OpenAI Realtime API
Speech-to-speechSub-600ms TTFT · barge-in · in-app voice copilots
ai voice agent development · live AI voice agents we ship to production.
AI voice agents we ship to production.
Realtime, telephony, mobile — picked per workload.
AI voice agent development company shipping production AI voice agents on three stacks: realtime voice ai on the OpenAI Realtime API (speech-to-speech), chained STT+LLM+TTS, and Deepgram Voice Agent. Sub-600ms first-token latency, $0.005–$0.06/min honest cost range, eval-gated against your real call recordings. Telephony (Twilio · Telnyx · LiveKit · Daily) and mobile-native voice inside Flutter / iOS / Android / web. First pilot live in 5–7 weeks, behind a feature flag, with a walk-away point if the metric won't move.
Definition
What is a sub-second AI voice agent?
A sub-second AI voice agent is a real-time conversational system that completes a full speech-in to speech-out round-trip in roughly 600 milliseconds, the latency budget at which a phone call feels natural rather than robotic. Built on streaming speech-to-speech models like OpenAI Realtime (gpt-realtime-2) or Gemini Live, it pairs server-side voice activity detection for interruption handling with function-calling against business systems mid-turn. Unlike interactive voice response (IVR) trees, a voice agent reasons over open-ended speech rather than matching a fixed phone-tree branch. Unlike traditional Whisper-plus-TTS chains, the streaming model handles speech end-to-end without per-leg latency stacking. Production deployments use Twilio for telephony, a hybrid retrieval index (pgvector + BM25) for grounded answers, and Langfuse for per-call eval logging.
<600ms
first-token latency on shipped Realtime pilots
3 stacks
Realtime · chained · unified, picked per workload
$0.005–$0.06
honest per-minute cost range we'll quote you
Fixed-fee
audit-to-roadmap before any voice agent build
voice ai agent · what we ship Six AI voice agents we ship
Six AI voice agents we ship
we've shipped to production.
Every voice ai agent pattern below has been shipped from this exact playbook: telephony inbound and outbound, in-app voice copilots, IVR replacement, multilingual support, voice-first ecommerce. Each one comes with an eval suite, audit logging, barge-in tuning, and a per-minute cost target — not a demo reel.
Inbound telephony voice agents — call deflection
The crown-jewel use case for voice ai for customer service. A real phone number (Twilio, Telnyx, or your existing SIP trunk) answered by a voice ai agent that handles tier-1 calls: appointment changes, order status, billing balance, store hours, password resets. Confidence gate at 0.7; anything below escalates to a human queue with a structured handoff summary, not a re-asked greeting. We ship these with barge-in, hold detection, and DTMF fallback for callers who hit a frustration ceiling.
Outbound calling agents — reminders, surveys, qualification
Outbound voice agents for appointment reminders, payment-due nudges, NPS surveys, and lead qualification. Per-call cost lands at $0.05–$0.15 (vs $4–$8 for a human callback) and you control the cadence + script + escalation rules. Compliance baked in: TCPA-aware time windows, opt-out detection on every turn, full call recording with consent prompts where required. Built on a chained stack to keep cost low at outbound scale.
In-app voice copilots — voice inside your mobile / web product
Voice agents that live inside your iOS / Android / Flutter / web app, not a phone call. Conversational ai voice for hands-free interfaces in food delivery, automotive, field service, and accessibility surfaces. Mobile-native delivery is a differentiator: we wrote the UI kit a lot of Flutter screens run on, so the voice surface ships with proper interrupt handling and mic-permission UX, not a half-baked WebRTC bolt-on.
IVR replacement — kill the press-1-for-sales tree
Direct replacement of legacy IVR menus with a single open-ended voice ai agent that routes by intent instead of by keypad. Average call shortens 40–90 seconds when callers can just say what they want, and the same agent can answer the question the caller would have been routed to anyway. Plays cleanly inside existing contact-center platforms (Genesys, Five9, Amazon Connect) via SIP REFER or media-streaming.
