AI Voice Agents: Receptionists, Phone Workflows and Helpdesk Voice

A voice agent is a phone system that listens, decides, and speaks. The decision that shapes a project is which calls the agent is allowed to finish, and which calls must reach a person. Three jobs get sold under the same name, and they do not share a shortlist.

The front desk. After-hours bookings, opening hours, new-patient intake, and a warm transfer when the caller asks for a person. That is a receptionist product. The full comparison, including handoff and recovery paths, is in the AI receptionist guide.

The custom phone stack. You want your own prompt, your own telephony, a calendar integration, and a transfer that fires when the caller interrupts. That is a platform build, covered in the AI phone agents guide.

The queue you already run. If your team already works in a helpdesk such as Zendesk, Gorgias, or Chatbase, voice is either included in the plan you pay for or sold as an add-on to it. Seat prices and resolution units are compared in the help desk comparison.

These products overlap in technology but differ in ownership, pricing, and failure modes. Choose the job before the vendor.

Watch a receptionist demo that shows what a human takeover looks like. The useful minute is the handoff, not the greeting:

How a voice agent works, end to end

Every voice agent is a pipeline. Knowing the parts explains the failure modes.

1. Telephony. The call arrives through a phone number, a VoIP trunk, or a carrier such as Twilio or Telnyx. This leg costs per minute on platform builds and is bundled into receptionist and helpdesk products.
2. Speech recognition (STT). The caller's audio becomes text. Accuracy drops with background noise, accents, and weak mobile connections.
3. Reasoning. The language model decides what to say next, using the transcript, your knowledge base, and any connected tools such as a calendar or order lookup.
4. Synthesis (TTS). The reply becomes speech. Quality ranges from robotic to near-human.
5. Latency. Every stage adds delay. As total delay approaches roughly half a second, callers are more likely to talk over the agent, so measure barge-in behavior on your own calls rather than relying on a fixed threshold.
6. Barge-in. The agent must detect that the caller is speaking and stop. If it does not, callers hang up.
7. Transfer. When the agent gives up, it needs a destination and a way to pass context to a human.

Usage-based pricing reflects this stack. Retell publishes "$0.07-$0.31 / min for AI Voice Agents" depending on the models you choose. Vapi breaks the same costs into a hosting fee plus provider costs for transcription, the language model, and the voice. On a platform build, the invoice is a sum of parts.

Three jobs, three cost models

Vendor pricing reveals which category a product belongs to.

Receptionist tools: per-customer or flat monthly

Receptionist products price for the job, not the compute. Goodcall charges "$79 a month" for a solo operator, covering "100 unique customers a month then 79¢ each," with no per-minute meter. Flat-rate AI receptionists remove the meter entirely, unlike live answering services, which Goodcall estimates at $1.50 to $1.90 a minute. Staffing-style receptionist services often quote per-minute or per-call bundles that include human backup. The recurring cost there is the transfer, not the greeting.

Phone platforms: per minute, per component

Platforms like Vapi and Retell charge for exactly what runs. Vapi's free tier starts with "$5 free credits" and "4 concurrent calls", then hosting runs "$0.05/min" with models passing through at cost. Retell warns about a subtle billing rule: "Billing covers the entire duration of the call because the speech-to-text engine remains active and listening throughout, even during silence." Hold time bills like talk time.

The platform model suits teams that will tune the prompt, build integrations, and accept operational ownership. It is a poor fit for someone who just wants the phone answered.

Helpdesk voice: included, or sold as an add-on

If you already pay for a helpdesk, voice is a feature, not necessarily a new vendor. Zendesk includes telephony in its Suite Team plan, adds an IVR phone menu at Suite Professional, and lists Generative AI for voice at Suite Enterprise. Gorgias sells voice as an add-on billed per voice ticket, with tiers that scale with call volume. Chatbase includes Voice on Basic and adds telephony with concurrent call capacity at Business.

When the caller's problem is the same as the ticket problem, this is the cheapest path: the voice call becomes a ticket in the system you already run. It is the wrong path when you want a polished standalone phone experience or deep custom logic, because you would be paying for helpdesk features you may not use.

Some helpdesks also attach an outcome-based AI model to voice. Gorgias, for example, prices its AI Agent by resolved interaction: you pay only when the AI completes a conversation on its own. That shifts cost from seats to successful completions, but it requires you to define what a resolution means before you can forecast it. Ask for a current quote for your mix, because the interactive pricing page does not display per-tier rates.

A booking call, step by step

Here is a hypothetical afternoon call to a dental clinic using a receptionist agent. The scenario is illustrative, not a measured deployment.

Caller: "I'd like to book a cleaning."

Agent: "Sure. Can I have the last name on the account?"

Caller: "Meyer."

Agent: "And a phone number to confirm the appointment?"

Caller: "555-0142."

Name and phone are lookup keys, not identity proof. The agent uses them to find the right record, then sends a secure verification link to the phone number on file. The caller taps it, proving they control that contact, and only then can the agent act on the account. Collecting passwords or codes in the call would be a red flag; verification should happen on a secure screen outside the conversation.

