Voicebot, voice agent or AI receptionist: which to choose and what does it cost?

Voicebot, voice agent, AI receptionist: what is the difference?

The names get used interchangeably, but they cover four different things.

Name	How it works	What it is good for
IVR	a recorded menu: "press 1, press 2"	routing a call to the right department
Voicebot (classic)	recognises speech and intent, follows a script written in advance	narrow, repetitive tasks: parcel status, appointment confirmation, a survey
Voice agent	a language model listens and speaks in real time and calls tools in your systems during the call	conversations you cannot draw as a tree: booking with follow-up questions, quoting, a caller who changes their mind mid-sentence
AI receptionist	a voice agent sold as a finished product: number, calendar, dashboard, recordings, transcripts	a small business that wants its phone answered from tomorrow

Two caveats. Vendors call both scripted systems and LLM-based agents "voicebots", so ask what is under the hood and what happens when a caller leaves the script. And "virtual receptionist" also describes services where humans answer the phone. This guide is about the AI kind.

The rest of the guide covers voice agents in two variants: bought as a product and built yourself. A classic scripted voicebot mostly makes sense where one already runs and does a single job.

What is a voice agent good at, and what is it not?

It works for calls that have a goal and end in an operation in a system:

• booking, moving and cancelling appointments at a clinic, garage or salon,
• answering the phone after hours and at peak times when the front desk cannot keep up,
• questions about order status, price, availability and opening hours,
• first-pass qualification of an enquiry and writing the lead to the CRM,
• taking a service request with a customer number and a description of the problem.

It works worse where the conversation itself is the value: an upset customer with a complaint, a negotiation, a medical question that needs judgement, a sale that depends on rapport. There the agent should collect the details and hand the call to a person rather than try to finish it.

The simplest test: if you can state what data the agent must collect and which operation it performs at the end, it is a call for an agent. If you cannot, fix the process first.

What does an off-the-shelf AI receptionist cost?

ElevenLabs Reception.ai has the best-documented price list. As of 17 September 2026, billed monthly:

Plan	Monthly price	Minutes included	Each extra minute
Basic	$29	75	$0.45
Plus	$79	275	$0.38
Premium	$199	1,000	$0.30

Billed annually it comes to $24, $66 and $166 a month. Moving up a plan pays off from about 186 minutes (Basic to Plus) and about 591 minutes (Plus to Premium), so the cost climbs in steps.

The price covers the whole product: call handling, a booking calendar and booking page, Google Calendar, a simple CRM with contact history, a knowledge base, recordings, transcripts, call summaries and analytics. You are paying for not having to build anything, not for a minute of model time.

Three things to check in a trial before planning a rollout outside the US:

• Phone numbers. According to the product page, numbers are currently provisioned only for US businesses; other countries are marked "coming soon".
• Which plan unlocks webhooks and MCP. The documentation contradicts itself: the webhooks page says Basic or higher, the integrations overview says Basic gets Google Calendar only, and the product page lists webhooks under Plus.
• Your language and GDPR. The page promises 70+ languages without listing them and cites ISO 27001 and ISO 42001, but says nothing about GDPR or EU data residency.

For a local reference point: Polish off-the-shelf services with public pricing cost roughly 900 to 3,000 PLN a month in September 2026, with bundles of 500 to 4,000 minutes, setup fees from zero to 5,000 PLN and overage between 0.40 and 1.09 PLN per minute. They cost more than Reception.ai, but the price includes a Polish number, onboarding in Polish and a contract with a local company.

What does a custom voice agent cost?

A custom agent has three line items: the voice model, telephony and a small server. All prices from official price lists, as of 17 September 2026.

Voice model. OpenAI's gpt-live-1 costs $0.05 per minute of session, billed per second. Backend tool calls and extra reasoning are billed separately at the rate of whichever model handles them. Google's Gemini 3.8 Live prices audio separately in each direction: $0.005 per minute of listening and $0.018 per minute of speaking. If the agent talks for about 40% of the call, that is roughly $0.012 per minute.

Telephony. Twilio in Poland: a mobile number at $4 a month, inbound calls at $0.0107 per minute, SIP interface $0.0040, Media Streams $0.0044, recording $0.0025. A geographic number requires an address inside the numbering zone plus company registration details; a mobile number has no such restriction.

One minute, all in:

Component	GPT-Live	Gemini 3.8 Live
voice model	$0.0500	about $0.0122
Twilio inbound	$0.0107	$0.0107
SIP or Media Streams	$0.0040	$0.0044
total per minute	about $0.065	about $0.027

Add about $14 a month in fixed fees: $4 for the number and around $10 for a small server. These are conservative assumptions. With Elastic SIP Trunking instead of Programmable Voice, a GPT-Live minute drops to about $0.057.

The architectural difference matters for the build cost. OpenAI accepts SIP calls natively, so audio goes from the carrier straight to the model and never crosses your server. Gemini Live speaks WebSocket, so you need your own audio bridge between Twilio and the model, and then you host it, monitor it and secure it.

