How Does an AI Receptionist Work? Architecture, Technology & Latency Explained (2026)

If you want the higher-level overview of what an AI receptionist is before diving into the architecture, that's the definitional companion to this page. Otherwise, here's the engineering view.

Most production AI receptionists in 2026 use what's called a “cascading architecture” — three specialised AI services running in sequence, with each handing off to the next:

Caller's voice  →  [STT: Speech-to-Text]
                   ↓
               [LLM: Large Language Model]
                   ↓
               [TTS: Text-to-Speech]  →  Caller hears response

This is the dominant architecture because each layer can be independently optimised, swapped, and improved. If a better speech-to-text model comes out, you switch only that layer. If you want a different voice, you change only the TTS provider. The trade-off is that the handoffs between layers add latency — typically 100-300 milliseconds total.

A newer alternative is unified speech-to-speech architecture — a single model that handles voice input to voice output without intermediate text steps. OpenAI's Realtime API, Google's Gemini Live, and xAI's Grok Voice ThinkFast use this approach. Total end-to-end latency targets sub-300ms, which is genuinely human-natural. The trade-off is less flexibility — you can't independently swap layers, and the unified model handles all decisions including tone, pacing, and intent recognition together.

Most Australian AI receptionist providers (including Aussie AI Agency, Sophiie AI, Johnni AI, AiDial, Smith.ai) currently use cascading architecture because:

• It's more mature and battle-tested in production
• Industry-specific compliance configuration is easier when the LLM layer is separately controllable
• Voice swapping (different accents, genders, languages) is easier
• Debugging is dramatically easier — you can inspect the transcript, the LLM's response, and the TTS output independently
• Function calling and API integration are more reliable

Unified speech-to-speech models will likely become the dominant architecture within 2-3 years as they mature, but for current production use, cascading remains the reliable choice for compliance-regulated industries.

A more complete view of the production architecture

In practice, AI receptionist systems include several additional layers beyond just STT/LLM/TTS:

• Telephony layer — SIP trunks, VoIP providers (commonly Twilio in Australia), call routing logic
• Voice activity detection (VAD) — distinguishes when the caller is speaking vs paused vs finished
• Turn-taking logic — handles interruptions, overlapping speech, and conversational pacing
• Context management — maintains the full transcript and business context across the call
• Function calling layer — translates LLM decisions into API calls to your business software
• Observability layer — logs every call, traces latency by stage, records quality metrics
• Escalation logic — confidence scoring that decides when to hand off to humans

When a vendor says “AI receptionist,” they're packaging all of these layers into a single product. The headline simplicity (“AI answers the phone”) hides genuine engineering complexity underneath.

Here's what happens in each step of a typical AI receptionist call, with realistic timing for each stage:

Step 1: Call routing — typically <100ms

Your business phone line is forwarded to a virtual phone number managed by the AI receptionist provider. Most Australian providers use Twilio for telephony (the dominant carrier-agnostic platform), though some use direct SIP trunk connections to Australian carriers like Telstra or Optus.

When a caller dials your usual business number, the call routes through SIP/VoIP to the AI's number. The caller doesn't see this redirect — they think they're calling you directly, because they are. Your number stays yours; only the call destination changes.

Step 2: Speech-to-text (STT) — typically 100-300ms per turn

The caller's audio is streamed in real time to a speech recognition service. Modern AI receptionists commonly use:

• OpenAI Whisper / Whisper Large v3 ↗ — high accuracy, supports 99+ languages
• Deepgram Nova-2 — purpose-built for real-time call audio, lowest latency
• Google Cloud Speech-to-Text — strong Australian English accent handling
• Azure Cognitive Services Speech — enterprise-grade, ISO compliance
• AssemblyAI — strong on accents and noisy environments

The best of these achieve word error rates around 4.9% on benchmark English audio (per NIST testing), with Australian accent accuracy typically 92-96% on clean phone audio. Background noise, strong accents, and conversational speech reduce accuracy — but well below the 2024 threshold where this was a regular problem.

Step 3: Large language model (LLM) — typically 300-500ms per turn

The transcribed text is fed to a large language model with a structured system prompt that defines:

• The business name, services, and opening hours
• The receptionist's persona (name, tone, pacing, escalation rules)
• The booking system and integrations available
• Industry-specific compliance rules (AHPRA framework for healthcare, AFSL/NCCP for finance, state-based conveyancing law, etc.)
• The full transcript of the conversation so far
• Available “tools” the LLM can call (book appointment, transfer call, send SMS, look up customer)

Common LLM choices for AI receptionists:

• OpenAI GPT-4 / GPT-4o — high quality, strong function calling
• Anthropic Claude (Sonnet / Opus) ↗ — strong instruction following, lower hallucination rates, preferred for compliance-sensitive applications
• Google Gemini Pro / Flash — competitive quality, Google ecosystem integration
• Meta Llama 3.1 / 3.3 (self-hosted) — open-source option for data sovereignty

The LLM generates a structured response — what to say next, what action to take (book, transfer, escalate), and any data to capture.