Multilingual support voice agents
Voice agents that operate across English, Spanish, Hindi, Portuguese, Arabic, Mandarin, French, and German out of the box. Chained stacks shine here: Deepgram or Whisper for STT in the caller's language, GPT-5 or Sonnet 4.6 for the reasoning step (both genuinely multilingual without separate models), and ElevenLabs or Cartesia for TTS that doesn't sound like a 2018 phone tree. Language detection on the first turn, no menu required.
Voice-first ecommerce + concierge
Voice agents for hands-busy commerce — kitchen-side reordering, drive-thru, in-car, accessibility-first browsing. Grounded in your Shopify / WooCommerce / commercetools catalog with function calls into order-management, payment-tokenization, and 3PL APIs. Realtime API is the right choice here when TTFT under 600ms is non-negotiable for the natural conversational feel customers expect.
ai voice agent development · the architecture decision Three voice architectures,
Three voice architectures,
honest tradeoffs on latency, cost, control.
The single biggest decision in voice ai development is which architecture you pick. Most listicles dodge it. OpenAI Realtime API, chained STT+LLM+TTS, and unified-vendor stacks each win different workloads. Per-dimension honest comparison below; the audit picks for your workload.
Dimension
You're here
OpenAI Realtime API Speech-to-speech · single model
Chained STT+LLM+TTS Composable · model-agnostic
Deepgram Voice Agent Unified vendor · bundled SLA
First-token latency (TTFT) How fast the agent starts speaking back. Conversational feel breaks above 800ms.
OpenAI Realtime API <600ms typical · best-in-class · speech-to-speech avoids two model hops
Chained STT+LLM+TTS 800ms–1.2s · improving with streaming Whisper + GPT-5-mini + ElevenLabs Flash
Deepgram Voice Agent ~700ms · vendor-tuned end-to-end · narrower than chained
Cost per minute in production Real numbers from shipped pilots at meaningful volume.
OpenAI Realtime API ~$0.06/min · paying for premium latency · adds up fast on outbound
Chained STT+LLM+TTS $0.005–$0.02/min · cheapest by 3–10× · routes to mini models per turn
Deepgram Voice Agent ~$0.04/min · bundled but locks you to one vendor's pricing curve
Model control + swap Can you change the reasoning model without re-architecting?
OpenAI Realtime API Reasoning is OpenAI-locked · you swap voices but not the brain
Chained STT+LLM+TTS Swap Claude ⇄ GPT ⇄ Llama per workload · best for ai voice agent development
Deepgram Voice Agent Vendor-locked stack · you trade flexibility for fast time-to-pilot
Barge-in + interruption handling Can the caller cut the agent off mid-sentence without breaking the turn?
OpenAI Realtime API Native · the API was designed for it · <300ms barge-in detection
Chained STT+LLM+TTS Doable · requires VAD + endpointing + cancellation plumbing on every layer
Deepgram Voice Agent Native · Deepgram's agent endpoint handles it · less control over thresholds
Multilingual coverage Production support beyond English.
OpenAI Realtime API Strong on top-10 languages · weaker on Indic + African languages today
Chained STT+LLM+TTS Best-in-class · Whisper + Sonnet 4.6 + ElevenLabs covers 30+ languages
Deepgram Voice Agent Solid English + EU · narrower than chained · improving quickly
Telephony fit How well it plugs into Twilio · Telnyx · LiveKit · Genesys.
OpenAI Realtime API Works via media-streaming bridge · adds ~50ms vs in-app · stable in production
Chained STT+LLM+TTS Native fit · every telephony provider has STT/TTS hooks for chained pipelines
Deepgram Voice Agent Twilio + LiveKit have first-party integrations · fastest telephony onboarding
Numbers are typical production traces from shipped pilots. Per-workload picks vary with your eval data and call volume.
voice cost economics How we cut a voice agent bill
How we cut a voice agent bill
from $0.18/min to $0.05/min.
Five tactics stacked, ordered by impact on chained voice stacks. Most voice pilots see per-minute cost drop 60–80% at the same eval-suite quality after this pass. Realtime API has a tighter optimization surface (~$0.06/min floor); chained is where the real cost engineering lives. This optimization pass is included in every voice agent pilot, post-cutover.
Naive chained baseline 0.18 $/min naive cost
What we ship 0.05 $/min same eval-suite quality
telephony · pick a stack Four telephony stacks,
Four telephony stacks,
one voice agent — pick yours.
