Voice agents are having a moment — and for good reason: voice can remove friction where typing is slow, impossible, or unnatural (phones, on-the-go users, call centers).
But voice also raises the bar: real-time latency, turn-taking, interruptions, escalation, and trust.
Why voice is different (business impact, not technical)
A voice interaction has three unforgiving properties:
- Users expect speed (awkward pauses kill trust)
- Mistakes feel more personal (especially in support/sales)
- Escalation must be smooth (otherwise it's worse than an IVR)
Frameworks and platforms in this space explicitly focus on turn-taking, interruptions, and realtime pipelines because those details drive user trust. (docs.livekit.io)
The "Should this be voice?" checklist
Voice is a good fit when:
- The workflow is already happening on phone calls
- Users are in motion (field ops, logistics, scheduling)
- Speed matters more than long-form depth
- The conversation can be bounded with a clear escalation path
Voice is a bad fit when:
- The user needs to compare lots of detail (tables, options, long docs)
- The workflow requires heavy back-and-forth troubleshooting
- Compliance/verification requirements aren't figured out
Voice-native use cases that usually win
1) Scheduling and rescheduling
Bounded task, clear success condition, high volume.
2) Triage + warm transfer (support or services)
Let the voice agent gather context, classify intent, and route to the right human team.
3) Order status / basic account questions (with safe verification)
Great deflection use case — but only if identity and data access are handled carefully.
4) Lead qualification + routing
Voice can feel more natural than web forms, especially for services businesses.
Where voice agents fail (common product mistakes)
- No clear scope ("it can help with anything")
- No graceful fallback ("sorry, I can't help" loops)
- Too much autonomy in high-risk flows
- No ops ownership (who monitors, updates, and handles escalations?)
Voice UX requirements you should demand (as a buyer)
If you're paying for a voice agent, insist on:
- Expectation-setting ("Here's what I can do…")
- Confirmation for critical details (dates, addresses, payments)
- Interruptibility (people talk over systems; systems must handle it)
- Fast human handoff with context (summary + intent + collected fields)
- Auditability (transcripts, outcomes, escalations)
Build vs buy (voice edition)
Level 1: Off-the-shelf voice/agent capabilities
Good for quick proof of value, but limited control.
Level 2: Configurable voice-agent platforms (fastest time-to-value)
Examples:
- Vapi supports building voice agents that can make/receive phone calls; it explicitly positions "Assistants" (single-agent) and "Squads" (multi-assistant orchestration) as primitives, plus real-time conversations and phone/web integration. (docs.vapi.ai)
- ElevenLabs Agents positions itself as an orchestration platform combining speech-to-text, an LLM, and text-to-speech, with interruption and turn-taking logic and knowledge bases. (help.elevenlabs.io)
Level 3: Developer-led voice agent systems (maximum control)
Example:
- LiveKit Agents is an open-source framework for realtime, programmable voice agents; it highlights handling realtime media pipelines and also offers an Agent Builder for prototyping/deploying in-browser. (docs.livekit.io)
At the model layer, you'll also see realtime speech APIs (e.g., OpenAI's Realtime API supports realtime audio sessions and predefined voices). (platform.openai.com)
Next step
If you're considering voice, the best starting point is a Voice Agent Feasibility Assessment:
- pick one call type (e.g., scheduling)
- map escalation rules
- define KPIs (deflection, AHT, booking rate, CSAT proxy)
- choose the right level (1/2/3)