The AI voice-agent stack: where automated phone support actually works

Phone is the channel most operators deprioritize in their AI roadmap — and the one their customers care about most. That gap is closing fast. A new generation of voice-agent infrastructure has matured enough to run inbound support queues at real scale, with containment rates and cost structures that finally justify the swap from human-heavy IVR queues. But the failure modes are nastier than chat, and the vendor landscape is split into three categories that barely overlap. Here’s how to read the map.

The two tiers: infrastructure vs. managed platforms

The voice-agent stack has a clean fault line: infrastructure-layer platforms (Vapi, Retell, Bland) and managed customer-service platforms (Intercom Fin Voice, Sierra, Decagon). Picking the wrong tier for your context is the most expensive mistake operators make.

On the infrastructure side, mid-market bring-your-own-key platforms like Vapi ($0.05/min) and Retell ($0.055/min) layer an explicit per-minute fee on raw component costs.

Vapi’s $0.05/min is a platform orchestration fee only — you connect your own STT, LLM, and TTS providers on top of it.

A basic setup (Deepgram + GPT-4o mini + PlayHT) runs around $0.14–$0.15/min all-in; switch to premium providers like ElevenLabs and Claude Sonnet and you’re looking at $0.25–$0.33/min. The flexibility is real — you choose the speech-to-text provider, the language model, the text-to-speech engine, and wire them together through the API — but so is the engineering overhead.

Retell takes a cleaner billing approach. There are no platform fees and no feature gating; every plan gets full access to the entire platform, and you pay only for what you consume.

Once a call is transferred to a human, the AI voice agent fee stops — only the telephony fee continues for the remainder of the transferred call.

Bland moved to subscription-based pricing in late 2025. After a December 2025 pricing increase, its free Start plan runs $0.14/min, the Build plan is $299/mo at $0.12/min, and Scale is $499/mo at $0.11/min.

Bland bundles LLM, STT, TTS, and telephony into a single rate but carries an estimated 15–40% markup on underlying components.

On the managed platform side, the economics are entirely different. Intercom Fin Voice is the incumbent’s answer: Fin Voice is Intercom’s AI agent for phone support — it answers every call instantly, uses your existing knowledge base to resolve questions, and when human help is needed, seamlessly hands off with full context.

As of now, Fin Voice is limited to select customers partnered with the Intercom sales team, supported on workspaces hosted in the US, EU, and Australia. The January 2026 Fin 3 launch added Procedures for complex multi-step query resolution and a Simulations testing suite — a powerful testing environment to ensure Fin is behaving as expected and catch issues before they reach customers.

Sierra is the enterprise-native alternative. In May 2026, Sierra raised $950M at a $15.8B valuation, led by Tiger Global and Google’s GV.

Sierra acquired Receptive AI in March 2026, bringing native voice capability to a platform that had been chat-focused at launch. Pricing is not published: contracts typically start at $150,000+ annually, covering platform licensing, implementation fees, and usage costs. Sierra’s pricing model is outcome-based — you pay only when the software achieves specific outcomes such as a resolved conversation, a saved cancellation, or an upsell; if the conversation is unresolved, in most cases there’s no charge.

Decagon is the fastest-growing entrant in this tier. Decagon launched AI voice agents in partnership with ElevenLabs, combining ElevenLabs’ expressive voice models with Decagon’s agentic platform to make phone support as seamless as live chat.

It’s a serious enterprise platform with strong voice, automation, and compliance credentials, with pricing appearing to start around $95K/year. One early customer, Chime, managed a 70% resolution rate across both chat and voice. Compare Decagon vs Sierra and Decagon vs Intercom Fin if you’re shortlisting in this tier.

Containment rate is the only metric that matters

Most teams obsess over per-minute costs. The number that actually determines ROI is containment.

Containment rate is the percentage of calls fully handled by AI without human escalation.

This ranges from 45% on poorly configured agents to 88% on well-trained ones with solid knowledge bases.

Getting from 65% to 80% containment on 1,000 calls per month saves hundreds of dollars monthly in escalation costs — more than switching platforms entirely.