Step 4: Text-to-speech (TTS) — typically 200-400ms per turn

The LLM's text response is sent to a voice synthesis service that produces natural-sounding audio:

• ElevenLabs ↗ — industry-leading voice quality, supports custom Australian voices
• Cartesia Sonic — purpose-built for low-latency streaming, sub-200ms first audio
• Azure Neural TTS — enterprise-grade with multi-region Australian deployment
• PlayHT — competitive on voice variety
• OpenAI TTS — bundled with GPT-4o for unified workflows

The audio streams back to the caller as it's generated — the caller doesn't wait for the full response to be synthesised before hearing the first words.

Step 5: Workflow actions — runs in parallel with the conversation

While the conversation continues, the AI triggers actions in your business software through function calling. Example workflow during a single call:

• Caller mentions they want to book — AI calls check_availability(date_range) against your practice management API
• Available slots returned — AI offers them to caller
• Caller confirms — AI calls create_booking(patient_id, slot_id) and gets confirmation
• AI calls send_sms(phone, confirmation_message) in parallel
• AI calls log_lead(name, contact, reason) in your CRM

The two-stage commit pattern is critical here: the AI never says “you're booked” until the booking API returns success. If the API times out or errors, the AI says “I'll confirm by SMS in the next minute” and handles the actual confirmation asynchronously — never telling the caller something happened that didn't.

Step 6: Call summary — generated within 60 seconds of call ending

When the call ends, the AI generates:

• A written summary in plain English (“Mark Henderson called to book a check-up. Booked Tuesday 2:30pm. Existing patient, file updated.”)
• A full transcript of the call
• Structured data extraction (caller name, contact details, reason for call, action taken, urgency level)
• A confidence score on whether anything needs human review

This is sent to your team within 60 seconds — typically by email, SMS, push notification, or directly into your practice management system / CRM.

The entire 6-step flow happens for every call. Most callers complete a booking in 60-90 seconds of conversation, which means 4-8 round-trips of STT → LLM → TTS, each adding latency.

“Latency” in AI voice context refers to mouth-to-ear turn gap — the time from when the caller stops speaking to when the AI's response reaches the caller's ear. It's the single most important technical metric for AI receptionist quality, because it determines whether the call feels natural or robotic.

The human reference points

• Natural conversational pause: 200-500ms — feels normal
• Slightly slow but acceptable: 500-1,000ms — noticeable but tolerable
• Awkward delay: 1,000-1,500ms — caller notices, feels “robotic”
• Broken: >1,500ms — caller may hang up or talk over the AI

Research from voice AI specialists like Cresta ↗ and customer experience studies cited by Twilio ↗ and MindStudio ↗ converge on the same threshold: sub-500ms feels human; over 1,000ms degrades the experience rapidly.

The latency budget breakdown in cascading architecture

Most production AI receptionists in 2026 operate in the 600-1,100ms total latency range. Best-in-class systems (Aussie AI Agency, top-tier configurations of Sophiie AI, AiDial, premium tiers of Smith.ai) sit at the lower end. Budget tools and poorly-configured systems often exceed 1,500ms.

Why latency varies so much

• Streaming vs batch processing — streaming each stage (sending words as they arrive vs waiting for full sentences) cuts latency significantly
• Geographic proximity — Australian-hosted infrastructure reduces network round-trip vs offshore hosting
• Model size — smaller LLMs respond faster but with lower quality; finding the right size matters
• VAD tuning — overly cautious end-of-speech detection adds 200-400ms of waiting after the caller stops talking
• Function call overhead — when the LLM needs to call your booking API mid-conversation, the round-trip adds 100-500ms

For a calling experience that feels human, latency optimisation is genuinely the highest-leverage engineering work in building an AI receptionist. The difference between 600ms and 1,200ms latency is the difference between “I can't tell it's AI” and “obviously a robot.”

The AI receptionist isn't useful on its own — its value comes from completing actions in the software you already use. There are three main integration patterns:

Pattern 1: Native API integration

The provider has built a direct connection to specific software. Example: Aussie AI Agency's Cliniko integration uses Cliniko's REST API to check appointment availability, create bookings, and update patient records in real time during the call.

• Best for: Common Australian business software (Cliniko, Halaxy, Karbon, Xero, ServiceM8, LEAP)
• Reliability: Highest — purpose-built connections handle edge cases
• Coverage: Limited to software the provider has built for

Pattern 2: Webhook + middleware (Zapier, Make, n8n)

The AI receptionist sends structured data via webhook to an automation platform, which routes it into your software.

