A voice agent is judged in the pause before every reply
Chat users forgive a slow answer. Callers do not. If your agent pauses too long, talks over the caller or mishears a date of birth, they hang up or ask for a human. Good AI voice agent development is mostly about the unglamorous parts: latency, interruptions, telephony, handoff, and hundreds of simulated calls before a real patient or customer hears it.
We have built this for ourselves. Clinic Answering Service, our AI receptionist for US medical practices, answers every call 24/7 in English and Spanish, books real appointments, checks insurance eligibility on the call and syncs with the practice’s EHR. That work shapes how we build AI voice agents for healthcare, from clinics and dental practices to labs, and for service businesses that live on appointments.
Voice is usually one channel among several. Many projects pair it with WhatsApp confirmations and an AI agent back end that serves chat as well, so a caller who hangs up can finish by message.
AI voice agent development services
Five common builds. Each includes call flows, integrations, a suite of simulated test calls and a handoff path to your staff.
AI receptionist for clinics
The problem
Phones ring through lunch and after hours, patients reach voicemail and book elsewhere, and front-desk staff juggle calls with the people standing in front of them.
How we build it
The agent answers, identifies the patient, books or moves appointments in your practice-management system or EHR, checks insurance where the integration allows, and transfers urgent or clinical calls to staff with a summary. For US practices, every vendor in the chain is covered by a BAA.
What you get
- Call flows for booking, rescheduling and FAQs
- PMS or EHR integration
- Urgent-call escalation rules
- Transcripts and call summaries
Appointment booking agents
The problem
Service businesses lose bookings because callers cannot get through, and staff spend hours on calls that end with a slot being picked.
How we build it
We connect the agent to your calendar or booking system, encode the rules (durations, buffers, who does what) in code, and confirm each booking by SMS or WhatsApp. The booking system prevents double-booking; the model is never trusted to.
What you get
- Booking tools with live availability
- Business rules enforced in code
- Confirmation by SMS or WhatsApp
- Daily booking report
Outbound reminder & follow-up calls
The problem
No-shows and unpaid balances need a call, but nobody has time to make hundreds of them.
How we build it
Calls go only to numbers with recorded consent, inside permitted hours, with an opt-out on every call. The agent confirms, reschedules or sends a payment link, and writes the outcome back to your system.
What you get
- Consent and opt-out checks before dialling
- Calling-window and retry rules
- Outcomes written back to your system
- One opt-out list across channels
Self-built voice stack
The problem
Hosted platforms add a per-minute fee on top of model and telephony costs, limit how far you can change the pipeline, or do not meet your data-residency rules.
How we build it
We build on open-source frameworks such as Pipecat or LiveKit Agents in your own cloud account, choose speech-to-text, model and text-to-speech providers per language, and run it with monitoring and autoscaling.
What you get
- Voice pipeline in your cloud account
- Provider choice per language
- Latency and error dashboards
- Runbook and on-call handover
Voice testing & evaluation
The problem
A voice agent sounds fine in five test calls, then fails on accents, background noise, interruptions and callers who change their minds.
How we build it
We run simulated callers with scripted personas, accents and noise against every change, score the transcripts against expected outcomes, and replay real calls (consented and redacted) as regression tests.
What you get
- Simulated-caller test suite
- Scored transcripts per release
- Regression set from real calls
- Latency report per turn
Voice AI platforms and tools we use
Hosted platforms get you live fastest. Open-source frameworks give you control. Vendor details as of October 2026; plans and terms change, so we confirm them per project.
Vapi
A developer-focused platform with APIs for assistants, tools and phone numbers. HIPAA support is offered as a paid add-on with a signed BAA, as of October 2026.
Retell AI
A voice agent platform with a visual flow builder and an API. Offers a self-serve BAA as of October 2026; check exactly what it covers on your plan.
Bland
An enterprise-focused platform with dedicated infrastructure options. A BAA is offered on enterprise plans as of October 2026.
Pipecat
A BSD-licensed Python framework for real-time voice and multimodal pipelines, version 1.x as of October 2026. You choose each provider and host it yourself.
LiveKit Agents
An Apache-licensed framework, version 1.x as of October 2026, built on LiveKit’s WebRTC media server, with phone calls through SIP.
Twilio and SIP trunks
Twilio Programmable Voice and Elastic SIP Trunking are HIPAA-eligible under a BAA for US voice traffic. In India we work through licensed telephony providers.
Tools and services we use alongside it
- Vapi
- Retell AI
- Bland
- Pipecat
- LiveKit Agents
- Twilio
- SIP trunking
- WebRTC
- Deepgram
- ElevenLabs
- Python
- Node.js
- Redis
- PostgreSQL
Versions and vendor details checked October 2026. We pin versions per project and upgrade on a schedule you approve.
