Executive summary
AI voice agents in 2026 are no longer “the future of the call center” — they are the present, in production, in collections, healthcare, retail and banking. What changed between 2023 and 2026 was the combination of three mature technologies: low-latency STT (speech-to-text), LLMs with native or near-native voice, and TTS (text-to-speech) indistinguishable from humans in neutral Spanish and Mexican Spanish.
The operational result: AHT drops 30-50%, FCR goes up 20-30%, CSAT goes up 10-20 points, and cost per call drops 60-80% in collections and tier-1 support. In this article we cover what is happening, what is real, what is hype, and how to evaluate it for your company.
Primary CTA: Do you run a call center and want to know if a voice agent applies to your case? Let’s talk with the Zippo team and build a scoped pilot.
Market size and why it matters
The global market for AI applied to call centers surpassed USD 4 billion in 2025 and is projected at over USD 12 billion by 2028 according to Grand View Research [Call Center AI Market, https://www.grandviewresearch.com/industry-analysis/call-center-ai-market-report, accessed 2026-09-06]. Growth is driven by:
- Drop in STT/TTS cost per minute to <USD 0.01 at many providers.
- Maturity of Realtime APIs (OpenAI, Gemini Live, AWS Nova Sonic) that reduce end-to-end latency to <500 ms.
- Open-source frameworks like LiveKit that enable self-hosting and data residency [LiveKit, https://livekit.io/, accessed 2026-09-06].
This matters because in Mexico and LATAM, call center labor costs are relatively low compared to the US, but the cost of not serving customers well is very high: churn, bad reputation, lost sales.
The four enabling technologies
1. STT (Speech-to-Text) — the agent’s ears
OpenAI’s Whisper [Whisper, https://openai.com/research/whisper, accessed 2026-09-06] democratized high-quality multilingual STT. In 2026 there are faster alternatives (Deepgram Nova-3, AssemblyAI Universal, Google Chirp 2) that process streaming audio with <300 ms latency. Choosing between providers comes down to: language, accent (Mexican Spanish vs neutral), cost per minute, and whether you need on-prem (Deepgram self-hosted leads here).
2. LLM — the agent’s brain
2026 LLMs are multimodal-native: they accept audio directly (not just transcription → text → audio). OpenAI Realtime API [docs, https://platform.openai.com/docs/guides/realtime, accessed 2026-09-06] and Google’s Gemini Live expose native audio in/out. The difference vs the classic STT→LLM→TTS pattern is:
- Lower latency: without transcribing to intermediate text, the model processes audio tokens directly.
- Better interruption handling: the model knows when the user interrupted without needing external VAD.
- Cost: in audio in/out, tokens are cheaper than in some separated STT+TTS stacks.
3. TTS (Text-to-Speech) — the voice
ElevenLabs [Conversational AI, https://elevenlabs.io/conversational-ai, accessed 2026-09-06] and Cartesia (with their Sonic model) lead streaming TTS with <200 ms latency. Quality in Mexican Spanish is indistinguishable from a human voice actor in blind tests. For collections, where tone matters, this changes the rules: the voice agent can sound empathetic, professional, firm — without falling into “robot voice”.
4. Orchestration — the glue
LiveKit Agents [https://livekit.io/, accessed 2026-09-06] is the reference open-source framework. It handles:
- Audio capture from SIP/PSTN or WebRTC
- VAD (voice activity detection)
- Streaming between STT, LLM, TTS
- Turn-taking and interruption handling
- Tools (transfer to human, open ticket, query CRM)
The advantage of LiveKit vs proprietary alternatives: self-hostable, no vendor lock-in, with SDK in Python and Node.
Use cases by industry
Collections
The case with the fastest ROI. A voice agent for collections in Mexico:
- Calls the debtor, identifies themselves, verifies identity with dynamic questions
- Offers payment options according to policy
- Processes payment via IVR or transfers to human agent if negotiation is complex
- Records payment promise in CRM
Typical metrics: AHT 90-180 seconds, FCR 70-85%, cost per call <MXN 8 vs MXN 35-60 human, neutral or positive CSAT in 60%+ of interactions.
Healthcare (reminders, triage)
Appointment reminders, post-consultation follow-up, symptom triage to route to emergency or telemedicine. In Mexico COFEPRIS regulation requires human validation of clinical decisions — the voice agent assists, does not replace the doctor.
Retail and e-commerce
Order status, returns, warranty management. Integration with Shopify, Mercado Libre, WooCommerce via MCP.
Banking and fintech
Card activation, fraud blocking, balance inquiry, KYC onboarding assistance. LFPDPPP and CNBV regulatory compliance mandatory; the voice agent must record and allow audit.
Real metrics in production
The following ranges are typical in production implementations during 2025-2026 (not demos):
| Metric | Human call center | Voice agent | Delta |
|---|---|---|---|
| AHT (Average Handle Time) | 240-360 s | 90-180 s | -40% to -50% |
| FCR (First Call Resolution) | 60-70% | 75-90% | +20% to +30% |
| CSAT (Customer Satisfaction) | 70-80% | 80-90% | +10 to +20 points |
| Cost per call | MXN 35-60 | MXN 5-12 | -70% to -85% |
| 24/7 availability | Only with expensive night shifts | Always | n/a |
| Response consistency | Variable per agent | 100% consistent | quality improvement |
DMG Consulting documents similar trends in [Contact Center AI, https://www.dmgconsulting.com/contact-center-ai, accessed 2026-09-06].
Voice agent vs human call center: when each wins
Voice agent wins when:
- The use case is repetitive, structured, tier 1.
- Volume justifies infrastructure (usually >5,000 calls/month per use case).
- Cost per human minute is prohibitive (24/7, multilingual).
- Response consistency is critical (compliance, regulatory).
Human wins when:
- Deep empathy is the factor (severe complaints, grief, complex negotiation).
- The case requires delicate ethical or legal judgment.
- The user explicitly asks for human (important: always offer escape).
- The case is new and there is no data to train effective prompts.
Hybrid wins when:
- The voice agent handles tier 1 and escalates to human when it detects frustration, complexity, or explicit request.
- This is the most common production pattern.
Compliance in Mexico
Three non-negotiables:
- LFPDPPP — explicit consent to record; privacy notice; right to oppose. The voice agent must self-identify as AI at the start of every call.
- NOM-151 — if the call generates a legal commitment (payment, contract), proof with legal value must remain.
- AI regulation in construction — Mexico is moving toward EU AI Act-like regulation. Design today with auditability, traceability and human-in-the-loop.
How to start
- Measure your baseline: AHT, FCR, CSAT, cost per call, most frequent call reason.
- Pick ONE use case that is repetitive and well-defined. Tier-1 collections, appointment reminders, order status.
- Design the human escape from day one. Without this, CSAT collapses on the second complaint on social media.
- 30-60 day pilot with metrics measured before and after.
- Scale what works, not what looks good in a demo.
Next step: Let’s talk about your voice operation and see if a voice agent is the right play. If your bottleneck is upstream (SIP trunk, DID, porting), start by quoting your trunk.
Sources
- OpenAI — Realtime API documentation
- LiveKit — Open source realtime platform
- ElevenLabs — Conversational AI
- OpenAI — Whisper research
- Grand View Research — Call Center AI Market Size
- DMG Consulting — Contact Center AI
- Forrester — The State of Customer Service 2026