What does it add up to per month at different volumes?

Assumption: the average call lasts four minutes. For Reception.ai each row uses the cheapest plan plus overage; for a custom agent, the per-minute rate plus $14 in fixed fees.

Calls per month	Minutes	Reception.ai	Custom agent, GPT-Live	Custom agent, Gemini
25	100	$40	$20	$17
50	200	$79	$27	$19
100	400	$127	$40	$25
150	600	$199	$53	$30
250	1,000	$199	$79	$41
500	2,000	$499	$143	$69
750	3,000	$799	$208	$96
1,250	5,000	$1,399	$338	$151

This is not a like-for-like comparison. Reception.ai includes the dashboard, calendar, CRM, recordings and analytics. The custom columns are running costs only; build and maintenance come on top.

When does a custom agent pay back?

The question is not which model is cheaper per minute. It is at what call volume the saving covers the build. A minimal custom agent on GPT-Live (SIP, sessions, a tool gateway, logging, error handling) takes me one to two days. A multi-tenant version with monitoring and a dashboard is a bigger project. For comparison, wiring n8n into Reception.ai takes a few hours, because you only build the workflows and the tool contracts.

Monthly saving of a custom GPT-Live agent against Reception.ai, and how long the build takes to pay back:

Calls per month	Monthly saving	Payback on a $2,000 build	Payback on a $5,000 build
50	$52	38 months	96 months
100	$87	23 months	58 months
150	$146	14 months	34 months
250	$120	17 months	42 months
500	$356	6 months	14 months
750	$591	3 months	8 months
1,250	$1,062	2 months	5 months

The 250-call row looks worse than the 150-call row because at 1,000 minutes Reception.ai still costs $199: that is the sweet spot of the Premium plan. Past 1,000 minutes every extra minute costs $0.30 and the custom agent pulls away fast.

The thresholds I use for a single business:

• Below about 50 calls a month: off-the-shelf. The saving is too small to justify your own infrastructure.
• 50 to 150: usually still off-the-shelf, unless you need unusual integrations or the product sounds poor in your language.
• 150 to 500: a custom agent starts to make sense. It comes down to the build cost, how long you will run it, and whether the core will serve other businesses or branches.
• Above 500: custom. At 2,000 minutes the gap exceeds $350 a month.

The maths changes when one core serves many businesses or locations. Ten clinics at 50 calls a month each cost 10 × $79 = $790 on the off-the-shelf service and about $270 on a custom agent. A $2,000 build then pays back in under four months, even though no single clinic crosses the threshold on its own.

When is a voice agent worth it at all?

The costs above are small next to a salary. But an agent rarely replaces a salary. It usually picks up what nobody picks up today: calls after hours, during breaks, at peak, while the receptionist is talking to a patient at the desk.

So I run the numbers from missed calls. Take last month's missed-call count from your phone bill. Estimate what share of them were new customers who never called back. Multiply by the average value of a first visit or order. If the result is several times the agent's monthly cost, the decision is easy. If the business already answers nearly everything and gets a dozen calls a month, there is nothing for a voice agent to recover.

An agent also does not pay off when the result of the call has nowhere to go. Without an online calendar, a CRM or at least a spreadsheet that receives the booking, you have an answering machine with a nice voice.

How do you connect a voice agent to n8n?

n8n does not touch the audio and should not. The voice model runs the real-time conversation; n8n is the layer where the business logic lives: CRM, calendar, price list, ERP. During the call the agent invokes a tool, n8n performs the operation and returns the result.

Caller
  ↓
voice agent (Reception.ai, GPT-Live or Gemini Live)
  ↓  tool call
n8n webhook
  ↓
CRM / calendar / price list / ERP
  ↓  short JSON
voice agent reads the answer aloud

An example from a garage. The caller says: "I'd like to book a brake pad replacement for a BMW 320i." The agent collects the registration number and calls a tool:

{
  "registration": "WA12345",
  "service": "brake_pads"
}

The n8n workflow finds the vehicle in the CRM, checks the price list and free slots, and returns:

{
  "vehicle": "BMW 320i",
  "price_from": 650,
  "slots": ["2026-09-18T15:30", "2026-09-19T09:00"]
}

The agent replies: "For that BMW the price starts at 650 zloty. I have tomorrow at 3:30 pm or Saturday at 9 am."

In Reception.ai you define a tool in the dashboard: name, description, URL, HTTP method, parameters and a response timeout. The description is part of the prompt, because the agent uses it to decide when to call the tool. I cover how to write those descriptions in the guide to the n8n AI Agent. In a custom agent the same tools are declared as model functions, and a small gateway forwards the calls to n8n webhooks.

A typical front-desk toolset: find_customer, check_availability, create_booking, reschedule_booking, cancel_booking, calculate_quote, create_lead, send_confirmation, transfer_call.

