TL;DR
Gartner predicts that by 2029 agentic AI will autonomously resolve 80% of common customer service issues without human intervention, cutting operational costs by 30%, and AI voice agents are how that shift reaches the phone line. A production AI voice agent answers in under 1.2 seconds, runs at about $0.12–$0.15 per connected minute, and handles the repetitive calls (booking, status, intake, after-hours) while escalating the rest to people. A custom build starts with a proof-of-concept from about $8,000–$25,000, and a single-workflow production agent typically lands at $35,000–$80,000. This guide covers how AI voice agents work, where they pay off by industry, what they cost, the TCPA and HIPAA rules that apply, and how to decide between building and buying.
Want an AI voice agent on your phones without the guesswork? Book a free 30-minute call with DestiLabs, top-ranked on Clutch for Voice & Speech Recognition. We'll pick your highest-volume call type and give you a costed plan with latency targets. → Book a call
What is an AI voice agent?
An AI voice agent is software that answers and makes phone calls in natural speech, understands what the caller wants, and completes the task in your systems, such as booking an appointment, checking an order, or taking a payment. It chains speech-to-text, a large language model, and text-to-speech into a loop that replies in under about 1.2 seconds.
That last part is what separates it from the phone tech most businesses already own. An IVR makes callers press 1 for billing. A basic voice bot can answer a question from a script. An AI voice agent listens to a sentence like "I need to move my Thursday cleaning to sometime next week, mornings if you have them," checks the calendar, offers two slots, books one, and texts a confirmation. If you're new to the category, our beginner explainer on what an AI voice agent is covers the basics. This page is the deeper guide: the mechanics, the money, and the rules.
| Feature | IVR phone menu | AI voice agent | Human agent |
|---|---|---|---|
| How callers interact | Keypad or fixed phrases | Free, natural speech | Free, natural speech |
| Resolves the request | Rarely; mostly routes | Yes, for routine calls | Yes, including complex ones |
| Availability | 24/7 | 24/7, unlimited concurrent calls | Shift hours, one call at a time |
| Typical cost | Low fixed platform fee | ~$0.12–$0.15 per connected minute | Fully loaded salary per seat |
| Best for | Simple routing | High-volume, repetitive calls | Judgment, empathy, exceptions |
If an IVR is what you're replacing, our guide to AI IVR walks through the migration.
How do AI voice agents work?
AI voice agents work as a streaming, real-time loop: listen, understand, act, speak, repeat. Every stage overlaps with the next, because waiting for each one to finish before starting the next would blow the latency budget on every turn.
- 1Telephony. The call arrives over SIP or a carrier like Twilio, and audio streams to the agent in small chunks.
- 2Speech-to-text (STT). A streaming recognizer transcribes the caller as they speak, and voice activity detection decides when they've finished their turn.
- 3Reasoning (LLM). A language model reads the transcript plus call context, decides what to do, and calls tools: look up the patient, check availability, create the booking.
- 4Retrieval and actions. The agent pulls verified answers from your knowledge base (RAG) and writes to your CRM, practice-management, or order system through APIs.
- 5Text-to-speech (TTS). The response is voiced as it's generated, so the caller hears the first words while the rest is still being produced.
Here's how a sub-1.2-second turn typically breaks down in a well-engineered stack:
| Pipeline stage | What it does | Typical latency budget |
|---|---|---|
| Network and telephony | Carries audio in and out | 50–150 ms |
| End-of-turn detection | Decides the caller has finished | 100–300 ms |
| Speech-to-text | Finalizes the transcript | 100–250 ms |
| LLM (time to first token) | Decides and starts the reply | 250–500 ms |
| Text-to-speech (time to first audio) | Starts speaking | 75–200 ms |
| Total, end of speech to first audio | All stages streamed and overlapped | Target under ~1.2 s |
Tool calls are the wildcard. A slow scheduling API can add a full second, which is why good agents say something natural ("Let me check what's open") while the lookup runs. DestiLabs voice deployments run at 0.99–1.2 seconds end to end. We publish how we measure that across production projects in our AI voice agent benchmark.
Why is latency the make-or-break metric?
Because callers judge the agent by the pause before it answers. Under about 1.2 seconds, the conversation feels natural. Past two seconds, people start talking over the agent or assume the line dropped, and hang-ups climb. Accuracy matters too, but a correct answer that arrives late still sounds broken. Latency is an architecture decision made during the build (streaming, model choice, where the servers sit relative to your callers), not something you tune afterward.
What can AI voice agents do? Use cases by industry
AI voice agents deliver the most value where call volume is high and conversations are repetitive: the calls that eat staff time without needing human judgment. The rule we give every client is the same: automate the routine, escalate the rest.
Across industries, the strongest jobs are inbound scheduling and rescheduling, reminders and confirmations, order and account status, new-customer intake, after-hours and overflow coverage, and outbound follow-ups to people who've already consented. The stakes of a missed call are real: just 61% of inbound business calls reach a live person, according to Invoca's 2025 analysis of more than 60 million calls. An agent that answers every call closes most of that gap on day one.
