Build your own AI phone receptionist (for tech-savvy gym owners)
Last checked 2026-10-16. Every claim links to its source.
Getting an AI to answer your gym's phone is easy now. Knowing it says the right class times, the right trial offer and the right cancellation policy, every time, is the hard part.
Written by a gym owner who has built his own tools. This one is for owners who are comfortable with a little code and want proof, not vibes.
The core difference: loading facts vs. proving them
Turnkey receptionists let you load your business info and go live fast. Frontdesk AI, for example, takes text, website URLs and PDF/DOCX/TXT uploads plus Q&A pairs. [source]
Some turnkey tools let you test before going live. Smith.ai lists a Quality Studio and simulated calls on every plan, and Goodcall's plans include an allowance of test runs. [source] [source]
What developer platforms add is a test suite you write, version and rerun as often as you like. Retell's simulation testing grades saved test cases pass or fail, runs them as batches, and is meant to be rerun after every prompt or flow change. [source]
Vapi's Evals let you script a conversation and check each important turn with exact matching, regex, an AI judge or tool-call checks, and run them from CI. [source]
Path 1: Turnkey receptionist
Examples: Smith.ai: Free plan with 25 calls/month ($3.00/call after), Pro $150/month. Dialzara: Business Lite $29/month (60 min), Business Pro $99/month (220 min). Goodcall: Starter $79/month. Frontdesk AI: $99/month or $79/month billed annually, 200 voice minutes. [source] [source] [source] [source]
What you get: Fast setup, a knowledge base, transfers and message taking. Testing is whatever the vendor's tools offer, inside their dashboard.
Call: Right for most gyms. You still need to call it yourself after every change to your schedule or offers.
Path 2: Developer platform (Retell AI or Vapi)
Pricing: Retell is pay-as-you-go, listed at $0.07–$0.31/minute, with Simulation Testing listed as a platform feature. Vapi lists $0.05/min for Vapi hosting, plus transcriber, model and voice costs, with $5 free credits to start. [source] [source]
What you get: Your own prompt, tools (booking, transfer, SMS), and test cases you control. Retell test cases can mock functions so a test doesn't create a real booking, and can be exported as JSON. [source]
Call: Worth it if you want to prove accuracy and you're willing to own the prompt, the tools and the test suite.
Path 3: Add an independent tester
Cekura: A hosted QA tool with documented integrations for Retell, Vapi, ElevenLabs, LiveKit, Pipecat, SIP and custom setups. Its pay-as-you-go plan starts at $0 and lists $0.25 per voice testing minute and $0.05 per monitored call. [source] [source]
voicetest: Open source (Apache 2.0). Imports agents from Retell, Vapi, LiveKit, Bland and Telnyx, simulates conversations, judges them with LLMs, and runs in GitHub Actions. [source] [source]
Call: Useful when you want tests that don't live inside one vendor, or want to compare platforms with the same test suite.
Vapi heads-up: use Evals and Simulations
Vapi's old Test Suites docs page now redirects to its voice-testing guide, which splits testing into Evals (check a decision at a known point in the conversation) and Simulations (check the outcome of a whole conversation). If you find an older tutorial built on Test Suites, start from the current guide instead. [source]
Build a gold set
A gold set is your answer key: the calls your agent must get right. Each item has four parts: what the caller says, the facts the answer must include, the claims it must never make, and the expected action (answer, book, transfer, or take a message).
Seed it from 50–100 real calls. Your call log already tells you what people ask; don't invent questions from a desk chair.
Test facts and behavior separately. "Says Saturday open gym is 9–11am" is a fact check. "Transfers an angry member instead of arguing" is a behavior check. Mixing them hides which one broke.
Keep the gold set in your own repo, as plain files. Platforms change; your answer key shouldn't be locked in one vendor's dashboard. Retell exports test cases as JSON, and voicetest reads tests from a tests.json file. [source] [source]
Rerun it on every change: new schedule, new promo, new prompt, new model. Retell's docs recommend rerunning the suite after every prompt or flow change and judging a scenario on its pass rate across runs, because the simulated caller and the grader are both LLMs. [source]
Example gold items for a gym
Class times. Caller: "What time is the Saturday kids class?" Required: the exact time from your current schedule. Forbidden: any time not on the schedule; "I think." Action: answer.
Trial offer. Caller: "Do you have a free trial?" Required: your actual offer and how to book it. Forbidden: a discount or length you don't offer. Action: book the trial (with the booking tool mocked in tests).
Cancellation policy. Caller: "How do I cancel my membership?" Required: your notice period and how to submit it. Forbidden: "You're cancelled"; waiving fees. Action: take a message or transfer, per your policy.
Angry member. Caller: "You charged me twice and I want my money back now." Required: acknowledge, collect name and callback number. Forbidden: promising a refund or blaming the member. Action: transfer, or take a message for a manager.
The babysitting bill
Every schedule change is a gold-set change. If you don't update the answer key, your tests will pass while the agent reads last season's timetable.
Test runs aren't free. Retell's simulated caller bills per message, and Goodcall meters test runs against an AI usage allowance. [source] [source]
Can Retell AI run regression tests on a phone agent?
Yes. Retell's simulation testing grades saved test cases pass or fail, runs them as batches, supports function mocks, and is available over the API for CI. [source]
Are Vapi Test Suites still supported?
Vapi's Test Suites docs page now redirects to its voice-testing guide, which covers Evals for turn-level decisions and Simulations for whole-conversation outcomes. [source]
How much does Smith.ai's AI receptionist cost?
Smith.ai lists a Free plan with 25 calls/month ($3.00 per call after that) and Pro at $150/month. [source]
How much does Dialzara cost?
Dialzara lists Business Lite at $29/month with 60 minutes and Business Pro at $99/month with 220 minutes. [source]
Things change. Check with the vendor before you buy.