AI Phone Agent 2026: What It Costs, How It Works and How to Roll It Out

In short: An AI phone agent answers calls, understands requests in natural speech and handles routine work on its own: answering questions, booking appointments, taking callback requests, transferring to a person. What it costs depends mostly on the pricing model. Developer platforms like Vapi and Retell bill in US dollars per minute, roughly $0.07 to $0.31 by their own figures, before your developers' time. Goodcall charges from $79 a month and counts callers instead of minutes. Synthflow lists enterprise contracts starting at $30,000 a year. EchoCall sells minute bundles from €75 for 250 minutes, which is €0.30 per minute, with no setup fee. And since 2 August 2026 the EU requires that anyone talking to an AI is told so, at the latest when the conversation starts.

The first version of this guide from June 2026 contained figures we could not back up, including a language count, a setup time, a share of calls resolved automatically and an overly broad statement about where data is processed. We rewrote the article in October 2026. Prices of other providers come from their pricing pages, retrieved on 6 October 2026. Sources are listed at the end.

What an AI phone agent is and what it is not

An AI phone agent is software that holds conversations on your phone number. You will also see the terms AI voice agent, AI call agent and AI receptionist. Unlike a phone menu with "press 1", it understands full sentences. Unlike voicemail, it answers, asks follow-up questions and gets something done. Typical tasks are:

• answering opening hours, prices, directions and other standard questions from your knowledge base
• checking free slots in your calendar and booking them directly
• taking down a request and passing it to your team as a callback
• recognising urgent calls and transferring them live to a person
• looking up data in your systems, such as the status of an order

An AI phone agent does not replace your team. It takes the calls that repeat and the calls that go unanswered today: in the evening, at the weekend, when every line is busy. Advice, complaints and anything that needs judgement stay with people. A good agent knows when to hand over. Our article on the AI receptionist shows what that looks like at the front desk of a practice or a small business.

How a call works under the hood

Whatever the provider, every call passes through the same four stations.

1. The call comes in. Customers either dial a number from the provider directly, or your existing number forwards the call, for example when the line is busy or nobody answers. Larger phone systems connect the agent through a SIP trunk.
2. The agent listens. Speech recognition turns what is said into text. What matters is not only whether it understands correctly but whether it notices when the caller has finished. Deepgram introduced its Flux model for this in October 2025 and calls it the first speech recognition model that knows when someone is actually done talking. [10]
3. The agent decides. A language model determines what happens next: answer, ask, check the calendar, transfer. To do that it calls tools connected to your systems.
4. The agent speaks. A synthetic voice reads out the answer. If the caller interrupts, it stops and the agent listens again.

This chain of speech recognition, language model and voice is called a cascade. Since 2025 a second design has taken hold: speech-to-speech models that understand audio and speak directly, without going through text. OpenAI made its Realtime API generally available on 28 August 2025, with a direct SIP connection for telephony. On 10 September 2026 it followed with GPT-Live 1, a model that listens and speaks at the same time and hands difficult tasks to a second model in the background. Google released Gemini 3.8 Live five days later. [7] [8]

Two things follow for you as a buyer. The building blocks are getting better and cheaper quickly. And they keep changing: OpenAI shut down the beta interface of its Realtime API on 12 May 2026, and anyone who had built on it had to rebuild. [7] With a developer platform you choose the models yourself and test again after every change. With a finished platform you leave that to the provider and only take care of the conversation: greeting, knowledge, rules, transfers.

What an AI phone agent costs: the line items

Every per-minute price is made of the same line items. Once you know them, you can compare offers that look unrelated at first glance.

What does that mean for one minute of conversation? Vapi's own calculator arrives at $82 to $129 for 1,000 minutes, so roughly $0.08 to $0.13 per minute, still without telephony. Retell lists a range of $0.07 to $0.31 per minute on its pricing page, depending on voice and language model. [1] [2] With a German Twilio number, add roughly one US cent per inbound minute.