The voice agent is half the story; the telephony layer underneath is the other half. Twilio for ubiquity (twilio voice ai via Media Streams is the most common production path we ship), Telnyx for margin on high-volume outbound, LiveKit for in-app voice copilots that share a codebase across mobile + web + phone, Daily for the most opinionated open-source stack via Pipecat. Pick a platform to see the integration code, auth model, and timeline.
Auth API key + auth token · per-account SID
Pattern Media Streams · WebSocket
Typical time-to-live 2–4 weeks from greenfield to pilot
from twilio.rest import Client
from twilio.twiml.voice_response import VoiceResponse, Connect
# 1. Incoming call → TwiML returns <Connect><Stream/> to our WS endpoint
def incoming_call(request):
response = VoiceResponse()
connect = Connect()
connect.stream(url="wss://voice.api.example.com/media")
response.append(connect)
return str(response)
# 2. Media WS receives 8kHz µ-law audio frames every 20ms
# 3. We re-encode to PCM, send into OpenAI Realtime or Whisper
# 4. TTS audio (ElevenLabs Flash) streams back to Twilio media WS
# 5. Barge-in cancels in-flight TTS within ~280ms via VAD signal
Real integration snippet — the names changed. Twilio · lives in your repo.
voice ai development · vendor stack The three layers behind a voice agent,
The three layers behind a voice agent,
picked per stage not per vendor.
A chained voice ai agent is three vendors stacked — STT in, LLM in the middle, TTS out. Each layer has 2–3 production-grade choices in 2026 with meaningful tradeoffs on cost, latency, and language coverage. Here's the default stack we ship; we re-pick per workload when the eval data demands it.
STT tier — speech in
Deepgram Nova-3 is our default for telephony (8kHz µ-law, streaming, sub-200ms partial transcripts, multilingual). Whisper-large-v3 still wins on rare-language coverage and noisy environments. We pick per workload. OpenAI Realtime API rolls its own STT internally so this layer disappears entirely when you go speech-to-speech. We benchmark on your real call audio before recommending; vendor benchmarks lie about phone-quality audio constantly.
LLM tier — the reasoning step
GPT-5 / Sonnet 4.6 for grounded multi-turn reasoning, function calling, and tool use (the part that turns a voice ai agent into something that does work, not just talks). GPT-5-mini or Haiku 4.5 for cheap classify + routing layers in front. Realtime API ties you to OpenAI's reasoning model — fine when latency dominates the requirement, less fine when your eval data says Claude wins for your domain. Model-agnostic by default; locked-in only when you ask for it.
TTS tier — speech out
ElevenLabs Flash v2.5 for sub-200ms TTFT and natural prosody (the default for in-app voice copilots and premium telephony). Cartesia Sonic for the cheapest-per-character production-grade voice (winning on long outbound calls where per-minute cost dominates). OpenAI TTS (gpt-4o-mini-tts) when you want a single-vendor stack. The choice is a $/min + brand-voice decision, not a quality one. All three sound human in 2026.
eval suite · 4 metrics Four metrics every voice pilot is gated on,
Four metrics every voice pilot is gated on,
before any production rollout.
The disqualifying signal when shortlisting any voice ai company is asking how they measure quality and getting marketing-speak back. Here are the four numbers we publish on every voice agent pilot, measured on a held-out eval set from your real call recordings. Below the targets, we don't roll out. We tune.
Intent-match accuracy
On a held-out eval set of 200+ real call snippets from your domain, what fraction does the agent classify and respond to correctly? Most production voice ai agents we see land at 84–93% on tier-1 intents after tuning. Below 80% on a frozen eval set means the prompt or retrieval needs work before rollout, not after.
Hallucination rate per 100 calls
How often does the agent invent a policy, a balance, a hours-of-operation, or a product feature that doesn't exist? Measured by sampled human review (200 calls/week minimum during pilot). Target is <2 per 100 calls before production rollout. Anything above 5 means your RAG grounding is missing the source documents the agent needs.
Barge-in latency (p50, p95)
When the caller interrupts the agent mid-sentence, how fast does the agent stop talking? Target: <300ms p50, <500ms p95. Above 500ms and the conversation feels robotic; callers start over-talking instead of waiting. This is the metric that decides whether a voice agent feels alive or feels like 2015 IVR.
