# P13 — Founder Guide: How Voice AI Actually Works

**Updated:** 30 August 2026 · Read after `BUSINESS_PLAN.md`. After this document you can hold a pilot call, read a platform dashboard, and understand every line of the cost math.

---

## 1. The pipeline: what happens inside one second of a call

When a caller speaks to an AI receptionist, three machines run in a relay, hundreds of times per call:

```
Caller speaks → ① STT ("ears") → ② LLM ("brain") → ③ TTS ("mouth") → Caller hears
                 speech-to-text     decides what       text-to-speech
                 transcribes the    to say / do        speaks the reply
                 audio to text      (book, answer,     in a natural voice
                                    transfer...)
```

| Stage | What it is | Typical providers | Typical cost |
|---|---|---|---|
| **STT** (speech-to-text) | Turns the caller's audio into text, live, while they are still talking | Deepgram, AssemblyAI, Whisper-class models | ~$0.004–0.01/min |
| **LLM** (large language model) | Reads the transcript + our instructions ("the prompt") and decides the reply and any action (book slot, take message, transfer) | GPT-4o-mini-class (budget) to GPT-4o/Claude-class (premium) | ~$0.003–0.16/min |
| **TTS** (text-to-speech) | Speaks the reply in a chosen voice | Deepgram Aura, Azure, Cartesia (budget); ElevenLabs (premium) | ~$0.01–0.10/min |
| **Orchestration** | The platform gluing all three together with the phone call, in real time | Vapi, Retell, Synthflow, Bland | ~$0.05–0.09/min |
| **Telephony** | The actual phone line and number | Twilio, Telnyx, platform-bundled | ~$0.01–0.02/min + number rental |

**Latency is the product.** A human notices a pause longer than ~1 second and starts saying "hello? hello?". Good 2026 platforms achieve ~500–800 ms voice-to-voice response. This is why we rent the platform instead of wiring STT+LLM+TTS ourselves: the platforms' entire engineering effort is squeezing that relay under one second. When you evaluate a platform or a config change, the first metric on the dashboard to check is **voice-to-voice latency**.

## 2. The vocabulary you'll actually use

| Term | Plain meaning | Why you care |
|---|---|---|
| **Barge-in / interruption handling** | The caller can talk over the AI and it stops talking and listens, like a polite human | The #1 difference between "wow" and "hate robots". Always on. Our answer to the robot objection |
| **Endpointing** | How the system decides the caller finished their sentence | Tuned wrong = AI interrupts mid-sentence or leaves awkward gaps |
| **Prompt** | The written instructions defining the assistant: identity, tone, what it knows, what it may and may not do | This is our craft. The prompt pack per vertical IS the product we accumulate |
| **Tools / function calling** | The LLM's ability to trigger real actions: check calendar availability, create a booking, send a WhatsApp, transfer the call | Booking = a tool call to Cal.com/Google Calendar. No tools = just a talking FAQ |
| **Warm transfer / escalation** | AI hands the live call to a human (the owner's mobile), optionally whispering context first | The safety valve. Every config has one |
| **Voicemail fallback** | If transfer fails and AI can't help, take a structured message and deliver it | Never dead-end a caller |
| **IVR** | The old "press 1 for sales" menu tree | What we are replacing. Useful contrast in the pitch |
| **SIP / VoIP** | Phone calls carried over the internet rather than the mobile network | Matters because some countries (UAE!) restrict it — see §4 |
| **Call forwarding / conditional forwarding** | Client's real number sends calls to our AI number — always, or only when busy/unanswered/after-hours | The standard onboarding pattern: client keeps their number, zero porting risk, instant rollback |
| **Concurrency** | How many simultaneous calls the assistant can take | Platforms include 10–20 free; an SMB rarely needs more |
| **Turn / conversation turn** | One caller-utterance + one AI-reply cycle | QA metric: bookings should complete in ≤ ~8 turns |