• Best for: Less-common software or custom workflows
• Reliability: Medium — depends on the middleware layer
• Coverage: Universal — Zapier supports 6,000+ apps

Pattern 3: Custom API integration

For enterprise or niche software, the provider builds a custom connector.

• Best for: Unusual proprietary systems
• Reliability: High — if built well
• Cost: Often a one-off integration fee ($500-$3,000)

The data flow during a typical call

For a medical clinic booking a new patient:

1. Call arrives → AI greets caller
2. Caller provides name, phone, reason for visit
3. AI calls lookup_patient(phone) → existing patient? No
4. AI captures structured intake data during natural conversation
5. AI calls check_availability(practitioner, date_range) → returns slots
6. AI offers slots, caller picks one
7. AI calls create_patient(name, phone, dob, reason) → patient ID returned
8. AI calls create_booking(patient_id, slot_id, type) → booking confirmed
9. AI calls send_sms(phone, confirmation_template, slot_details) → SMS sent
10. AI confirms verbally to caller
11. AI calls log_call(transcript, summary, structured_data) → archived

That's 7 API calls in a 90-second conversation, all happening in parallel with the natural-sounding voice interaction.

The integration quality directly determines the AI receptionist's value. A provider with excellent voice quality but weak Cliniko integration is less useful to a medical clinic than a provider with decent voice and deep, reliable Cliniko integration. When evaluating providers, the integration depth question matters as much as the voice quality question.

Knowing when to stop and hand off is one of the hardest engineering problems in AI receptionist design. A poorly designed system tries to answer everything and hallucinates; a well-designed system explicitly escalates when:

Trigger 1: Confidence scoring below threshold

Modern LLMs can produce a confidence score on their own response quality. When confidence drops below a configured threshold (e.g., 70%), the system flags the response for human review rather than committing to an answer.

Trigger 2: Keyword/intent-based escalation

Specific phrases trigger automatic escalation regardless of LLM confidence:

• “Emergency” / “urgent” / “ambulance” / “000” — immediate human handoff
• “Complaint” / “lawyer” / “lawsuit” — escalation to senior team
• Healthcare-specific: “chest pain” / “bleeding” / “overdose” — immediate clinical escalation, 000 directive
• Pharmacy-specific: S4/S8 medication names — escalation to pharmacist
• Financial: “fraud” / “scam” / “stolen” — escalation to compliance

Trigger 3: Caller request

When a caller asks for a human (“Can I speak to a real person?”), the AI immediately offers transfer or callback options without trying to convince the caller to continue with AI.

Trigger 4: Out-of-scope topic detection

For compliance-regulated industries, the LLM is configured to recognise when callers ask for advice the AI is not authorised to provide:

• Medical: clinical diagnosis questions → escalation to practitioner (AHPRA-regulated healthcare)
• Financial: product recommendations → escalation to AFSL-credentialed adviser (AFSL-regulated financial planning) or NCCP-licensed broker (NCCP-regulated mortgage broking)
• Legal: specific legal opinion → escalation to solicitor (state-based conveyancing)
• Pharmacy: dosing questions → escalation to pharmacist (S4/S8 medication escalation)

Trigger 5: Sentiment detection

When the caller's tone indicates frustration, distress, or anger, the system can route to humans even if the call would otherwise be routine. Aussie AI Agency's healthcare overrides specifically include this trigger — bereavement-tone or panic-tone callers always reach a human.

The compliance distinction matters most in Australian regulated industries. A generic global AI receptionist may handle escalation poorly because it wasn't designed for AHPRA, AFSL, NCCP, or state-licensed legal scope. Industry-specific AI receptionists bake the escalation rules into the system design from day one. Geographic considerations also apply — Sydney businesses and Melbourne businesses face state-specific compliance variations.

This guide is published by Aussie AI Agency — we're a Sydney-based AI receptionist provider, so we have a commercial interest in helping you understand the category.

Where the technical details come from

The technology citations (OpenAI Whisper, Anthropic Claude, ElevenLabs, Deepgram, Cartesia, Twilio, Azure Speech) are all real production services that AI receptionist providers use. Specific latency figures and word error rates are sourced from public vendor documentation, NIST benchmarks, and engineering blogs from Twilio, AssemblyAI, Cresta, and MindStudio (cited where used).

Where we draw the line on technical disclosure

We don't disclose Aussie AI Agency's specific stack choices in detail — that's competitive engineering, not category education. What we do share: cascading architecture is standard, our latency targets sub-800ms, our integrations are native rather than middleware-only, our data is hosted in AWS Sydney region.

A caveat on technology change rates

Voice AI technology moves fast. Models cited today (GPT-4o, Claude Sonnet, ElevenLabs Turbo, Cartesia Sonic) may be superseded in 6-12 months. This page is reviewed quarterly to keep citations current.

If you spot a technical inaccuracy, email niel@aussieaiagency.com.au.