Four rules that prevent most of the trouble:

1. Provider-agnostic tools. I design the input and output contract so that it does not depend on the voice model. Today GPT-Live is the best in Polish; in six months that may change. Then I swap the adapter and the n8n workflows stay as they are. The same toolset also serves Reception.ai at a smaller business.
2. Fast, short responses. The caller is waiting in silence, so the workflow answers through Respond to Webhook as soon as it has the result, and sends the SMS or writes to the spreadsheet afterwards. The JSON holds only what the agent needs to say. Gemini 3.8 Live calls tools asynchronously by default and can keep talking meanwhile ("I have the car details, just checking the slots"), which suits longer workflows.
3. An error is also an answer. When the CRM is down or there are no slots, the workflow returns a readable message and the agent offers a callback or transfers to a person. In n8n that means error outputs on nodes and an error workflow with an alert. n8n-mcp data shows that only 2% of workflows built by AI agents have error handling, and on a phone call you hear its absence immediately.
4. A secured webhook. A token header on every tool and no public unauthenticated URLs. A cancel_booking tool open to anyone who knows the URL is asking for trouble. I go deeper into securing webhooks in the guide on n8n self-hosted vs cloud.

Webhooks or MCP?

Reception.ai supports both: tools as webhooks, and an MCP server over SSE or Streamable HTTP, with a choice of which tools the receptionist may call. n8n can be that MCP side through its MCP Server Trigger node.

For a first deployment I pick webhooks. They are simpler, predictable and easy to debug: one tool, one workflow, one execution history. MCP earns its place when the tool catalogue grows and has to serve several agents at once, say the phone receptionist, the website chat and an internal assistant.

Which model speaks Polish best?

In September 2026 I tested all three live. It was a quick test, not a formal benchmark, but for a receptionist a natural conversation in the caller's language is the whole product, so I take it seriously. If your callers speak another language, run the same test in that language.

• GPT-Live: the best. The most natural voice, a coherent conversation, never switched language.
• Reception.ai: fine, but in the public demo the voice changed to a different one mid-call. A demo may be configured differently from your own account, so check it in a trial.
• Gemini 3.8 Live: the weakest. It started in Polish and then switched to German by itself. A strict prompt instruction may fix it ("speak Polish only; foreign proper names are not a request to change language"), but for a production receptionist it is a red flag.

Gemini is clearly cheaper, but the gap is small in absolute terms: about $37 a month at 1,000 minutes and about $187 at 5,000. One customer who hangs up because the front desk started speaking German costs more. I do not pick a voice model on price per minute.

Before going live, run 10 to 30 identical scenarios on each candidate: interrupting the agent, changing your mind mid-sentence, phone and registration numbers, local surnames, dates and times, background noise, a long pause, two or three tool calls in one conversation, a tool error, no availability, handoff to a person. Score whether the agent keeps its language and voice, how long responses take, whether the data in the system is correct, and how many calls needed a human.

What about GDPR and disclosing that it is an AI?

Two obligations, both handled in the greeting.

AI Act, Article 50. From 2 August 2026, a person interacting with an AI system must be told so clearly and at the latest at the start of the interaction, unless it is obvious from context. A line in the terms of service or a notice after the call does not count. Source: Regulation (EU) 2024/1689.

Recording. Recording a customer call is processing of personal data. The caller should hear at the start that the call is recorded, who the controller is and why the recording is made, with the full privacy notice available on your website. Add a data processing agreement with the agent vendor and find out where recordings and transcripts physically end up.

One sentence does it: "Hello, this is the clinic's virtual assistant, this call is recorded." This describes practice, not legal advice. With medical or financial data, talk to a lawyer before choosing a vendor, not after.

Which to choose: a decision tree

1. Can you state what data the agent must collect and which operation it performs? No: fix the process first. An agent will not repair it.
2. Fewer than about 50 calls a month and you want to start fast? Yes: an off-the-shelf AI receptionist, with n8n connected through webhooks if you need your own logic. Check phone numbers and language quality in a trial first.
3. More than 500 calls a month, or many locations on one core? Yes: a custom agent on GPT-Live with native SIP, tools in n8n.
4. In between? Work out the payback from the table above for your build cost and decide whether conversation quality in your language is worth more to you than a ready dashboard.
5. Gemini 3.8 Live is worth watching. It is the cheapest and has good asynchronous tool calling, but it came last in my Polish test and needs an extra audio bridge.

Whichever variant you pick, the business logic stays in n8n and the voice model is a replaceable part. That way choosing option 2 today does not block the move to option 3 when the phone gets busier.

Sources

Prices and parameters checked on 17 September 2026 on vendor pages:

• Reception.ai: plans and pricing, webhooks, MCP server, integrations overview, product page
• OpenAI: gpt-live-1 model, Live API
• Google: Gemini API pricing, Gemini 3.8 Live, Live API capabilities
• Twilio: voice pricing in Poland, Elastic SIP Trunking in Poland, regulatory requirements for Polish numbers
• AI Act, Regulation (EU) 2024/1689

Dollar amounts exclude VAT. Model and telephony prices change often, so check the current price lists before you decide.