Guest Feedback Voice Agent — Cost-Engineered Phone AI for Hospitality
TL;DR
This is a cost-engineered voice agent for hospitality that calls guests after checkout to collect feedback — built so the per-call cost stays near zero. The trick is pre-caching: the agent's standard voice lines are generated once with ElevenLabs and reused, so text-to-speech spend on a typical call is almost nothing. Claude Haiku, a small fast model, steps in only when a guest says something off-script and a dynamic response is genuinely needed. Calls run over Twilio, both inbound and scheduled outbound, and cohosts get instant WhatsApp notifications with the feedback summary. Most voice-agent projects die on unit economics — this one was engineered backwards from cost per call. If you want always-on phone automation but the per-minute pricing of typical voice AI stacks scares you, this architecture is the answer. It also pairs naturally with the rental platforms I build, closing the loop from booking to post-stay feedback automatically.
How is the per-call cost kept near zero?
Three decisions. First, pre-cached ElevenLabs audio for every standard line — greeting, questions, confirmations — generated once, replayed forever. Second, Claude Haiku instead of a large model, invoked only for off-script replies rather than every turn. Third, tight call design: the feedback flow is short and structured, so most calls never leave the cached path.
What does the guest experience look like?
Shortly after checkout, the guest gets a friendly call asking about their stay. Answers are captured and summarized, and the cohost receives a WhatsApp notification immediately — so a bad-stay signal can be acted on the same day, not discovered in a review a week later.
Can the same pattern do more than feedback?
Yes — the inbound + scheduled-outbound Twilio architecture generalizes to booking confirmations, pre-arrival calls, and upsells. Feedback was the first deployment because it has clear value and zero risk to the booking funnel.
Stack
Outcomes
FAQ
What does a call actually cost to run?
Telephony minutes are the main cost; AI spend is engineered close to zero via cached voice lines and a small model used only when needed. Exact numbers depend on call volume and geography — covered in scoping.
Need something similar?
I ship a working MVP in days and iterate weekly — AI agents, voice AI, CRM implementations, and automation pipelines, fully remote.
hammadnaazkhan@gmail.com