Mean-time-to-action (MTTA)
From call start to the moment the agent does the thing the caller wanted (booked the appointment, looked up the order, queued the refund). Target depends on the workflow, but we publish the baseline you'll be measured against during the audit. Most replacements of human tier-1 calls cut MTTA by 40–70% just by removing the queue wait.
engagement models Three ways to start.
Three ways to start.
Audit, pilot, or continuous.
Same pricing as our other engagements. Most voice ai company shortlists hide pricing. We publish it. Audit first to scope architecture + telephony, run a 5–7 week pilot on the highest-ROI workflow, then continuous if you want to ship the next 2–3.
1–2 weeks
Voice agent audit
Find the voice workflow worth shipping, in the right architecture, before any build commitment.
Fixed-fee fixed
- Use-case shortlist (inbound · outbound · in-app · IVR replacement)
- Architecture recommendation (Realtime · chained · unified vendor)
- Telephony stack pick (Twilio · Telnyx · LiveKit · Daily)
- Per-minute cost projection in $/min terms, not vague ranges
- Eval-set design — 50–200 call snippets from your real domain
- 90-day voice-agent roadmap with named pilots + walk-away criteria
5–7 weeks
Voice agent pilot
One voice ai agent shipped end-to-end on your chosen architecture and telephony stack, with eval data, not a demo reel.
Fixed-bid fixed price
- Architecture build — Realtime · chained · or unified per audit decision
- Telephony integration — your numbers, your trunks, your IVR fallback
- RAG grounding over your real knowledge base (policies, catalog, tickets)
- Eval pass on the 4 metrics — intent-match · hallucination · barge-in · MTTA
- Pilot rollout 10% → 50% → 100% gated by metric movement
- Walk-away point — if the metric won't move, no phase 2
Monthly
Continuous voice team
Embedded squad shipping new voice workflows + tuning the live ones.
monthly per month
- Voice engineer + PM + ops analyst, embedded
- Monthly eval drift + barge-in + MTTA report
- Per-call cost-of-ownership tracking and optimization passes
- New workflow rollouts on cadence (outbound, multilingual, mobile)
- Cancel any month — no annual contract
Your numbers, your data Realtime · chained · unified — your pick Eval-gated rollouts Walk-away point in every pilot
voice agent build playbook How we ship a production voice agent
How we ship a production voice agent
in 5–7 weeks, evaled + flagged.
Four stages, milestone-billed, with a walk-away point at the vendor-pick decision. Most voice agent failures happen because the team picked the architecture by ideology, not eval. Both are in week 1 and week 2 here, not after the pilot bill arrives.
- Week 1
Eval set + scope
Harvest 50–200 real call snippets from your call recordings (or run a structured intake if you're greenfield). Build the eval set the voice ai agent will be measured against — intent-match accuracy on each call type, expected MTTA per workflow. Scope locked: which call types in, which out, who escalates to whom.
Eval set + scope doc + architecture recommendation - Week 2
Telephony + STT/TTS pick
Wire the chosen telephony stack (Twilio · Telnyx · LiveKit · Daily) into a sandbox number. Run STT options against your real call audio (Whisper vs Deepgram vs Realtime) and pick per WER on YOUR audio, not vendor benchmark audio. TTS A/B for brand voice. RAG corpus ingested if grounding is in scope.
Sandbox call live · vendor picks lockedWalk-away point - Weeks 3–4
Build + barge-in tuning
Wire the full anatomy: telephony → STT → LLM (with tool use + RAG) → TTS → guardrails → log. Tune the hard parts: barge-in thresholds (<300ms p50), endpointing (when has the caller stopped talking?), hold-music handling, DTMF fallback, escalation handoff with structured summary. Behind a feature flag on a sub-100-call cohort.
Voice agent live on shadow cohort - Weeks 5–7
Eval pass + rollout
Run the 4-metric eval against shadow traffic for 2 weeks. Hallucination review by human sample. Roll out 10% → 50% → 100% gated by eval movement. Token-optimization pass post-cutover: route to cheaper models per turn, cache the system prompt, idle-trim. Most voice pilots see per-minute cost drop 60–80% at the same eval-suite quality.
