What a customer service voice agent actually does
Answers an inbound support call, retrieves the caller's account and order data through tool calls, and resolves the issue on the line or warm-transfers it to a human with the full context attached.
A customer service voice agent sits on your support line and runs the same loop a chat agent does, with one layer added: real-time speech. It transcribes the caller as they talk, understands the intent, pulls account or order data through a tool call against your CRM or billing system, and either resolves the issue on the call or warm-transfers it with a spoken summary so the human picking up doesn't make the caller repeat everything. The gap between a text support agent and a voice one is latency: every extra 300-500ms of dead air between the caller finishing a sentence and the agent replying reads as an awkward pause on a phone call in a way it never does in chat.
Voice-first support consolidated around outcome pricing in 2026, the same shift chat-based agents went through a year earlier. Fin now handles chat, email and voice from a single $0.99-per-outcome agent, and Sierra raised a $950M round in May 2026 at roughly a $15.8B valuation on close to $200M in reported ARR, almost entirely from enterprise contact centers replacing IVR trees and offshore call-center seats with outcome-priced voice agents.
Where it still breaks: a genuinely upset caller who needs a human's tone rather than a script, calls that require judgment outside the documented policy, and the authentication problem. Voice deepfakes are a real enough threat that FINRA's 2026 oversight report specifically flags caller-authentication risk for voice agents handling account changes, and a support agent that skips strong identity verification before touching an account is one cloned voice away from a costly mistake.
AI capabilities involved
Real-time speech-to-text on the live call
Understanding intent and holding the conversation
Natural-sounding voice output (text-to-speech)
Account and order lookups mid-call via tool calls
Who uses this
- Support and contact-center teams fielding 500-5,000 inbound calls a month on billing, order status or account questions
- Businesses already paying for an answering service or offshore call-center seats and comparing the per-minute math
- Companies whose call volume spikes seasonally and need overflow coverage without hiring temp staff
- Teams running a text support agent already and extending the same resolution logic to the phone channel
What the customer service voice agent handles (and what it shouldn't)
| Task | Agent handles it? | Systems it touches | Human role |
|---|---|---|---|
| Order status and shipment lookups by phone | Fully | Order system, ShipStation/AfterShip | None |
| Account and billing questions | Fully | CRM, billing platform | None unless the charge is disputed |
| Password or account-access resets by phone | With approval | Auth provider | Confirms caller identity first |
| Refunds and returns under a set dollar threshold | With approval | Order system, payment processor | Approves anything over the threshold |
| Caller identity verification before account changes | Fully | Auth provider, CRM | Steps in if verification fails |
| Angry or already-escalated callers | No | Call transfer | Takes the call over immediately |
| Complex multi-account or enterprise billing disputes | No | CRM | Handles the full case |
| After-hours and overflow call coverage | Fully | Telephony, calendar | Reviews flagged calls the next business day |
Off-the-shelf AI agents on the market
Real products you can sign up for today — with current 2026 pricing, honest pros and cons.
Fin (formerly Intercom), now owned by Salesforce
Teams already on Intercom or Zendesk adding a voice channel
14-day trial, no card required
$0.99 per resolved outcome across chat, email and voice; a 50-outcome monthly minimum applies when running standalone on another helpdesk
Custom outcome-volume discounts for larger teams
Pros
- +One agent handles chat, email and voice with the same knowledge base
- +Works standalone on top of Zendesk or Salesforce, not just Intercom
- +Backed by Salesforce's resources since the September 2026 acquisition closed
Cons
- −"Outcome" is defined broadly, so the vendor's claimed 67% resolution rate may not match your real numbers
- −Independent teardowns report 45-53% resolution in production
Sierra
Large contact centers replacing legacy IVR and offshore seats
None public
Contact sales; third parties report starting contracts around $150K/yr with the exact per-resolution rate negotiated privately
Custom, scoped to call volume and complexity
Pros
- +Purpose-built "Agent OS" for enterprise voice and chat, including back-office actions
- +Escalations to a human are free under its outcome-based model
- +Raised at a $15.8B valuation on close to $200M ARR in May 2026, signaling real enterprise traction
Cons
- −No self-serve pricing or trial, so evaluating it requires a sales process
- −Deployment timelines run months, not weeks, for enterprise contact centers
Decagon
Mid-to-large support organizations past the SaaS self-serve tier
None
Contact sales; third parties report roughly a $50K/yr platform fee plus about $0.99 per conversation
Custom, enterprise-only
Pros
- +"Agent Operating Procedures" let you write escalation and resolution logic in plain language instead of a workflow builder
- +Chat, email and voice from one platform
- +Reached a $4.5B valuation on a $250M Series D in January 2026
Cons
- −No public pricing page; the platform fee alone puts it out of reach for small teams
- −Enterprise-only sales process with no self-serve trial
Dialzara
Very small support teams with light call volume and simple resolution needs
No free tier
Lite $29/mo (60 min); Pro $99/mo (220 min); Plus $199/mo (500 min); Elite $349/mo (1,000 min)
Done-for-you tier from $1,500+/mo
Pros
- +Cheapest entry point of any vendor on this page, by a wide margin
- +Supports English and Spanish out of the box
- +Outbound calling add-on available from $750/mo
Cons
- −Overage runs $0.48 down to $0.35/minute once you exceed the included minutes
- −No native CRM or order-system integrations beyond Zapier
The AI stack
A production voice support agent needs a real-time voice layer, telephony to actually carry the call, a reasoning model, and observability tuned for the extra failure modes voice adds on top of text.
