The moment voice AI stopped being a demo

Voice AI has been "almost there" for five years. Text-to-speech got good. Speech-to-text got good. LLMs got good. But the glue — the real-time, low-latency, interruption-aware, context-managing orchestration between them — kept falling short. Any real production deployment stumbled on the same problem: the human would say something, the bot would think, the pause would be too long, and the human would either interrupt or hang up.

In 2026, three platforms fixed this well enough for real enterprise deployment: Retell AI, Vapi, and LiveKit Agents. We've now shipped voice agents on all three for clients in the Middle East and India, and we've built one from scratch on top of LiveKit's lower-level primitives. This is what we learned, what each one is good at, and how to pick.

Why voice is fundamentally harder than text

Before comparing platforms, understand what makes voice agents an engineering-hard problem.

Latency budget is unforgiving. Humans expect a reply within 500-800 milliseconds. Anything longer and the conversation feels wrong. Text agents can take 3-4 seconds and feel fine. Voice cannot.

Turn-taking is a UX design problem, not just a technical one. When does the agent talk? When does it listen? What happens when the user interrupts? What about background noise, silence, someone thinking? All of this is non-trivial.

State management is real-time and stateful. Text agents can restart a conversation from a log. Voice agents can't recover a lost audio stream. Connections drop. Networks stutter. Users hang up mid-sentence. All of this must be handled gracefully.

Emotion and tone matter. A robotic-sounding voice loses trust faster than a robotic-sounding text agent. Prosody, pace, intonation — these are product features, not audio settings.

Compliance is different. Recording, consent, PII handling — voice-specific regulations apply on top of general privacy law. GDPR, DPDP, and regional telecom regulations all touch voice differently.

The platforms below each solved a version of these problems. What they optimize for differs.

Retell AI

The pitch: The fastest way to ship a working voice agent. Managed platform. Batteries included. Configure prompts, connect tools, deploy.

What actually works well.

Retell's latency is genuinely impressive. Sub-800ms first-response is achievable out of the box, and their turn-taking model is polished. Users describe conversations as "feeling natural" — a rare compliment for voice AI.

The developer experience is the smoothest of the three. A workable agent from scratch takes an afternoon. The dashboard is well-designed. Deployment is one click.

Retell's approach to interruption handling — they call it "backchanneling" — sounds subtle but changes the feel of a call. The agent produces small confirmatory sounds ("mm-hmm", "I see") that make interruption natural instead of jarring.

Where it stumbles.

Cost at scale is real. A high-volume voice deployment on Retell can hit thousands of dollars per week. For consumer scale, you'll want negotiated pricing or an alternative.

Customization ceilings show up on complex workflows. If you need deeply custom logic — multi-agent handoffs, sophisticated tool routing, integration with a specific IVR — you'll hit the platform's limits.

Non-English performance is decent but not the best. For Arabic and Indian languages, plan to evaluate carefully.

Best for: Startups and mid-sized teams that want to ship a good voice agent fast, especially in English-heavy customer support contexts.

Vapi

The pitch: Similar target to Retell — managed voice agent platform — with somewhat lower prices and a more flexible programming model.

What actually works well.

Vapi's programmability is a step above Retell. You can inject custom logic more easily, mix models at will, and hook into external systems with less friction. For a team with engineers who like control, Vapi feels better.

Pricing is more transparent and generally lower than Retell for comparable workloads.

Multi-language support is respectable. Vapi has invested in non-English voice quality; Arabic and Hindi are usable, though still not perfect.

Where it stumbles.

The latency envelope is slightly wider than Retell. Not by much, but enough to notice on some conversations.

Documentation is thinner. Ecosystem is smaller. When you hit an unusual edge case, you're often on your own.

Turn-taking is functional but less polished than Retell's. The agent will occasionally step on the user's words in ways that feel unnatural.

Best for: Engineering-heavy teams that want more control than Retell offers, or teams operating on tighter budgets.

LiveKit Agents

The pitch: Lower-level primitives. LiveKit is a WebRTC infrastructure platform; their Agents SDK is a framework for building voice agents on top. Not fully managed — you run the agent logic.

What actually works well.

Full control. Your agent code runs where you want it. Your model choices are unlimited. Your latency budget is only limited by your infrastructure.

The WebRTC foundation is genuinely excellent. Audio quality, connection reliability, and geographic distribution are best-in-class.

For teams that need on-prem or in-VPC deployment — regulated industries, KSA/UAE data-residency compliance — LiveKit is the pragmatic default. Retell and Vapi have less mature enterprise deployment stories.

Multi-agent coordination is easier to implement here than on the fully-managed alternatives. If your call flow requires handoffs between multiple specialist agents, LiveKit's architecture makes this natural.

Where it stumbles.

You build more of the agent yourself. This is exactly what you want if you have the engineering to invest and exactly what you don't want if you don't.

The tooling for prompt iteration, evaluation, and monitoring is less polished than the managed platforms. You'll want to build or integrate additional layers.