Healthcare
Clinics use AI voice agents for appointment booking, reminders, prescription-refill routing, and intake, reaching patients who'd rather call than use a portal. The HIPAA details, EHR integrations, and ROI math are in our guide to AI voice agents for healthcare, and dental practices have their own playbook in AI voice agents for dental practices.
Real estate
Brokerages and property managers lose leads every time an inquiry hits voicemail on a Saturday. A voice agent answers listing questions, captures buyer details, books showings, and logs maintenance requests. See AI voice agents for real estate for costs and setup.
Contact centers
High-volume support and service teams deploy voice agents as the first tier: authenticating callers, resolving status and account questions, and passing warm transfers with a summary so the human doesn't start from zero. The staffing and cost model for that setup is in contact center AI voice agents.
Fintech, ecommerce, and professional services
Fintech teams use voice for balance and transaction status, card controls, and identity verification, where accuracy and security come first (AI for fintech). Ecommerce brands use it for order tracking, returns, and after-hours sales calls. Law firms use it for 24/7 client intake, covered in our AI receptionist for law firms guide.
How much do AI voice agents cost in 2026?
AI voice agent cost has two parts: what you pay to build or subscribe, and what you pay per minute to run. The run cost for a well-built agent is about $0.12–$0.15 per connected minute, bundling speech recognition, the model, voice synthesis, and telephony into one number. That's the figure to compare across vendors, because headline platform rates usually leave something out.
| Path | Typical 2026 cost | What you get |
|---|---|---|
| Off-the-shelf platform | ~$50–$500+/mo, plus metered minutes | Templates, limited integrations, shared infrastructure |
| Custom proof-of-concept | From ~$8,000–$25,000 | One call type on real calls, measured latency and accuracy |
| Single-workflow production build | ~$35,000–$80,000 | One workflow, deep integrations, monitoring, guardrails |
| Multi-workflow or enterprise build | $80,000–$200,000+ | Several call types, compliance, multilingual, failover |
A quick worked example: a clinic group fielding 6,000 calls a month at an average of three minutes each runs 18,000 minutes. At $0.13 per minute, that's about $2,340 a month to answer every call, day and night. If the agent fully resolves 60% of those calls, it handles 3,600 calls a month that staff no longer touch, and recovers the after-hours bookings that used to go to voicemail. Against a $35,000–$80,000 build, payback usually lands inside the first year.
For the full breakdown, including hidden fees, platform markups, and per-minute math by vendor type, see our AI voice agent pricing guide.
Want real numbers for your call volume? Book a call and we'll scope your first voice workflow with build cost, per-minute run cost, and a latency target. → Book a call
Is using an AI voice legal? Compliance rules for AI voice agents
Yes, using an AI voice agent is legal, but the rules depend on which direction the call goes and what data it touches. Inbound calls (a customer rings you, the agent answers) carry the lightest burden. Outbound calls are where most of the legal risk sits.
TCPA and the FCC's 2024 ruling. In February 2024, the FCC issued a declaratory ruling confirming that AI-generated voices, including cloned voices, count as "artificial" voices under the Telephone Consumer Protection Act. In practice, an outbound call that uses an AI voice is treated like a prerecorded robocall: you need the recipient's prior express consent, and prior express written consent for marketing calls. Violations carry statutory damages of $500 to $1,500 per call, so outbound campaigns need consent records, opt-out handling, and calling-hour rules built into the agent, not bolted on later.
Call recording consent. Federal law allows one-party consent, but about a dozen states, including California, Florida, Illinois, Pennsylvania, and Washington, require every party to agree. The simple fix is a short disclosure at the start of the call: "This call is with an AI assistant and may be recorded." Many states and the FCC are also moving toward requiring AI disclosure, so we build it in by default.
HIPAA for healthcare. If the agent hears or handles protected health information, every vendor that creates, receives, stores, or processes that information (typically the speech, model, voice, and hosting layers) needs to sign a Business Associate Agreement. A carrier that only transmits the audio with transient access can qualify as a conduit that doesn't need a BAA, but that exception ends the moment it records, transcribes, or stores calls. Beyond BAAs, data needs encryption in transit and at rest, and call recordings and transcripts need retention and access controls. This is one of the main reasons healthcare teams choose custom builds over consumer-grade platforms.
Payments and finance. Agents that take card numbers need PCI DSS scoping (usually by pausing recording and routing card capture through a compliant processor), and financial-services calls may fall under GLBA and state rules on identity verification.
This isn't legal advice, and the rules keep moving. Have counsel review your outbound scripts and consent flows before launch.
AI voice bot vs. AI voice agent: what's the difference?
An AI voice bot usually means a system that answers questions or routes calls from a script. An AI voice agent is the more capable version: it reasons about an open-ended request and takes actions in your systems before deciding whether a human is needed. The terms get used interchangeably in vendor marketing, so ask a direct question in demos: "Show me the agent completing a booking, end to end, in our system." A bot can't; an agent can.
Should you build or buy an AI voice agent?
Buy when speed matters more than fit. Build when fit, compliance, or unit economics matter more than speed.
Buy a platform when: you have one simple call type (hours, directions, basic FAQs), your systems have native integrations with the platform, volume is modest, and you're fine with templated behavior. You can be live in days.