Full rollout + monthly cost target locked
capability patterns Voice agents we've shipped.
Voice agents we've shipped.
Three architectures, three workloads.
Three anonymized capability patterns drawn from real voice ai agent engagements — one on the Realtime API (mobile-native), one chained (outbound telephony), one chained-multilingual (IVR replacement). Named references shared under NDA once we know what you're building.
gpt-realtime-2 tier-1 voice agent at $0.10/call
Problem
Tier-1 voice queue averaging 4-minute wait at peak. Five inbound questions accounted for 62% of call volume. Existing IVR bouncing 80%+ to a human. The binding constraint was sub-700ms first-token; anything slower and US callers report a robot.
Approach
gpt-realtime-2 speech-to-speech over a help-center pgvector RAG, Twilio + Cloudflare Workers for sub-60ms ingress, handoff_to_human as a function-calling tool with confidence gating at 0.7. Whisper fallback for accent + noise. Kill point caught + fixed at week 5 (multilingual cache invalidation).
gpt-realtime-2Whisper-large-v3pgvector 0.7Twilio VoiceCloudflare Workers
Outcome
Read the full case study
≈ 38% tier-1 deflection (95% CI 33%–43% · n=11,400 calls) · published $0.10/call (2026-Q1)
Realtime voice copilot inside a Flutter food-delivery app
Problem
Customers re-ordering favourites mid-commute or hands-busy in the kitchen had a 7-tap nested-menu reorder flow. Drop-off at step 3 was 38%. Voice was the obvious primitive, but every hosted voice SDK added 200–400ms of round-trip that broke the conversational feel.
Approach
OpenAI Realtime API integrated directly into the Flutter shell via a thin Dart-to-WebRTC bridge. Function calls into the existing order-management API (no rewrite required). Sub-600ms TTFT measured end-to-end on the user's actual device, not on a wired desktop. Barge-in handled natively. Mic-permission UX and on-screen visual feedback shipped in the same release.
OpenAI Realtime APIFlutter / DartWebRTC bridgeOrder-management APIRAG over catalog
Outcome
Read the full case study
+18% order-completion lift on voice cohort vs tap cohort (2026-Q1)
Appointment-reminder + reschedule voice agent
Problem
Multi-location clinic running 4,000 outbound reminder calls weekly through a human contact-center, each call $4–6 fully-loaded. Reschedule rate captured at the call was <30% because reps couldn't see the calendar live.
Approach
Chained stack (Telnyx + Deepgram Nova-3 + GPT-5-mini + ElevenLabs Flash) for outbound cost economics. Function calls into the EHR's scheduling API to offer real open slots live during the call. TCPA-aware calling windows, opt-out detection per turn, audit logging of every consent prompt. Escalation to a human queue when the caller asked anything outside reschedule scope.
TelnyxDeepgram Nova-3GPT-5-miniElevenLabs FlashEHR scheduling API
Outcome
~$0.10–$0.15 per-call cost on this chained stack (published math · vs $4–6 live-agent baseline)
Multilingual support IVR replaced with a single voice ai agent
Problem
Legacy IVR for a multi-country retail brand routed calls across 6 languages with 4-level keypad menus. Average pre-agent wait was 90 seconds; 22% of callers abandoned before reaching a human; non-English callers fared worst because the menu was English-first.
Approach
Single open-ended voice agent answers in the caller's language (detected on first turn) and routes by intent, not by keypad. Chained stack: Whisper-large-v3 for STT (best rare-language coverage), Sonnet 4.6 for multilingual reasoning, ElevenLabs Multilingual v2 for TTS. Integrated into the existing Genesys contact center via SIP REFER so the human-handoff queues didn't change.
Whisper-large-v3Sonnet 4.6ElevenLabs Multilingual v2Genesys SIP REFERPinecone RAG
Outcome
40–70s pre-agent wait reduction target on this pattern (range, not a single shipped figure)
frequently asked Questions voice agent buyers ask most.
Questions voice agent buyers ask most.
Real answers, no hedging.
What does an AI voice agent development company actually deliver?