The latency budget for one turn
People leave very short gaps between turns. A cross-language study in PNAS (Stivers et al., 2009) found the most common gap in all ten languages it studied was between 0 and 200 milliseconds. No AI pipeline matches that, but every stage adds delay, so we budget each one and measure it on real calls.
| Stage | What happens | Typical share of the wait | How we cut it |
|---|---|---|---|
| Telephony and network | Audio travels from the caller’s phone through the carrier to your servers and back | Small but fixed | Host close to the carrier’s media region; avoid extra hops |
| End-of-turn detection | Deciding the caller has finished speaking | Often the largest single wait when tuned cautiously | Turn-detection models, tuned per call type |
| Speech-to-text | Finalising the transcript | Short with streaming recognition | Streaming models chosen per language and accent |
| Language model | Time to the first words of the reply | The most variable stage | Short prompts, prompt caching, faster models for routine turns, streaming |
| Tool calls | Calendar, EHR or CRM lookups | From negligible to seconds | Prefetch likely data, cache, and say “let me check” while waiting |
| Text-to-speech | Time to the first audio | Short with streaming synthesis | Start speaking on the first sentence, not the whole reply |
Interruptions matter as much as speed. When a caller talks over the agent, it should stop within a moment and listen, not finish its sentence.
Hosted platform or self-built?
| Option | Control | Cost model | Compliance options (as of October 2026) |
|---|---|---|---|
| Vapi | High for a hosted service: bring your own models and voices, deep API | Per-minute platform fee plus model, speech and telephony costs | HIPAA mode with BAA as a paid add-on |
| Retell AI | Medium: flow builder plus API | Per-minute pricing that varies with the model and voice chosen | Self-serve BAA |
| Bland | Medium to high on enterprise plans | Per-minute pricing; enterprise terms | BAA on enterprise plans |
| Pipecat (self-built) | Full: every stage is your code | No platform fee; you pay each provider, hosting and engineering time | Whatever your providers and cloud sign; you hold the architecture |
| LiveKit Agents (self-built) | Full, with a managed media layer available | As above, or LiveKit’s cloud for the media layer | As above; check LiveKit Cloud terms if you use it |
Hosted wins for a first launch and modest volumes. Self-built wins when volume is high enough that per-minute platform fees hurt, when you need a language or provider a platform does not support, or when data must stay in a specific country.
Telephony in the US and India
In the US we use Twilio or a carrier SIP trunk, and either port the practice’s number or forward to the agent so callers keep dialling the number they know. Twilio signs a BAA for its HIPAA-eligible products, which cover US voice traffic.
India is different. Calls run through licensed cloud-telephony providers or operator SIP trunks, and outbound calling is shaped by TRAI’s commercial-communication regulations (TCCCPR), amended in February 2025. Promotional calls must come from 140-series numbers, by businesses registered with their telecom operator under that framework, and callers can block them through the do-not-disturb registry. Entities regulated by RBI, SEBI, IRDAI or PFRDA, and government bodies, must use 1600-series numbers for service and transactional calls to existing customers and citizens. Confirm the current rules with your telephony provider before any outbound campaign.
A HIPAA checklist: the BAA chain
A voice call carries protected health information through more vendors than people expect. Every one of these needs a BAA, or must never see PHI:
- Telephony carrier and number provider
- Voice platform, if hosted
- Speech-to-text provider
- Language model provider
- Text-to-speech provider, which receives the names and appointment details it speaks
- Cloud host, database and backups. HHS guidance says a provider is still a business associate even if it stores only encrypted data without the key
- Logging, tracing and analytics tools that store transcripts
- Recording storage, plus any integration middleware between the agent and the EHR
Beyond the paperwork: retention limits for recordings and transcripts, role-based access, audit logs, and a recording notice where state law requires every party to consent. Our HIPAA readiness work covers the rest of the programme; we help you prepare, we do not certify.
Designing the human handoff
- Triggers: the caller asks for a person, the agent misunderstands twice, the call is clinical or urgent, or the caller is upset.
- Emergencies: the agent tells the caller to hang up and dial 911 (or 112 in India) and flags the call. It never tries to triage.
- Warm transfer: staff hear or see a short summary before they pick up, so the caller does not repeat themselves.
- Nobody free: the agent takes a callback request with a summary and gives an honest callback window. It never loops.