Time-to-first-agent is longer. Plan for a week or two of engineering before you have a shippable prototype, versus an afternoon on Retell.

Best for: Teams with real engineering depth, regulated deployments, complex call flows, or high volume where the cost math justifies the extra work.

Building from raw components

For a specific client — a KSA fintech with strict data residency and heavy Arabic call volume — we skipped all three platforms and built on lower-level primitives: LiveKit for transport, Whisper for STT, an in-house TTS model for Arabic, Claude 3.5 for reasoning, our own state machine for turn-taking.

Reasons this made sense:

• Full control over the Arabic pronunciation and dialect.
• Data never left the KSA cloud region.
• Custom compliance logging that no platform supported.
• Deep integration with the client's IVR and CRM.

Reasons it doesn't usually make sense:

• 3-4 months of engineering to reach production quality.
• Ongoing maintenance of custom voice infrastructure.
• Duplicating what the platforms give you for free.

If your requirements are unusual enough to justify custom, do it. Otherwise, one of the three platforms will save you months.

The head-to-head

| Dimension | Retell | Vapi | LiveKit Agents | Custom | |-----------|--------|------|----------------|--------| | Time to first working agent | 4 hours | 6-8 hours | 1-2 weeks | 3-4 months | | Latency (first response) | Excellent | Very good | Good (depends on infra) | Depends | | Turn-taking polish | Best | Good | Good | Depends on work | | Language support | English-first, others OK | Multi-language investment | Model-dependent | You choose | | Cost at scale | High | Moderate | Low (if efficient) | Variable | | Customization | Limited | Moderate | Full | Full | | Data residency options | Limited | Limited | Full | Full | | Best for | Speed-to-ship | Balanced choice | Regulated / complex / high-volume | Unusual requirements |

When each platform pays back

Retell: startup building a voice product with English-speaking users, focus on time-to-market.

Vapi: small-to-mid team, more engineering than Retell customers, cost-conscious.

LiveKit Agents: enterprise with real engineering depth, regulated data, complex call flows, or high volume.

Custom: unusual language or regulatory requirements, deep integration needs, willingness to invest 3+ months.

The design patterns that matter more than the platform

Regardless of which platform you pick, the same design patterns produce good voice agents. Get these right; the platform choice matters less.

Design turn-taking explicitly. When does the agent talk? When does it interrupt? What sounds does it make while listening? These are product design decisions. Don't leave them to defaults.

Handle silence gracefully. Some users think before speaking. Some are on bad connections. Your agent's response to five seconds of silence should not be "Are you still there?" on repeat.

Fail toward humans. If confidence drops, if the tool call errors, if the user seems frustrated — hand off to a human politely. The best voice agents know when they're beaten.

Keep prompts short. Voice UX punishes verbose agents. Every sentence should be earning its place.

Design your voice like a product. Pace, tone, formality. Test with real users. The wrong voice makes the right content feel wrong.

Test on real network conditions. Everyone tests on office wifi. Test on 4G, on 3G, on packet loss. That's where voice agents fall apart.

Frequently asked questions

How much does a production voice agent cost per call? Depends on call length and platform. Rough estimates for a 5-minute call: Retell $0.60-1.20, Vapi $0.30-0.80, LiveKit Agents self-managed $0.10-0.40. Custom builds can be lower still but amortize huge upfront cost.

What about compliance and recording consent? Every jurisdiction has different rules. In India and the GCC, most jurisdictions allow single-party consent recording, but disclosure at call start is best practice. The platforms provide recording; you're responsible for the legal framework.

Can voice agents replace human call center staff? Not entirely. Best deployments are hybrid — voice agents handle tier-one, hand off to humans for anything complex or emotional. Full replacement typically fails on customer trust and complex cases.

How do these handle emotion? Better than they used to. The voice quality of GPT-4o-realtime, Claude's voice mode, and Deepgram's Aura is genuinely convincing on average. Poor emotional tuning is now a design problem more than a technology problem.

What languages actually work? English works excellently everywhere. Arabic works reasonably well but requires attention to dialect. Hindi is usable, other Indian languages varying quality. Chinese and Japanese are strong on some platforms. Test on your real language mix.

The strategic read

Voice agents in 2026 are a legitimate product category. Customer support, sales qualification, appointment scheduling, and outbound calls all have working, deployable voice AI solutions. The bar to entry has dropped from "18-month engineering project" to "afternoon with Retell." That changes the calculus for many businesses.

Whether you build on a platform or roll your own is now genuinely a strategic choice, not a forced hand.

At Xenolve we ship voice agents for enterprise clients across India, KSA, and UAE — including regulated deployments where data residency and Arabic/Hindi quality matter. If your team is evaluating voice AI and wants an outside perspective on platform choice, integration approach, or Arabic/multilingual quality, get in touch. We've built enough of these to know where the traps are.

Voice AI works now. The question is what you build with it.