> ## Documentation Index > Fetch the complete documentation index at: https://www.vobiz.ai/docs/llms.txt > Use this file to discover all available pages before exploring further. # What Is an AI Phone Call? How AI Calling Works > An AI phone call is a real telephone call where the conversation is handled by an AI voice agent — not a human. How AI phone calls work, the technology stack (ASR, LLM, TTS), use cases, and what distinguishes them from robocalls and IVR systems. *November 2026 · Vobiz Team* **An AI phone call is a real telephone call where the conversation is handled by an AI voice agent — not a pre-recorded message and not a human.** The AI listens to what the caller says, understands it in context, and responds in natural-sounding speech, in real time. Unlike a robocall playing a fixed script, an AI phone call is dynamic: the AI can follow the conversation wherever it goes, answer questions, and take actions like booking appointments or looking up account details. > [!NOTE] > > **Key takeaways** > > * AI phone calls use ASR (speech recognition) + LLM (reasoning) + TTS (speech synthesis) over real telephony infrastructure > * They are interactive and adaptive — unlike robocalls that play fixed recordings > * Both inbound (AI answers calls) and outbound (AI initiates calls) are common business deployments > * In the US, AI phone calls require the same TCPA written consent as pre-recorded robocalls (2024 FCC ruling) > * Latency — the gap between when a caller stops speaking and when the AI responds — is the primary quality variable ## How an AI phone call works A single AI phone call involves several technology layers working together in under a second: ``` Caller speaks ↓ SIP trunk (telephony carrier) → Vobiz media server ↓ Voice Activity Detection (VAD) detects end of speech ↓ ASR — speech-to-text transcription ↓ LLM — understands intent, generates response text, calls tools if needed ↓ TTS — synthesises natural-sounding speech from response text ↓ Vobiz media server → SIP trunk → caller hears response ``` **Media server** (the often overlooked layer): Before audio reaches ASR, it passes through a media server that handles noise cancellation, voice isolation, codec conversion (G.722/Opus for HD audio), and Packet Loss Concealment. Poor audio quality degrades ASR accuracy — a 10% increase in word error rate translates directly to AI misunderstandings and poor call quality. ## AI phone call vs robocall vs IVR | System | Interactive? | Dynamic response? | Natural language? | Tool access? | | - | - | - | - | - | | **Robocall** | No | No | No | No | | **IVR (press 1 for...)** | Limited | No | Keyword only | No | | **AI phone call** | Yes | Yes | Full natural language | Yes | A robocall plays audio and collects at most a keypress. An IVR routes based on button presses or simple voice commands. An AI phone call conducts a real conversation — the caller can say anything, and the AI responds appropriately. ## Inbound AI phone calls An AI agent answering inbound calls can: * **Replace IVR menus** — callers speak naturally instead of pressing buttons * **Handle complete requests** — look up account info, process a payment, reschedule an appointment, without transferring to a human * **Triage complex issues** — gather context, determine urgency, then route to the right human with a full call summary * **Handle after-hours calls** — the AI is available 24/7; calls outside business hours are handled or logged with follow-up promises **Example**: A patient calls a dental clinic at 8pm. The AI answers, confirms their identity, checks the appointment system, and offers two slots for a cancellation opening tomorrow. The patient confirms. The AI books the appointment and sends an SMS confirmation — zero human involvement. ## Outbound AI phone calls An AI agent initiating outbound calls can: * **Payment reminders** — call customers with overdue accounts, present the amount due, process payment over the call * **Appointment reminders** — call patients or customers to confirm, reschedule if needed * **Lead qualification** — call inbound leads within seconds of form submission, ask qualifying questions, update CRM * **Post-interaction surveys** — call after a service interaction, collect NPS or CSAT verbally * **Delivery notifications** — proactive calls for order status updates or delivery exceptions **Scale**: An AI agent can run hundreds of simultaneous outbound calls from a single campaign. A human agent can handle one call at a time. ## What determines call quality **Latency**: The delay between when a caller stops speaking and when the AI responds. A well-optimised AI calling stack achieves 800ms–1.2 seconds end-to-end. Poor infrastructure can push this to 