## 3. The platform landscape (researched 30 Aug 2026 — full pricing in `PRICING.md` and `COMPETITOR_ANALYSIS.md`)

Two layers exist. Never confuse them in a pitch:

**Layer 1 — Infrastructure platforms (what WE buy):**

| Platform | Model | Notes |
|---|---|---|
| **Retell AI** | $0.07/min voice engine + LLM + telephony ≈ $0.10–0.18/min all-in; pay-as-you-go, $2/mo numbers, free concurrency | Developer-friendly, strong dashboards (latency, sentiment), native booking integrations |
| **Vapi** | $0.05/min orchestration + you assemble STT/LLM/TTS ≈ $0.07–0.25/min all-in; cheapest with your own provider keys | Maximum control of cost and stack; more knobs to hold wrong |
| **Synthflow** | Usage pricing ≈ $0.12/min effective; **white-label/reseller toolkit ≈ $2,000/mo** | No-code, agency-oriented; the white-label price only makes sense at 15–20+ clients |
| **Bland** | $0.09/min base, tiered plans push $0.11–0.12; extra fees (transfer time, per-attempt) | Self-hosted stack pitch; fee structure needs careful reading |

**Layer 2 — Managed services (what we COMPETE with):** Smith.ai ($95–800/mo, per-call billing, human backup), Goodcall (from $79/mo), Rosie (from $49/mo), Slang.ai ($399–599/mo, restaurants), plus hundreds of agencies reselling Synthflow/Vapi wrappers. We are Layer 2, built on Layer 1 — exactly like they are.

## 4. Telephony basics — and the three-market reality

A voice agent needs a phone number the platform can control. That is trivial in some countries and near-impossible in others:

| Market | Reality (verified 30 Aug 2026 — details & sources in `COMPATIBILITY.md`) |
|---|---|
| **US** | Easy. Twilio local number ~$1.15/mo, inbound ~$0.0085/min, buy in minutes via API. This is why the LLC channel is the scale market |
| **UAE / Gulf** | Hard. Twilio does not sell UAE local numbers; the UAE actively blocks/penalizes VoIP-style calling on its numbers (TDRA regime). KSA numbers exist on Twilio but with heavy documentation requirements. Workarounds: **WhatsApp Business Calling API** (calls happen inside WhatsApp — no local number needed, inbound calls free) or an international/toll-free design with the partner |
| **Bangladesh** | Hard. BTRC regulates VoIP tightly; Twilio BD numbers carry documentation requirements and practical limits. Workaround: WhatsApp calling (WhatsApp is the default business channel in BD anyway) or conditional forwarding designs — assessed honestly per pilot |

**The pattern to remember:** the client always keeps their existing number. We add either (a) conditional forwarding to a platform number (US), or (b) a WhatsApp call button/number (BD/Gulf). Rollback is one dial-code — that is the go-live safety story in `DEPLOYMENT.md`.

## 5. Per-minute cost anatomy (one real example)

Budget stack on Retell, US client, standard voice, fast small LLM:

| Component | $/min |
|---|---|
| Retell voice engine (includes STT+TTS orchestration) | 0.070 |
| LLM (GPT-4o-mini-class) | 0.006 |
| Telephony (Twilio via platform) | 0.015 |
| **All-in per minute** | **≈ 0.091** |
| Premium swap: ElevenLabs voice + GPT-4o-class LLM | ≈ 0.16–0.20 |

A 250-min month ≈ **$23–25 infra**; a 750-min month ≈ **$68–75**. Hold those two numbers in your head — they are the whole reason the tiers in `PRICING.md` have minute caps.