Agent: "Thanks, Ms. Meyer. I've found your record. I've sent a verification link to the phone number on file. Please tap it to confirm it's you."

Caller: (completes verification through the secure link)

Agent: "Thank you. I have Thursday at 10 a.m. or Friday at 3 p.m."

Caller: "I wanted Tuesday."

Agent: "Let me check Tuesday." Pause. "Tuesday is fully booked. Would Thursday at 10 work?"

Caller: "Okay."

Agent: "Great, I'll book Thursday at 10 for you. You'll get a confirmation to the number on file. Anything else?"

Caller: "What will my copay be?"

Agent: "I don't have payment details for that. One moment, I'll get someone who can help."

The agent transfers to a staffed receptionist, passing a summary: record verified by secure link, Thursday 10 a.m. booked, copay question unanswered. The human picks up with that context already on screen. This call leans on the two features that matter most: booking only from a live calendar, and a transfer that carries context.

Notice what the agent did not do. It did not invent a Tuesday slot, it did not answer a question outside its scope, and it did not treat the phone number as proof of identity. It flagged a gap and escalated.

Failures that show up in the first ten recordings

Play back the first ten real calls. Four failures appear before any dashboard does.

No barge-in. The caller talks over the agent and the agent finishes its sentence anyway. Interruption handling is the difference between natural and unbearable. If the transcript shows overlapping speech where the agent kept going, the configuration is not ready.

Transfer to nowhere. The word "representative" does not reach a person. Write down the transfer destination, the hours it rings, and the voicemail path before go-live. A transfer that rings a closed office is a missed call with extra steps.

Confident invention. The agent states a price or a slot that is not in the calendar or the rate card. Constrain it to those two sources. A confident wrong answer is the failure customers remember, and it erodes trust faster than a human mistake.

The hard-to-undo booking. A wrong booking is difficult to reverse. The receptionist guide covers recovery rules: confirm the slot aloud, write the reservation only after confirmation, and give callers a way to correct a misheard date.

For helpdesk voice, the equivalent failures differ: the transcription is wrong, the ticket is misrouted, or the agent resolves a call without ever creating a ticket. Check that voice calls land in the right inbox with the right tags.

Outbound is a different project

Dialing a list adds consent, calling-hour rules, and do-not-call obligations. Those are not settings you discover after the first campaign. If the job is qualifying a list, read the AI SDR guide and keep the phone vendor out of the decision until the list and script are legal to run. An inbound receptionist and an outbound dialer can share a voice; they should not share a project plan.

A pilot with one number

Give the agent one number, one job, and ten business days. Count completed bookings or answered questions, transfers, and calls where the caller hung up during the greeting. Use the recordings, not the dashboard. If transfers are most of the volume, the agent is a filter, which can still be worth paying for when the human only hears the hard calls. Write that result down before adding a second number. A second number waits until the first has a named owner who can turn the agent off without taking the phone line down.

FAQ

What is the difference between an AI receptionist and a voice platform?

An AI receptionist is a finished product for booking, intake, and handoff. A voice platform gives you components, telephony, speech recognition, a model, and a voice, which you assemble into a custom agent. Receptionists ship fast and cost less to operate; platforms offer control and require ownership.

Can one voice agent handle both booking and support?

Technically yes, but the jobs have different data sources and escalation rules. Booking needs live calendar access and a secure way to verify the caller. Support needs order or ticket history. Combining them risks agents answering support questions with booking knowledge. Start with one job.

How much does an AI voice agent cost?

From about $0.07 to $0.31 per minute on usage-based platforms like Retell, with Vapi charging a $0.05 per minute hosting fee plus model costs. Receptionist products often use flat monthly rates, from about $79 a month for low-volume use. Helpdesk voice is usually bundled into seats or billed per ticket. Ask for both the agent minute and the transfer cost before comparing.

What happens when the AI cannot answer?

It should transfer to a human with context. Test that the transfer rings the right destination during your hours. A transfer to voicemail after hours may be acceptable, but it is a designed choice, not a default.

Do I need to replace my helpdesk to add voice?

No. Several helpdesks, including Zendesk and Gorgias, offer voice as a native or add-on feature. Compare the cost of the add-on against a standalone receptionist for your call volume. If your calls are mostly booking and simple intake, a receptionist may win. If the call is a ticket, the helpdesk add-on keeps one inbox.

What counts as a completed call in the contract?

Read the vendor's definitions. Some bill per minute, some per unique customer, some per voice ticket, and some per resolved outcome. Retell, for example, stops its agent fee after transfer but keeps charging telephony. Silence and hold time bill as talk time on usage platforms. These definitions determine whether your estimate is accurate.

Conclusion

The right voice agent is the one matched to the call job. Bookings and intake point to a receptionist product. Deep customization points to a platform, with an owner who maintains it. Calls that are tickets point to your existing helpdesk's voice add-on. Price the whole operation: the agent minutes, the transfers, and the human who still answers the calls the agent cannot finish. Run a ten-day pilot on one number, listen to the recordings, and decide based on what actually happens.