Raw model prices deserve a second look. Google bills by audio length on both sides, OpenAI and xAI by conversation minute. What ends up on the invoice depends on how long each side talks and how much context the model rereads with every sentence. A per-minute price from a model price list is therefore never the price of your phone agent.

What the pricing page does not show

The largest costs of an AI phone agent appear in no per-minute table.

• Development. On developer platforms you usually connect calendar, CRM and transfers yourself through APIs and webhooks. At $0.10 per minute, one developer hour at €100 roughly equals a thousand minutes of conversation.
• Maintenance. Opening hours change, and so do prices. New model versions behave differently from old ones, and interfaces get shut down. Someone has to notice and follow up.
• Add-ons. Retell charges an extra $0.005 per minute for its knowledge base and $0.01 per minute for removing personal data, a number costs $2 and a verified number $10 per month. Vapi charges $2,000 per month for its HIPAA package. [1] [2]
• Currency. US providers bill in dollars. Exchange rate swings end up with you.
• Your own time. Writing down knowledge, defining rules, reading the first calls. This happens with every provider, including us. Plan a few hours for the first call type.

Four pricing models compared

Four ways of billing an AI phone agent have become established. None is better in principle. It depends on who builds the agent at your company and what your calls look like.

Developer platform per minute. Vapi and Retell are built for developers. You pick a provider for each part of the conversation and pay for each building block separately. Per minute this is the cheapest route if you do the work yourself. Our comparisons with Vapi and Retell show the calculation in detail.

Flat fee by callers. Goodcall bills neither minutes nor calls but "unique customers": every phone number that talks to the agent in a month counts once, however often it calls. [3] That is easy to plan if a few customers talk for a long time. With many short calls from different numbers it gets more expensive. The pricing page does not say whether German numbers and German-language calls are available.

Enterprise contract. Synthflow writes that final pricing is scoped around call volume, concurrency, telephony setup, integrations, security needs and launch support. Parloa describes a usage-based model on its pricing page but publishes no amounts. [4] [5] For contact centers with hundreds of seats that is common and sensible. For a medical practice or a trades business it is too big. More in our comparison with Parloa.

Minute bundle on a finished platform. This is how EchoCall bills. You pay a fixed monthly amount in euros for a minute allowance and get speech models, voices, transfers and the ready-made integrations. Per minute this is more expensive than a self-built agent on Vapi. In return you need no developers and get one invoice instead of six. All bundles are on our pricing page.

Worked example: 500 minutes a month

A business receives 500 minutes of calls a month on its German number, spread across 150 different callers. Our own calculation with list prices from 6 October 2026, Goodcall checked again on 9 October 2026:

The developer platform wins on price as long as you leave development out. Count it in and the picture flips quickly. The flat fee is predictable as long as the number of callers stays stable. The bundle is the most expensive route per minute and the shortest route to the first real call.

Cost per resolved call

The minute price alone says little. What matters is the cost of a call the agent actually handles. An example with assumed values: calls last three minutes, the agent costs €0.30 per minute and resolves 70 of 100 calls. Then 100 calls cost 300 minutes × €0.30 = €90, and each resolved call costs €90 ÷ 70 = about €1.29, not counting the phone number. If the agent resolves only 40 of 100, it is €2.25. The resolution rate moves the result more than any discount on the minute price. You can enter your own call volumes in the ROI calculator.