Five things, in this order. (1) An audit (1–2 weeks, fixed-fee) that picks the voice workflow worth shipping, the architecture (Realtime · chained · unified vendor), and the telephony stack (Twilio · Telnyx · LiveKit · Daily). (2) A pilot (5–7 weeks, fixed-bid fixed) that ships one voice ai agent end-to-end on your chosen stack with the full anatomy wired (telephony → STT → LLM → TTS → guardrails → log) and gated by eval data, not a demo. (3) The 4-metric eval pass (intent-match accuracy, hallucination per 100 calls, barge-in latency, mean-time-to-action) run on a held-out set from your real call recordings. (4) Pilot rollout 10% → 50% → 100% gated by metric movement, with a walk-away point. (5) Continuous tier monthlynth if you want to ship the next workflow or tune the live one. We're a voice ai company that publishes pricing and architecture choices openly. Most listicle SERPs hide both.
Realtime API vs chained voice agents — which architecture wins?
Neither wins universally. Here's the honest decision. OpenAI Realtime API wins when first-token latency is the dominant requirement (in-app voice copilots, premium telephony, anything where conversational feel breaks above 800ms) and you can pay ~$0.06/min in production. Chained STT+LLM+TTS wins on cost (10–30× cheaper at scale), model flexibility (swap Claude ⇄ GPT per turn), and multilingual coverage — the right pick for high-volume outbound, multilingual support, and any workflow where per-minute economics dominate. Deepgram Voice Agent (unified vendor) wins when you want the fastest time-to-pilot with a single SLA and don't need to swap models. Our audit picks per workload, not per ideology. The openai realtime api development path is the right answer about a third of the time in our shipped pilots.
What does a voice agent cost per minute in production?
Honest numbers from shipped pilots. OpenAI Realtime API: ~$0.06/min all-in (model + audio in/out + telephony). Chained Whisper + GPT-5 + ElevenLabs Multilingual: ~$0.18/min naively, down to $0.05/min after the standard optimization pass (route to mini models per turn, cache the system prompt, idle-trim silence, swap Whisper for Deepgram Nova-3). Chained Deepgram + GPT-5-mini + Cartesia: ~$0.005–$0.02/min, the cheapest stack we ship, used on high-volume outbound. Deepgram Voice Agent: ~$0.04/min bundled. Per-minute cost is published in every audit deliverable in $/min terms, not vague "depends on usage" ranges. The waterfall above shows the optimization pass we run on every pilot.
Can voice agents do agentic workflows — tool use, multi-step actions, retrieval?
Yes. That's the modern definition. A voice ai agent that just answers FAQs is a 2018 IVR with better TTS. Agentic ai development for voice means the agent can call your APIs (refund a charge, book an appointment, look up an order), retrieve from your RAG corpus per turn (catalog, policy docs, ticket history), make multi-step decisions (verify identity → check policy → execute action), and hand off to a human with a structured summary when policy says it should. We build voice agents on the same agentic-workflows stack as our AI agent development work. The only difference is the interface is voice, not chat. Most agentic process automation projects benefit from a voice surface for a subset of the workflow; we'll tell you which during the audit.
How do you measure voice agent quality before rollout?
Four metrics, evaluated on a held-out set drawn from your real call recordings. (1) Intent-match accuracy: does the agent classify and respond to each call type correctly? Target >85% on tier-1 intents. (2) Hallucination rate per 100 calls: how often does the agent invent a policy, balance, or feature that doesn't exist? Target <2 per 100 calls before rollout. (3) Barge-in latency: when the caller interrupts, how fast does the agent stop? Target <300ms p50, <500ms p95. (4) Mean-time-to-action (MTTA): from call start to the moment the agent does the thing the caller wanted. Baseline measured during the audit, target movement defined per workflow. We sample 200 calls/week minimum during pilot for human review. Listicle competitors don't publish eval methodology because they don't run it. This is the disqualifying signal when shortlisting any voice agent platform vendor.
When is voice the WRONG primitive?
