reliability above 95%. I have built five Vapi agents now and the per-build time has dropped from two weeks the first time to about six focused hours, which is what I budget on a Sprout Sage AI automation engagement now.

Bland — what I found

Bland is the platform I would pick if I were running outbound campaigns more than inbound. It has the most reliable dialer (95%+ connection rate on outbound at scale in my testing), the lowest per-minute cost on outbound specifically ($0.06/min), and the cleanest answer-detection logic so you do not burn voice budget on voicemails.

On the inbound side of this test, Bland posted 11 confirmed bookings out of 100 calls. First-token latency was 780ms on average, slightly behind Vapi but still under the one-second perception threshold for most callers. The function-calling reliability ran around 91% in my test, meaning roughly one in ten booking attempts failed mid-flow and required a handoff to recover. Most of those failures were calendar-API edge cases (same-day slots that became unavailable between the lookup and the booking call) that I patched with a re-check before confirm pattern in week three of the test.

The qualified-handoff count was higher than Vapi at 37, which is partly because Bland’s confidence-scoring is more conservative; the agent kicks to handoff faster when the conversation goes off the script it has been trained on. That is a feature, not a bug, on medspa work where a wrong answer about a medical question is worse than a transferred call.

The voice quality on Bland is closer to “generic call center” than “warm receptionist” out of the box, even with ElevenLabs in the loop. I had to tune prosody and pacing more aggressively to get past the uncanny-valley feeling. After tuning, the patient-recognition rate (callers asking “is this AI?”) was actually lower on Bland than on Vapi, at about 9%. My read: when the voice quality is a hair worse, callers stop questioning whether it is a person because they have already decided it is not, but they continue the conversation anyway because the agent is competent.

Where Bland really earned its slot in my stack is on outbound. I have a separate Bland deployment running no-show recovery at the same medspa now: T+72hr outbound call to any patient who did not respond to the SMS recovery sequence in Flow 2 of my no-show framework. Connection rate on outbound: 89%. Rebook conversion on connected calls: 38%. That single workflow recovers about $4,400 a month in otherwise-lost revenue at this practice.

Synthflow — what I found

Synthflow is the platform I recommend to medspa owners who want to build an agent themselves over a weekend without writing code. The UI is the cleanest of the three, the templates for booking workflows are usable out of the box (with editing), and the price floor is friendly: their starter tier covers 200 minutes a month for $29.

The catch is what you give up at the top end. First-token latency averaged 950ms in my test, right at the edge of the perception threshold; about 1 in 4 callers noticed a beat of dead air at the start of each turn. Function-calling reliability was 88%, the lowest of the three. The booking integration into Aesthetic Record required a Zapier intermediate layer because Synthflow does not have a native webhook to the depth Vapi does, and the Zapier hop added 400 to 600ms of latency to every booking step.

That said, Synthflow booked 14 appointments out of 100 calls in the test. Not as good as Vapi, better than Bland on inbound. The handoff logic is the simplest to configure of the three, which I think helps the booking number indirectly: the agent kicks to a human earlier on borderline calls, and the human closes some of them.

For a medspa owner who wants to set up an AI voice agent without an agency or an engineer, Synthflow is the right answer. The total weekend-build time is about 6 to 10 hours if you have your service list, pricing, and FAQ already documented. The platform fee at production volume sits around $99 to $149 a month. The trade-off you accept is some latency and slightly lower booking reliability, both of which matter less than having the system live and capturing leads at all. A 14% booking rate on after-hours calls that previously got nothing is still a massive uplift.

The side-by-side numbers

If you are a medspa owner reading this and trying to make a decision: pick Vapi if you have someone technical who can configure it (or hire someone like me); pick Synthflow if you want to ship something live this weekend; add Bland on top of either one for outbound campaigns. The stack I install on Sprout Sage builds is Vapi inbound plus Bland outbound, both routed through the same GoHighLevel sub-account.

Quick gut check

Before you call me about this, ask yourself: how many calls is my practice missing right now after 6pm and on weekends? If you do not know the number, your Twilio dashboard or your phone provider’s call log has it. Most medspas I audit are missing 200 to 400 calls a month. At a $200 avg ticket and a 25% booking rate, that is $10,000 to $20,000 a month in revenue sitting in voicemail. Book a 30-min audit and I will pull your number and show you the math on your specific practice.

Where AI voice agents still lose to humans

I want to be clear about what AI does not do well, because over-promising on this is how agencies lose medspa clients in month three.

First, sensitive consultations. A first-time Botox patient who is nervous, asking five layered questions about risk and recovery, is a conversation a human handles better. The AI can capture the intent and book a free consult, but the close on the consult is human work. I script my agents to detect “new patient” plus “consultation” intent and route directly to the front desk’s calendar without trying to upsell.

Second, complex pricing on combination packages. A patient asking “how much for Botox plus filler plus a hydrafacial on the same visit” is not a calculation the agent should do live, even with pricing in the knowledge base, because medspa pricing tends to have provider-level and promotional overlays that change. The agent says “let me have a treatment coordinator call you with exact pricing in the next hour” and dispatches an SMS. Conversion on that handoff in my tracking is 71%, vs. about 12% on a phone-tag voicemail.