Voice (speech-to-text + text-to-speech)
Converts the caller's speech to text for the model and the model's reply back to natural speech
ElevenLabs Agents
$0.08/min on every self-serve plan; $0.16/min burst over the concurrency capVoice quality as the top priority
Retell AI (bundled)
Voice engine $0.055/min plus LLM, TTS and telephony, roughly $0.15/min all-inTeams that want one predictable bill
OpenAI Realtime (gpt-realtime-2.1)
$32/M audio-in, $64/M audio-out tokens, roughly $0.04-$0.10/min in practiceLowest-latency custom builds
Our pick: Retell for a single predictable bill, or a chained ElevenLabs plus Claude Sonnet 5.5 stack when you want to swap components independently.
Telephony
Carries the phone call and connects it to your voice pipeline
Retell's built-in telephony
Roughly $0.015/min included in the all-in rateFast production builds
Twilio direct
$0.008/min inbound, $0.014/min outboundFully custom builds
Our pick: Let your orchestration platform (Retell or Vapi) handle telephony pass-through rather than managing Twilio directly, unless you already run Twilio for other systems.
Reasoning model
Understands the caller's intent, decides when to resolve versus transfer, and calls tools for account lookups
Claude Sonnet 5.5
$2/M input, $10/M outputThe main call-handling loop
Claude Haiku 5.5
$0.10/M input tokens (up to 100K), $0.50/M outputStraightforward order-status and FAQ calls
Our pick: Sonnet 5.5 for the main resolution loop, with Haiku 5.5 handling simple status checks to cut per-call cost.
Observability & evals
Reviews call recordings and transcripts for missed verifications, bad transfers and policy drift
Vendor call logs and transcripts
Included on most SaaS tiersBuy-SaaS
Langfuse
Free up to 50K units/mo, then $29+/moCustom builds
Our pick: Review a random sample of real call recordings weekly, with extra attention to any call that touched account verification or a refund.
Reference architecture
The hardest engineering problem is verifying who's actually on the line before the agent touches an account: voice deepfakes make a caller's voice alone an unreliable identity signal, so the architecture needs a second verification factor (an account PIN, a one-time code) before any account change, not just confident-sounding speech.
Call connects
Telephony (Twilio/Retell-native)The caller's phone call routes to the voice agent's number
Speech-to-text
STT engineThe caller's audio is transcribed to text in real time as they speak
Identity check
Verification stepFor anything touching the account, the agent asks for a PIN, order number or one-time code before proceeding
Intent and reasoning
Reasoning modelThe model reads the transcript, decides the intent, and calls a tool for account or order data
Tool call
CRM/billing integrationAn order lookup, refund check or account update runs against your real system
Text-to-speech
TTS engineThe model's reply is converted to natural speech and streamed to the caller
Transfer or logging
Call-transfer logic, observability toolUnresolved or flagged calls warm-transfer to a human; every call is logged for weekly review
Estimated cost per request
~$0.11-$0.18 per minute of call, including voice, model and telephony costs before any vendor per-outcome fee
Cost calculator
Drag the sliders to model your actual usage. The numbers update in real time so you can stress-test economics before writing a single line of code.
Estimate your monthly cost at your own call volume and average call length.
Estimated monthly cost
$1,610
≈ $19.3k per year
Calculator notes
- The Fin-style line assumes close to a 1-to-1 outcome rate; move the average-call-length slider to see how longer calls change the DIY line, since Fin bills per outcome regardless of call length.
- The DIY line assumes a 4-minute average call on an all-in Retell/ElevenLabs/telephony stack at roughly $0.15/min.
- A custom agency build replaces both lines with your own voice-stack bill, typically $100-$500/mo at this volume, plus the upfront build cost.
- Sierra and Decagon's negotiated enterprise rates aren't modeled here; they only make financial sense well above this calculator's call-volume range.