## 6. What a configuration actually consists of (your ~2 weeks per pilot)

1. **Number & routing** — buy/assign number, set the client's forwarding rules (1–2 h).
2. **The prompt pack** — identity, greeting, services/menu/knowledge, booking rules, escalation rules, forbidden topics (the bulk of the craft; starts from our vertical template).
3. **Tools** — calendar availability + booking (Cal.com or Google Calendar), lead webhook to CRM/Sheet, transfer number, WhatsApp/SMS confirmation.
4. **The 20-call QA script** — we roleplay 20 scenarios and fix what breaks (`TEST_PLAN.md`).
5. **Go-live + watch week** — daily transcript review for week 1, then weekly (`TESTING.md`).

## 7. Glossary quick card

**ASR** = same as STT · **Voice-to-voice latency** = caller stops talking → AI audio starts · **Utterance** = one continuous piece of caller speech · **Webhook** = platform calls a URL we set when something happens (call ended, booking made) · **DTMF** = touch-tone digits (still needed for "press 1" fallbacks) · **CNAM/Caller ID** = the name/number shown to the caller · **PII** = personally identifiable information (names, numbers, health details — see `SECURITY.md`) · **Containment rate** = % of calls fully handled without a human.

## 8. Sources (all accessed 29–30 Aug 2026)

| Topic | Source |
|---|---|
| Vapi pricing ($0.05/min + components; $0.07–0.25 all-in) | cloudtalk.io/blog/vapi-ai-pricing/ · cekura.ai/blogs/vapi-ai-pricing · zeeg.me/en/blog/post/vapi-ai-pricing |
| Retell pricing ($0.07/min; $0.13–0.31 typical all-in; $2/mo numbers) | cloudtalk.io/retell-ai-pricing/ · cekura.ai/blogs/retell-ai-pricing-per-minute · layer3labs.io/guides/retell-ai-pricing |
| Bland pricing ($0.09 base, tiered $0.11–0.12, ancillary fees) | cloudtalk.io/blog/bland-ai-pricing/ · zeeg.me/en/blog/post/bland-ai-pricing |
| Synthflow usage pricing + $2k/mo white-label toolkit | autocalls.ai/article/synthflow-review · zeeg.me/en/blog/post/synthflow-ai-pricing · voiceaiwrapper.com/compare/synthflow-alternative |
| Managed-service prices (Smith.ai, Goodcall, Rosie, Slang.ai) | withallo.com/blog/ai-receptionist · agentzap.ai/blog/ai-receptionist-pricing-complete-cost-guide-2025 · bitebuddy.ai/blog/slang-ai-pricing · workflowstackai.com/blog/ai-receptionist-comparison-2026 |
| White-label/agency landscape (Vapi/Retell have no native white-label; wrappers exist) | trillet.ai/blogs/top-10-white-label-voice-ai-platforms-for-agencies-2026 · voiceaiwrapper.com |
| Telephony regulation (UAE blocks VoIP numbers; BD/KSA doc requirements) | twilio.com/en-us/guidelines/ae/regulatory · twilio.com/en-us/guidelines/bd/regulatory · twilio.com/en-us/guidelines/sa/regulatory · support.outreach.io (UAE VoIP blocking practice) |
| WhatsApp Business Calling API (inbound calls free; voice agents on WhatsApp) | twilio.com/en-us/blog/guide-to-whatsapp-business-calling-voice · respond.io/blog/whatsapp-ai-voice-agent · developers.facebook.com/documentation/business-messaging/whatsapp/calling |
| Calendar integrations (Vapi↔Google Calendar native tool; Cal.com API patterns) | docs.vapi.ai/tools/google-calendar · docs.vapi.ai/assistants/examples/appointment-scheduling · n8n.io/workflows/6895 (Cal.com) |
| Recording consent law (UAE all-party + criminal penalty; 12 two-party US states; BD) | recordinglaw.com/world-laws/world-recording-laws/uae-recording-laws/ · avoma.com/blog/call-recording-laws · frejun.com/call-recording-laws-uae/ |

Anything marked **[verify]** in these documents must be re-checked against the platform's own pricing page before a client quote — third-party pricing round-ups drift.