Third, complaints and medical concerns. Any caller mentioning post-procedure pain, swelling beyond expectations, allergic reaction, or any medical concern gets routed immediately to the on-call provider’s number with full context attached. The agent does not attempt clinical reassurance. This is the single most important guardrail you build into a medspa AI agent and the place where badly-configured agents create real legal exposure.

Fourth, the front-desk relationship. The best medspa front desks know their patients by name, remember preferences, ask about kids, soften the experience. The AI cannot do that and should not try to fake it. I script my agents to be transactional, warm, and efficient, not to pretend they remember the patient’s last visit. Patients respect a competent AI more than they respect a clumsy one trying to sound like a friend.

The integration stack I run on a Sprout Sage build

Here is the exact wiring I use when I install AI voice on a medspa, end to end. I am putting this in the post because the platform choice alone does not produce results; the integration is where most builds fall apart.

1. Twilio number with 10DLC registration as the inbound and outbound layer. The client owns the number, not me, not GoHighLevel. This matters when contracts end.
2. GoHighLevel sub-account (“Sage OS” snapshot) as the CRM and orchestration layer. Every call gets logged here. Every booking ID lives here. Every outbound campaign fires from here.
3. Vapi agent wired to the Twilio number for inbound, with function-call endpoints into GHL for contact lookup and into Aesthetic Record for calendar availability and booking.
4. Bland agent wired separately for outbound campaigns (no-show recovery, dormant reactivation, post-treatment check-in), triggered by GHL workflows.
5. Aesthetic Record as the PMS and clinical system. Bookings created by either agent land here; all chart data stays here.
6. Klaviyo for email nurture before and after the call (see my breakdown of the seven flows that took one practice from $8K to $28K a month in email).
7. NiceJob for review request automation, triggered when AR marks a treatment complete.
8. Stripe wired through GHL for deposit collection at the moment of booking on treatments above $300.

The full stack at scale costs the medspa $515 to $1,172 a month in software, plus my retainer. My implementation fee is $7,500 flat for the 60-day build, plus a $1,997/mo management retainer with a 3-month minimum. The math on the anchor case study (linked in the next section) shows this stack hitting a 30% revenue lift inside the first 60 days at a typical $80K/mo practice.

The 30% revenue lift math, in case you have not seen it

The AI voice agent on its own does not get you to 30%. It contributes about 4 percentage points (after-hours capture of leads that would otherwise have gone to voicemail). The other 26 come from the full stack: SMS reminders (6 points from no-show reduction), rebook flows (8 points), reactivation (5 points), review-driven inbound (3 points), and membership conversion (4 points). I broke this down in detail on the medspa AI 30% lift case study, lever by lever, with the source benchmarks for each number.

The point of putting it here is to set expectations honestly: an AI voice agent is one lever in a stack, not a magic product. The medspas that get 30% lifts are the ones that install the whole stack, not the ones that bolt one agent onto a broken funnel.

What to ask any vendor pitching you AI voice in 2026

Five questions I would ask before signing any AI voice agent contract:

• Whose phone number is it? If the agency owns the Twilio number, you do not own the asset. You should always own the number. Period.
• Does the agent book directly into my PMS or does it just send a lead form? The first answer is what you want. The second is a CRM with extra steps.
• What is the handoff protocol when the agent cannot answer? If they cannot explain this in a sentence, they have not built one.
• Show me three real call recordings from production medspa accounts. Not demos. Real calls. If they cannot share (NDA reasons), the agent is probably not in production anywhere yet.
• What happens to my data if I cancel? Call recordings, conversation logs, lead records. You should have export rights to all of it.

The 60-day build I run

If you want this installed and live at your practice, the 60-day build I run looks like this:

• Week 1: Discovery, access audit, Twilio 10DLC registration submitted (this is the long lead item, do it day one)
• Week 2: GoHighLevel sub-account configured, Aesthetic Record sync, Klaviyo connected
• Week 3: Vapi agent built and trained on the FAQ + booking rules + handoff scripts
• Week 4: End-to-end testing on real (but contained) call volume, staff training
• Week 5: Full launch, all inbound routes pointed to the new stack
• Week 6: First reporting, agent tuning based on the first 100 real calls
• Week 7: Bland outbound layer (no-show recovery, reactivation) goes live
• Week 8: Handoff, dashboards trained, monthly retainer begins

The full week-by-week breakdown with deliverables is on the AI automation service page. I take 5 medspa clients at a time, maximum. If you want a slot, the easiest path is to book a free 30-minute audit.

FAQ

Want me to install this at your practice?

I run five medspa engagements at a time, maximum. The 60-day build is $7,500 setup plus $1,997 a month after that (3-month minimum). What you get: the full Vapi inbound agent, the Bland outbound agent, the six core automations (booking confirmation, no-show recovery, post-treatment follow-up, review request, membership upsell, dormant reactivation), and the GoHighLevel + Aesthetic Record + Klaviyo + NiceJob stack wired into your existing phone line. You own everything.

Book a free 30-min call →   or call me directly at +91 97297 12388 or WhatsApp.