Outbound calls and consent in the US
On 8 February 2024 the FCC ruled that AI-generated voices count as “artificial” under the TCPA. Informational calls such as reminders generally need the person’s prior express consent, and telemarketing calls need prior express written consent. Since 11 April 2025, people can revoke consent by any reasonable means, and requests must be honoured within 10 business days. State laws add calling hours and their own consent rules. We build the consent store and opt-out handling (including “stop calling me” said mid-call); your counsel signs off the campaign.
Testing with simulated callers
Before go-live we run hundreds of simulated calls: elderly callers, noisy streets, strong accents, people who interrupt, change their minds, give a wrong date of birth or try to get details about another patient. Each call is scored against the expected outcome. The suite reruns on every change, and live monitoring tracks transfer rate, booking completion and response latency.
What drives the cost
- Minutes: call volume × call length, multiplied across platform, model, speech and telephony charges.
- Model and voice choice: larger models and premium voices cost more per minute.
- Telephony: numbers, carrier minutes and, for outbound in India, registration.
- Integrations: EHR and practice-system work is often the largest build cost.
- Compliance: BAA add-ons, recording storage and reviews.
Indian languages: what to expect
Speech recognition and synthesis for Hindi, Tamil, Malayalam, Telugu and other Indian languages have improved quickly, with providers such as Sarvam and open models from AI4Bharat focusing on them. Code-mixed speech, names, addresses and numbers over narrowband phone audio are still hard. Start with one language, let callers choose at the start, confirm key details by SMS or WhatsApp, test with real speakers of the local dialect, and keep a human fallback.
How to choose a voice agent partner
- Can they show per-stage latency from a live system?
- Will they list every vendor that hears your callers and its BAA or DPA status?
- Do they test with simulated callers before your patients hear it?
- Is the handoff designed, or just “transfer to reception”?
- Will they say when a hosted platform is cheaper than a custom build?
Voice AI we build and run
Live products that take real calls, which is why our latency, handoff and compliance advice comes from production rather than demos.
Clinic Answering Service
HIPAA-compliant AI receptionist for US practices, 24/7 Answers every call 24/7 for US practices in English and Spanish, books real appointments, checks insurance eligibility during the call and syncs with the EHR. Visit clinicansweringservice.com
Appointment Reminder App
WhatsApp & SMS reminders that grow into an AI inbox Adds voice AI calls and payment recovery to WhatsApp and SMS reminders, with consent records kept for every contact. Visit appointmentreminderapp.comA typical first engagement: one call type, live and supervised
A fixed-scope project of about six weeks that ends with a voice agent handling one or two call types under staff supervision.
Call audit
We review a sample of calls or call logs, group them by reason and pick the two or three call types worth automating first.
Flows, integrations & vendors
We write the call flows and handoff rules, connect the booking system, and choose hosted or self-built with the BAA or DPA chain checked.
Simulated-caller testing
Hundreds of simulated calls across personas, accents and noise, plus staff test calls, until the agent passes the agreed test set.
Supervised go-live
Overflow and after-hours calls first, with staff reviewing transcripts daily, then wider hours once the data supports it.
Guides and related pages
Guides
Questions about AI voice agent projects
We quote after a short discovery and do not publish prices. Build cost depends on the number of call flows, the systems to integrate (EHR and practice systems take the most work), the languages and the compliance requirements. Running cost is mostly per minute: platform, model, speech and telephony charges multiplied by your call volume. Both figures are set out in the written proposal.
Yes, for booking, rescheduling, directions, hours, intake and routine questions. Clinical questions and emergencies go to staff or emergency services. In the US every vendor in the call path must sign a BAA, and patients should be told they are speaking with an AI assistant.
A hosted platform is usually right for a first launch and moderate volumes, because it is faster to ship and someone else runs the media layer. A self-built stack on Pipecat or LiveKit Agents makes sense at high volume, for unsupported languages or providers, or when data must stay in a particular country.
Compliance comes from the whole setup, not a product label: a BAA with every vendor that handles PHI, access controls, audit logs, retention limits and staff policies. We design and document that setup and help you prepare for audits. We do not certify anyone.
It can, with caveats. Indian-language speech models have improved a lot, but mixed-language speech, names and numbers on phone-quality audio still cause errors. We start with one language, confirm key details by SMS or WhatsApp and test with real local speakers before launch.
It transfers the caller to staff with a short summary, or, if nobody is available, takes a callback request and tells the caller when to expect it. Emergencies are directed to emergency services immediately. The agent is designed never to trap a caller in a loop.
Missing calls you should be answering?
Tell us your call volume, your booking system and your busiest hours. We will tell you what a voice agent can realistically handle.






Compliance-first delivery