--- title: Voice AI Agent Vapi category: product entity_type: skill price: $15 canonical: https://forgehouse.ai/skills/voice-ai-agent-vapi/ lang: en hreflang_alt: https://forgehouse.ai/tr/skiller/voice-ai-agent-vapi/ last_updated: 2026-06-20 --- # Voice AI Agent Vapi > Vapi.ai + Bland.ai + Retell AI sesli AI agent kurulumu A production blueprint for building Turkish-language voice AI agents on Vapi.ai, with Bland.ai and Retell AI compared on cost and features. It wires appointment booking, inbound and outbound call automation, low-latency turn-taking, barge-in handling, consent-first call recording, calendar tool calling, and transcript storage into one operational pipeline measured by the minute. ## Use cases - Replacing reception phone load with an inbound appointment bot - Taking orders during seasonal or campaign call surges - Running outbound reminder and no-show prevention calls - Handing off to a live operator after repeated misunderstandings - Routing to WhatsApp fallback outside hours or on failure - Storing call transcripts with sentiment and intent analysis ## Benefits - Sub-500ms turn-taking keeps the conversation feeling human, not robotic - Callers are never trapped in a loop: three strikes triggers live transfer - Call recording stays consent-first, with PII redacted after the call - Phone-call handling cost drops to cents per minute versus staffed lines ## What’s included - Three-provider comparison across price, voice quality, latency, and webhooks - Full Vapi assistant config with Turkish neural voice and function calling - Signed webhook handler that records calls and transcripts to a database - Calendar tool-calling integration for live appointment booking - Consent flow that captures recording approval in the first seconds - Post-call sentiment and intent analysis with structured output ## Who it’s for Agencies and service businesses: clinics, salons, restaurants, florists, automating Turkish-language inbound and outbound calls with a consent-compliant voice agent. ## How it runs Past 800ms of response delay, callers decide they are talking to a robot. Building a Turkish-speaking phone agent that books appointments instead of losing calls means tuning for that number throughout: 1. Decides the provider against a measured comparison matrix (per-minute cost, Turkish voice quality, LLM flexibility, latency, webhook coverage, EU edge location, data processing agreement availability), with Vapi as the default backbone and the others reserved for bulk outbound or A/B testing. 2. Configures the assistant for natural Turkish conversation: Azure Neural Turkish voices matched to the sector demographic, low temperature for tool calling so an appointment never lands on a hallucinated date, endpointing around 350ms and a 100ms interruption threshold so the bot stops the instant the caller talks over it. Above 800ms response time the caller feels they are talking to a robot. 3. Places recording consent in the first three seconds as a state machine: the opening line asks for consent, keyword detection classifies the answer as positive, negative or unclear, a negative answer stops recording but continues the call anonymously, and an unclear answer re-asks once. 4. Wires the action layer through function calling: the booking tool writes directly into the calendar API with sanitized attendee data, and a 3-strike fallback transfers to a live human line the moment intent goes unrecognized three turns in a row or frustration keywords appear. The transfer counts as a success metric, not a failure. 5. Receives every call event on an HMAC-verified webhook and writes the full lifecycle to the database: call start, partial transcript chunks, function calls, end reason, duration, cost and recording URL, with the state log kept as a debugging trail. 6. Runs post-call analysis only when consent was positive: a small model classifies intent and sentiment, produces a 3-sentence summary, redacts personal identifiers from the stored transcript, flags calls needing human follow-up and triggers the downstream automation when an appointment was booked. ## FAQ ### My business takes calls in English, not Turkish, does this blueprint transfer? The architecture transfers, the tuning doesn't fully. Webhook handling, calendar tool calling, consent flow, and the three-provider comparison are language-neutral, but the assistant config ships with a Turkish neural voice and Turkish-first prompts you'd need to re-tune for another language. ### How does it keep calls from feeling robotic or trapping callers in loops? Through concrete mechanics, not model magic: sub-500ms turn-taking and barge-in handling keep the rhythm human, repeated misunderstandings trigger a three-strikes handoff to a live operator, and out-of-hours or failure cases route to a WhatsApp fallback instead of a dead end. ### Does it include the telephony minutes or a provider account? No. You open and pay your own Vapi, Bland, or Retell account, and per-minute call costs are billed by the provider. The blueprint supplies the assistant configuration, signed webhook handler, consent-first recording flow, and transcript pipeline around that account. ## Price $15, one-time, no subscription. VAT included. Related guide: [AI and LLM engineering](https://forgehouse.ai/guides/ai-llm-engineering/)