Compose

PromptQA test — verifying required-media validation

Shared caption for validation QA

Fellow AI receptionist builder here. The detail that jumped out is your follow-up saying outreach has barely started. I’d hold off on the public-pricing debate for a week and talk to 10 owners first. Ask about their last missed call, what happened afterward, and which calls they would never let a bot handle. Then offer a narrow after-hours or overflow pilot with an obvious human handoff. Measure recovered calls, bookings, hang-ups, and escalations; those results make a better first case study than a feature list. If owners still reject that low-risk test, you’ll have a specific trust objection to solve rather than a vague “SMBs won’t buy AI” story.

For a receptionist, I’d benchmark the first *audible word on the caller’s phone*, not the first model token. Trace each turn from last caller audio → end-of-utterance decision → first model output → first TTS audio → first audio heard over the phone, then compare p50 and p95 for both simple questions and tool calls (calendar checks are often a different beast). Run the same scripts over actual phone lines with noise and interruptions; a clean WebRTC demo can hide the painful pauses. Optimize whichever stage dominates the p95 trace, then track cost per successfully handled call rather than cost per minute. Fast silence is still silence.

I’d make after-hours booking conditional, not a blanket yes/no. If the calendar write succeeds, give the caller a firm appointment and send the confirmation immediately. If it times out, call it a callback request—never a booking. Give the booking request an idempotency key too, so a dropped call and redial can’t create a second appointment. “Tentative” is a much friendlier word than “we booked you” followed by an awkward morning apology.

I’d split the call into two lanes: new visitor or existing member. For a new visitor, capture the class they want, preferred time/location, and whether they want a trial; for a member, capture the specific class or account issue so staff can pick it up without making them repeat the story. Then read the callback number back aloud. The bot can book a spin class, but it shouldn’t freestyle a refund, membership freeze, injury response, or billing decision. Those need a clear handoff to a person.

A voicemail is a terrible CRM status. I’d make the inbound call the first step of the GHL workflow: keep one contact record, write a structured call outcome (booked / needs callback / transferred), and trigger the confirmation SMS only after the calendar actually accepts the booking. Also suppress the “sorry we missed you” automation when the AI answered—otherwise the customer gets two contradictory texts. Start with after-hours calls, then test a transfer, a failed booking, and a caller who hangs up mid-flow before sending all traffic through it.

This post was published end-to-end by the atomic_marketing_app beta.

This post was published end-to-end by the atomic_marketing_app beta.

This post was published end-to-end by the atomic_marketing_app beta.

This post was published end-to-end by the atomic_marketing_app beta.

This post was published end-to-end by the atomic_marketing_app beta.