Build it yourself with vibe-coding tools
Retell lets you assemble a working customer service voice agent from a phone number, a system prompt and a CRM tool without stitching together separate STT, LLM and TTS vendors yourself.
Time to MVP
A few days for a basic version, 2-4 weeks to get verification, transfers and consent language production-ready
Total cost to MVP
$0-$100 (Retell's pay-as-you-go plan with $10 free credit, plus usage)
You'll need
Starter prompt
Create a new Agent in Retell using Claude Sonnet 5.5 as the model, Retell's bundled voice engine for STT and TTS, and your carrier (Retell-native or Twilio) for telephony. Set the greeting to: 'Thanks for calling {Business Name} support, this is an AI assistant, how can I help?' so disclosure happens before anything else is said. Give the agent this system prompt: 'You are a customer service voice agent for {Business Name}. Greet the caller and identify their intent: order status, billing question, account change, or something else. Before looking up or changing any account, ask the caller to confirm their order number or account email, and treat that as the minimum identity check; never make an account change based on voice alone. Use the order-lookup tool for shipment and order questions, and the account-lookup tool for billing questions. If the caller requests a refund over $50, sounds upset, or asks for something outside what you can do, say: "Let me connect you with someone who can help with that right away," and use the transfer_call tool with a one-sentence spoken summary for whoever picks up. Never claim to be human if asked directly. This call is recorded for quality purposes, consistent with the greeting already given.' Add three tools: order_lookup and account_lookup pointed at your systems' APIs, and transfer_call pointed at your human support line. Turn on call recording, and log every call's transcript and outcome to a database or sheet for weekly review.
Paste this into Retell AI
Follow-up prompts (run in order)
- 1
Add a one-time-code verification step before any account change, sent by SMS during the call
- 2
Add a sentiment-detection step that escalates anything angry regardless of confidence
- 3
Add a daily digest summarizing every call the agent transferred and why
- 4
Add DTMF keypad capture so a caller can pay a deposit without speaking a card number aloud
Expected output
A phone number that answers support calls within one ring, resolves order-status and simple billing questions on the line, and warm-transfers anything it can't verify or handle with a spoken summary attached.
Known gotchas
- !Latency stacks up fast across STT, the model, TTS and telephony; callers notice anything over about half a second of dead air
- !Vendor-advertised latency figures are often optimistic; independent tests of similar platforms have measured 600-1,050ms where the vendor claimed sub-500ms
- !Skipping identity verification before account changes is the single biggest security gap in a DIY voice agent, and voice alone isn't a safe identity check
- !Forgetting the AI-disclosure line at call start is both a compliance problem and the most common caller complaint about these systems
- !Background noise on real phone lines degrades transcription accuracy far more than it does in a quiet demo
- !HIPAA or PCI-regulated use cases need a paid BAA add-on on most voice platforms, which changes the DIY cost math substantially
Compliance & risk reality check
A phone-based support agent carries more compliance weight than a chat agent because calls trigger recording-consent and telecom-specific rules on top of ordinary AI-disclosure rules.
TCPA and FCC AI-voice rules
The FCC's February 2024 Declaratory Ruling (24-17) holds that AI-generated voices count as an "artificial voice" under the TCPA, requiring prior consent for outbound calls; this mainly hits any outbound follow-up or callback the agent places rather than inbound answering.
Mitigation: If the agent ever calls out (a callback, a follow-up), capture written consent naming AI-generated calls and disclose the AI at the start of every outbound call.
Call recording and wiretap law
Ambriz v. Google and Taylor v. ConverseNow both survived motions to dismiss in 2025 on the theory that an AI vendor capable of using call recordings to train its own models can be an unlawful third-party eavesdropper, even without proving it actually happened.
Mitigation: Play a recording-and-AI notice at the very start of every call, and put a no-training clause in every vendor contract covering your call data.
Caller authentication and voice-deepfake risk
A caller's voice alone is not a reliable identity signal; FINRA's 2026 oversight report specifically flags authentication risk for voice agents handling account changes.
Mitigation: Require a second factor (account PIN, order number, one-time SMS code) before the agent makes any account change, never on voice alone.
AI disclosure on the call itself
California's BOT Act, Utah's AI Policy Act and the EU's Article 50 transparency rule (in force from August 2, 2026) all push toward telling callers they're talking to AI.
Mitigation: State "this is an AI assistant" in the first line of the greeting, and never claim to be human if a caller asks directly.
PCI DSS if the agent takes payment details by phone
If a caller speaks a card number to pay a balance, the recording, transcript and underlying model calls all enter cardholder-data scope under PCI DSS 4.0.1, fully mandatory since March 31, 2025.