Five honest scenarios where we'll tell you not to build a voice agent. (1) Low-bandwidth or noisy environments: factory floors, transit, anywhere the STT layer can't get clean audio reliably. Text or a structured form wins. (2) Regulated read-backs: when the caller must read or confirm specific text verbatim (financial disclosures, prescription instructions), voice is slower and error-prone vs a screen with a checkbox. (3) Complex UI flows: flight selection, multi-product configuration, anything that needs a comparison grid. Voice forces serial enumeration of options. (4) Cost-sensitive workflows at massive scale: if your per-interaction budget is sub-$0.01 and volume is in the tens of millions, even the cheapest chained stack will outprice a chat or async channel. (5) Users who don't want to talk: most under-35 cohorts in B2C SaaS, anyone in an open-plan office, anyone who's already on a video call. The right voice strategy is often a voice surface for some of your traffic, not all of it.
Can you put voice agents inside a Flutter or mobile app, not just on a phone line?
Yes, and that's a meaningful differentiator. Most voice ai company SERPs are telephony-centric (Vapi, Retell, Synthflow all start at a phone number). We ship the ai voice assistant pattern as a voice copilot inside iOS, Android, Flutter, and web apps via LiveKit Agents SDK or a direct Realtime-API WebRTC bridge. We wrote the UI kit that a lot of Flutter screens run on, so the voice surface ships with proper mic-permission UX, on-screen visual feedback during turn-taking, and interrupt handling that doesn't fight the OS. See the food-delivery capability pattern above for a shipped example — sub-600ms TTFT measured on real devices, +18% order-completion lift on the voice cohort. For mobile-specific work, the cross-link is Flutter app development.
How does this compare to platforms like Vapi, Retell, or Synthflow — should we just buy one?
Honest answer: sometimes yes. Buy a platform (Vapi, Retell, Synthflow, Bland, Air) when your workflow is standard (inbound deflection, outbound reminders, appointment scheduling), your call volume is moderate (under ~100K minutes/month), and you don't have a unique RAG corpus or unusual tool surface. You'll be live in days, not weeks, and the per-minute cost premium (typically 30–60% above a custom build) is worth it for the time-to-value. Build custom when you need (a) deep RAG over a proprietary corpus, (b) tool-use against internal systems no platform integrates with, (c) a mobile-native or in-app voice surface, (d) per-minute economics that platform pricing can't hit at your volume, or (e) a vendor-agnostic stack you can swap as the model landscape moves. Our audit will tell you which side you're on. We've recommended "go buy Vapi" to about a quarter of voice audit clients and that's the right answer for them. AI integration services often picks up the platform integration work when the answer is buy.
keep exploring Related pages.
Related pages.
Pick where you are.
Voice is rarely the only AI surface you're building. These sibling pages cover the adjacent decisions: text channels, multi-step agents, the Realtime API itself, and mobile-native delivery.
AI Chatbot Development
The text-channel sibling — when async beats voice or voice is wrong primitive.
AI Agent Development
Multi-step agents with tool use — voice is one interface to the same agentic stack.
OpenAI Development
Realtime API depth and openai realtime api development — the speech-to-speech path.
AI Integration Services
Plug your voice agent into Salesforce, Zendesk, HubSpot, EHR, OMS.
Flutter App Development
Mobile-native voice copilots inside your Flutter / iOS / Android app.
AI Automation Agency
When the voice turn is one step in a larger agentic process automation pipeline.
AI Consulting
Pre-build discovery audit when voice may or may not be the right primitive (often, it's not — we'll say so).
Generative AI Development Services
When voice is one channel in a wider custom AI build — app shell, retrieval, agent layer, monitoring. The umbrella AI development company pillar.
Agent reliability benchmark (2026-Q3)
Voice agents run the same multi-step planning loop as text agents. The reliability rubric (pass@1, error recovery, mean steps) carries over directly.
Ready to ship Hire an AI voice agent development team
Hire an AI voice agent development team
that ships eval data, not demo reels.
Book a free AI voice agent audit. We'll review your call recordings or use-case shortlist, recommend the architecture (Realtime · chained · unified) and telephony stack (Twilio · Telnyx · LiveKit · Daily), project per-minute cost in $/min terms, design the 4-metric eval set, and give you a 90-day voice agent roadmap. No deck, no obligation to build.
30 min, async or live Per-minute cost projection included Architecture + telephony pick + eval-set design
Updated May 23, 2026 · By Navin Sharma