2–3 seconds — perceptible and disruptive to conversation flow. **Audio quality**: HD Voice (G.722/Opus codec) vs standard telephony (G.711 narrowband). AI ASR models perform significantly better on wideband audio. Noise cancellation and voice isolation at the media server reduce word error rates in noisy environments. **ASR accuracy**: Word error rate (WER) in the target language and accent. Models trained on the dominant accent (US English, UK English, Indian English) vary significantly. Proper ASR selection for the caller population is critical. **LLM quality and latency**: The model's ability to understand context, handle interruptions, and use tools correctly — combined with low first-token latency so the AI doesn't pause unnaturally before responding. **Voice naturalness**: TTS voice quality — prosody, pacing, handling of numbers and proper nouns. Callers notice synthetic voices, but modern TTS (ElevenLabs, Cartesia, Deepgram Aura) has narrowed the gap considerably. ## Caller identification and trust A key challenge for outbound AI phone calls is answer rate. Carriers use STIR/SHAKEN to authenticate caller ID, and analytics platforms (Hiya, First Orion, TNS) label unknown or suspicious numbers as "Potential Spam" on the recipient's screen. Best practices: * Use numbers with **A-level STIR/SHAKEN attestation** — Vobiz numbers on the Vobiz network carry A-level attestation * Maintain consistent caller ID — do not rotate numbers across campaigns * Register numbers with Hiya, First Orion, and TNS for branded caller ID display * Use local numbers in the recipient's area code where possible (local presence) ## Compliance for AI phone calls AI phone calls are regulated as automated calls under TCPA (US), TRAI TCCCPR (India), and equivalent frameworks in other markets. **United States (TCPA)**: * The FCC confirmed in February 2024 that AI-generated voice calls = pre-recorded robocalls for consent purposes * Prior express **written** consent required before any outbound AI call to a US consumer mobile or residential number * DNC registry scrubbing required before every campaign * Calling hours: 8am–9pm recipient local time **India (TRAI TCCCPR)**: * 140-series caller ID mandatory for all commercial outbound calls (including AI) * TCCCPR Third Amendment (Sep 2026): declaration of AI-initiated calls required * 9am–9pm promotional call window * NDNC registry scrubbing required See [What is TCPA?](https://www.vobiz.ai/docs/blogs/what-is-tcpa.md) and [India Calling Regulations](https://www.vobiz.ai/docs/compliance/india/calling-regulations.md) for full detail. ## Platforms for building AI phone call systems **AI frameworks** (handle the ASR/LLM/TTS pipeline): * Retell AI, Vapi, ElevenLabs Conversational AI, LiveKit, Pipecat, OpenAI Realtime API **Telephony layer** (handle SIP trunking, PSTN numbers, media processing): * Vobiz, Twilio, Plivo, Telnyx Vobiz provides the telephony and media infrastructure — SIP trunking, number provisioning, bidirectional WebSocket audio streaming, AMD (answering machine detection), and an AI-optimised media server — that AI frameworks connect to for real PSTN call delivery. ## AMD — answering machine detection For outbound AI calls, **AMD (Answering Machine Detection)** determines whether the call was answered by a human or a voicemail system before the AI starts speaking. Without AMD, the AI may deliver its opening line to a voicemail inbox — wasting the call and potentially leaving an awkward recording. Vobiz provides per-call AMD via `machine_detection=Enable`. When AMD detects a human, the AI is connected. When a machine is detected, the call can be dropped or the AI can leave a structured voicemail. ## Related * [What is AI calling?](https://www.vobiz.ai/docs/blogs/what-is-ai-calling.md) — AI-powered calling overview * [What is a voice agent?](https://www.vobiz.ai/docs/blogs/what-is-a-voice-agent.md) — AI voice agent fundamentals * [What is agentic AI?](https://www.vobiz.ai/docs/blogs/what-is-agentic-ai.md) — how AI agents plan and act autonomously * [What is AI outbound calling?](https://www.vobiz.ai/docs/blogs/what-is-ai-outbound-calling.md) — outbound AI call campaigns * [What is phone call automation?](https://www.vobiz.ai/docs/blogs/what-is-phone-call-automation.md) — automating calls at scale * [What is answering machine detection?](https://www.vobiz.ai/docs/blogs/what-is-answering-machine-detection.md) — AMD for AI calling * [US calling regulations](https://www.vobiz.ai/docs/compliance/us/calling-regulations.md) — TCPA consent for AI calls * [India calling regulations](https://www.vobiz.ai/docs/compliance/india/calling-regulations.md) — TRAI rules for AI calls This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.