Mitigation: Use DTMF keypad capture or a pay-by-SMS link so the card number never reaches the AI, the recording or the transcript.
Build vs buy: the real math
3-6 weeks for a RapidDev build; a few days for a basic Retell version
Custom build time
$13K-$25K for an agency build (up to $35K for multi-system or compliance-heavy work); $0-$100/mo for DIY
One-time investment
Roughly 1,500-2,000 calls a month at a 4-minute average, where Fin's per-outcome fee starts to exceed the amortized first-year cost of a custom build
Breakeven vs buying
At 1,000 calls a month, Fin's $0.99-per-outcome pricing runs close to $990/mo assuming most calls resolve, plus the Intercom or Zendesk seats you're likely already paying for. A $13K-$25K custom build amortized over a year costs $1,083-$2,083/mo in year one, similar on paper, but the custom agent's marginal cost per extra call is just voice-stack tokens (around $0.11-$0.18/min) instead of a flat per-outcome fee. Past about 2,000 calls a month the per-outcome fee starts to outpace the amortized build cost, and past several thousand calls a month you're squarely in Sierra and Decagon's enterprise territory, where negotiated contracts take over the math entirely.
Skip the DIY — RapidDev builds the production version
A Lovable MVP gets you a demo. Production needs auth that doesn't leak data, AI calls that don't bankrupt you, observability when models drift, and code you can audit. That's what we ship through our AI app development service.
Discovery call (free)
30 minWe map your exact customer service voice agent use case: who uses it, target volume, AI model choice, integrations, compliance scope. You get a detailed scope document and fixed-price quote within 48 hours.
AI-accelerated build
3-6 weeks for a RapidDev build; a few days for a basic Retell versionOur engineers use Claude Code, Lovable, and custom tooling to ship 3–5x faster than agencies. You see weekly progress in a staging environment — not a black box.
Launch + handoff
1 weekWe deploy to your infrastructure, transfer the GitHub repo, set up CI/CD and monitoring, and train your team. You own 100% of the source code, prompts, and model configurations.
What you get
Timeline
3-6 weeks for a RapidDev build; a few days for a basic Retell version
Investment
$13K-$25K for an agency build (up to $35K for multi-system or compliance-heavy work); $0-$100/mo for DIY
vs SaaS
ROI in Roughly 1,500-2,000 calls a month at a 4-minute average, where Fin's per-outcome fee starts to exceed the amortized first-year cost of a custom build
30-min call. Fixed-price quote within 48 hours. No commitment.
Frequently asked questions
How much does an AI voice agent for customer service cost?
SaaS pricing starts at $0.99 per outcome on Fin, with a 50-outcome monthly minimum, up to a negotiated enterprise rate on Sierra or Decagon. A custom RapidDev build runs $13K-$25K upfront (up to $35K for complex, multi-system work), plus $100-$500/mo in voice-stack costs. A DIY build on Retell costs roughly $0.15 per minute of call time.
How long does it take to set up?
Fin connected to an existing Intercom or Zendesk helpdesk can go live in 1-4 weeks. An enterprise contract with Sierra or Decagon takes 2-3 months to scope and deploy. A custom agency build takes 3-6 weeks, and a basic Retell version can work within a few days, though production-ready verification and transfer handling takes 2-4 weeks.
Can RapidDev build this for us?
Yes. RapidDev has shipped 600+ apps and offers a free 30-minute consultation to scope your call volume, systems and compliance needs before you commit to anything.
Will callers know they're talking to AI?
They should hear it in the first line of the greeting. Stating "this is an AI assistant" satisfies disclosure expectations under California's BOT Act, Utah's AI Policy Act and the EU's Article 50 transparency rule at once, and it heads off the awkwardness of a caller figuring it out mid-call.
What's a realistic resolution rate for a voice agent?
Vendors like Fin advertise around 67% resolution across channels, but independent teardowns of production voice deployments report figures closer to 45-53%. Plan staffing and your cost calculator around the lower, independently reported range.
Can it work with Zendesk, Intercom or Salesforce?
Yes. Fin runs standalone on Intercom, Zendesk or Salesforce helpdesks, and Sierra and Decagon both build custom integrations into whatever CRM or support platform an enterprise contact center already runs. A custom build can connect to any of these plus systems no SaaS vendor supports.
Is it safe to verify a caller's identity by voice alone?
No. Voice deepfakes are convincing enough that FINRA's 2026 oversight report specifically flags voice-only authentication as a risk, so any agent making account changes should require a second factor, such as an account PIN or a one-time SMS code, before acting on what the caller says.
Want the production version?
- Delivered in 3-6 weeks for a RapidDev build; a few days for a basic Retell version
- You own 100% of the code
- AI cost monitoring built in
30-min call. No commitment.