The best-in-class enterprise deployments hit the top of this range. PolyAI demos have shown natural-sounding multi-turn conversations managing up to 80% call containment on transactional workflows like booking updates and account verification. The ceiling is real, but so is the floor: anything below 60% suggests knowledge gaps or poor call flow design.

The knowledge base is the lever that operators most underestimate. Unlike chat, callers expect immediate, conversational answers — so Fin works best when your knowledge base is complete, up to date, and written in plain language. Thin or outdated content is the most common reason Fin escalates calls unnecessarily.

Integration depth matters as much as content quality. An AI agent that cannot book appointments is a sophisticated FAQ bot. The value of voice AI is in completing the task. Decagon’s AI Actions model reflects this: through integrations with Stripe, Shopify, and Salesforce, the platform can process refunds, update orders, verify identity, and create tickets — without escalating to a human.

Latency: the failure mode nobody plans for until it’s on customers

Measured across production calls: Retell AI averages 580–620ms, Vapi hits 500–600ms with optimized provider pairings. For reference, natural human conversation turn-taking occurs at 200–300ms — anything below 700ms feels conversational, but above 900ms, callers notice and disengage.

Based on analysis of 4M+ production calls, target P50 under 1.5 seconds and P95 under 5 seconds for cascading architectures (STT → LLM → TTS). The key insight: average latency hides problems — a 500ms average can mask 10% of calls spiking to 3+ seconds.

Sierra’s multi-model “constellation” architecture creates a specific trade-off: routing through multiple AI models for accuracy checking can introduce latency, and third-party analysis notes that even a 700ms+ delay feels long and can create awkward pauses in live voice environments.

Where escalation breaks — and what good looks like

Two failure patterns dominate in production. The most common failure is the dead-end escalation: the agent recognizes it cannot help, says so, and then either loops back to the beginning of the menu or drops the caller with a hold time.

The second most common failure is the context-free handoff: the human agent picks up with no record of what was said, what was tried, or what the customer’s current state is.

Good escalation design is architecturally different from bolting on a “transfer to human” button. Retell AI’s call transfer passes full conversation context to the human agent, so the caller does not repeat themselves.

Intercom Fin’s escalation continues through the existing workflow — forwarding to a team, offering a callback, or checking availability — so the experience stays coherent.

Effective escalation design means the agent detects escalation signals early — rising sentiment frustration, repeated reformulation of the same request, explicit requests for a human — and routes before the experience deteriorates further.

Security is the underrated third failure mode. Voice prompt injection works the same way as text prompt injection — hidden instructions override the AI’s intended behavior — but the consequences are immediate: tool calls execute in the real world across payments, bookings, and account updates. The highest-risk vector is indirect injection via retrieved CRM data that reaches the LLM context silently.

Choosing your tier

Infrastructure (Vapi, Retell, Bland) — right for engineering teams that want full model control, are comfortable managing four to five vendor relationships, and are running high-volume or highly customized call flows. Expect $0.14–$0.33/min all-in depending on model choices. Budget engineering cycles for ongoing optimization; the platform won’t do it for you.

Managed enterprise (Sierra, Decagon, Intercom Fin Voice) — right for ops teams that want a vendor to own deployment and tuning. Legacy CX providers are trapped in a conflict: their revenue depends on seat-based pricing, so the more effective their AI becomes, the fewer seats clients need — undermining their own revenue model. The newer managed platforms (Sierra, Decagon) are structured around outcomes instead. The cost is real: Sierra starts around $150K/year in reported contracts, Decagon around $95K/year. The value proposition is the managed tuning and the CRM action layer, not just the conversation.

For mid-market teams that already run Intercom, Fin Voice is the lowest-friction path — it uses your existing knowledge base and your existing workflow routing, though availability is still limited to sales-partnered accounts as of June 2026.

---

The AI voice-agent stack is genuinely production-ready for high-volume, routine inbound support — billing inquiries, status checks, appointment scheduling, account updates. The ceiling breaks when queries require policy judgment calls, genuinely novel situations, or emotional de-escalation that the training data never anticipated. Build your containment strategy around that honest ceiling, invest in escalation context-passing before you invest in voice quality, and the economics work.