Build custom when: you need deep integrations with a specific EHR, PMS, CRM, or legacy system; you're in a regulated industry and need BAAs and audit trails across the stack; you want your own voice, escalation logic, and data ownership; or your volume is high enough that platform markups cost more per minute than owning the pipeline. A custom agent is an asset you own, tuned to your calls.
The low-risk path is the same either way: start with a scoped proof-of-concept on your highest-volume call type, measure latency, containment, and accuracy on real calls, then scale. Our AI voice agent development team builds and runs that loop. If you'd rather compare vendors first, our ranking of top AI voice agent companies lays out who builds what.
How do you choose an AI voice agent partner?
Choose on proof, not polish. Any vendor can run a scripted demo; few can show sub-1.2-second response time and stable accuracy on your real call patterns and integrations. Score candidates on six things:
- Measured latency. Ask for p50 and p95 end-of-speech to first-audio numbers from production, not a demo.
- All-in cost per connected minute. Speech, model, voice, and telephony in one figure, at your expected volume.
- Real-call accuracy and containment. What share of calls the agent resolves without a human, measured on unfiltered traffic.
- Integration depth. Can it write to your actual systems, or only read from a knowledge base?
- Escalation design. Warm transfers with a summary, clear triggers, and graceful recovery when the agent is unsure.
- Compliance posture. BAAs, consent capture, AI disclosure, recording controls, and data residency.
A partner who answers with benchmarks, like our published 0.99–1.2s latency and $0.12–$0.15 per minute, is far more likely to ship something that holds up in production. Our case studies show the receipts.
Which businesses should deploy an AI voice agent first?
Businesses whose phones are both busy and predictable see the fastest payback. Before any build, we score the use case on six factors, each from 1 to 5: call volume, repetitiveness, channel fit (do customers actually call?), integration readiness, escalation clarity, and brand tolerance for an AI voice. High scores on volume, repetitiveness, and channel fit with a clean handoff path make an ideal first build. Scheduling and status calls almost always top the list. Emotionally charged or high-stakes calls score low on brand tolerance by design, and stay with humans.
Clinics, dental and specialty practices, home services, property management, law firm intake, and ecommerce support all fit the profile well. If your call volume is low or every call needs judgment, a voice agent isn't the first thing to build.
Frequently Asked Questions
What is an AI voice agent?
An AI voice agent is software that holds natural spoken conversations over the phone and completes tasks in your systems, such as booking an appointment or checking an order. It chains speech-to-text, a language model, and text-to-speech into a loop that answers in under about 1.2 seconds, and it hands complex calls to a human.
How much do AI voice agents cost?
Run cost for a well-built AI voice agent is about $0.12–$0.15 per connected minute, covering speech recognition, the model, voice synthesis, and telephony. A custom build starts with a proof-of-concept from about $8,000–$25,000, and a single-workflow production agent typically costs $35,000–$80,000. Off-the-shelf platforms start around $50–$500 a month plus metered usage.
Is using an AI voice illegal?
No. Using an AI voice agent to answer inbound calls is legal. For outbound calls in the US, the FCC ruled in February 2024 that AI-generated voices count as artificial voices under the TCPA, so you need the same prior express consent required for prerecorded calls. Recording calls also requires consent from every party in about a dozen states, including California and Florida.
What is an AI voice bot?
AI voice bot is an older name for the same idea, but it usually describes a system that only answers questions or routes calls. An AI voice agent goes further: it understands free speech, reasons about the request, and takes actions in your CRM, calendar, or order system mid-call before escalating anything it cannot resolve.
How fast does an AI voice agent need to respond?
Under about 1.2 seconds from the moment the caller stops speaking. Past two seconds, callers talk over the agent or assume the line dropped. DestiLabs production deployments run at 0.99–1.2 seconds, which requires streaming every stage of the pipeline rather than waiting for each to finish.
Should I build or buy an AI voice agent?
Buy a platform if you need one simple call type live this month and your systems have native integrations. Build a custom agent when you need deep integrations, HIPAA or PCI compliance, your own brand voice, or when volume makes platform per-minute markups more expensive than owning the stack. A $8,000–$25,000 proof-of-concept answers the question on real calls.
Key Takeaways
- AI voice agents hold natural phone conversations and complete tasks in your systems; that's what separates them from IVRs and scripted voice bots.
- Latency decides whether callers stay: target under 1.2 seconds end to end. DestiLabs runs at 0.99–1.2s.
- Run cost is about $0.12–$0.15 per connected minute; custom builds start with an $8,000–$25,000 proof-of-concept and $35,000–$80,000 for a single-workflow production agent.
- Outbound AI voice calls fall under the TCPA after the FCC's February 2024 ruling; healthcare deployments need HIPAA BAAs with every vendor that processes or stores patient data.
- Start with your highest-volume, most repetitive call type, prove it on real calls, then scale.
Ready to put an AI voice agent on your phones? Our voice AI engineers will identify your highest-volume call type and give you a costed plan with latency targets. → Book a call
We build what you're reading about
Custom AI agents, voicebots and chatbots that cut costs, unlock growth, and deliver results you can see.