When an AI phone agent does not pay off

A bundle is a fixed monthly amount. Whether it pays off depends on how many calls the agent actually takes. On our smallest plan at €75 a month for 250 minutes, a call costs €3.75 if 20 calls come in a month, and €0.75 at 100 calls of two and a half minutes each. In four cases we advise against an AI phone agent, ours included:

• You only get a few calls. With a handful of calls a week, voicemail with fixed callback times is often enough. Our article on missed calls shows how to work out what a missed call costs you.
• Almost every call needs personal advice. If callers mainly want a legal, medical or tax assessment, the agent only handles appointments and routine questions. The advice stays with your team, and the savings stay small.
• The agent is supposed to decide about people on its own. Checking creditworthiness, deciding on public assistance benefits, assessing risks in life and health insurance: the AI Act lists AI systems for such decisions as high-risk in Annex III. [11] A phone agent should only take down such requests and pass them to your team.
• It is about your personal phone. EchoCall only contracts with businesses. For private calls, the call screening some smartphones now include is the better fit.

Rolling it out in five steps

An AI phone agent rarely fails because of the technology. It fails because nobody decided what it should do. These five steps have proven themselves.

1. Pick one call type. Start with the call that comes most often and needs the least judgement: appointment booking, opening hours, status questions. For one week, note down which five questions come up again and again.
2. Write down knowledge and rules. What may the agent promise, and what not? When does it transfer, to whom, and what happens if nobody picks up there? Write it in full sentences, the way you would brief a new temp.
3. Test before a customer calls. Play difficult callers: unclear, impatient, with two requests at once, in another language. With EchoCall you test the agent in the browser before any number points to it. For trying it out there are 30 voice minutes and 30 chat conversations for 14 days, with no credit card required.
4. Go live small. At first, connect the agent only outside opening hours or only when the line is busy, through call forwarding. That way your team stays the first voice, and the agent catches what would otherwise be lost.
5. Read transcripts and sharpen. For the first two weeks, read ten conversations every day. Where did the agent misunderstand, where did it transfer too early, where too late? Every correction goes into knowledge or rules. With EchoCall the transcript of every call is in the EchoHub.

Always keep the route to a person open. In a Gartner survey of 3,566 customers, 87 percent said that a company using generative AI in customer service must offer a way to reach a human. [16] Forrester predicted that in 2026 a third of companies will harm experiences with frustrating AI self-service. [17] An agent that transfers when it gets stuck is better than one that tries to do everything alone.

Seven metrics that show whether the agent works

Measure from day one. Otherwise, after two weeks all you know is that it feels good.

Add the business outcome you introduced the agent for, such as appointments booked or jobs taken.

How to recognise a good provider

Every provider should answer these questions before you sign:

• Are prices public and in euros on the website? What does a minute really cost, including telephony?
• Is there a trial without a sales call?
• Where is the platform operated, is there a data processing agreement under Art. 28 GDPR, and which subprocessors are involved?
• Can you set how long conversations are stored?
• Which calendar, CRM and shop systems are connected out of the box, and what only works through webhooks?
• Can the agent transfer live to a person, and what happens if nobody picks up there?
• Do you keep your number, through call forwarding or SIP?
• Which languages does the agent understand on the phone, and does it notice when the caller switches language?
• Who adapts the agent when a model is shut down?

For comparison, here are the answers for EchoCall. Prices are listed in euros on the pricing page, setup is free, and five seats for your team are included. EchoCall runs at IONOS in data centers in Frankfurt and Karlsruhe, and the data processing agreement under Art. 28 GDPR is ready to sign. You set how long conversations are kept per agent, from 1 to 3,650 days, with 30 days as the default. For particularly sensitive requests there is a zero-PII mode that does not store personal data permanently. Google and Outlook calendars, Cal.com, Calendly, HubSpot, Salesforce, Zendesk, Freshdesk, Shopify and Shopware are connected out of the box, and everything else runs through Zapier, Make, n8n or webhooks. The agent transfers live to your team. You keep your number through call forwarding or through SIP with Twilio, Telnyx, Vonage or Sinch. On the phone it understands and speaks more than 30 languages and recognises which language the caller is speaking. Product details are on the AI voice agent page, with examples for trades businesses and medical practices.

What the law requires

The AI disclosure since 2 August 2026

Article 50 of the EU AI Act has applied since 2 August 2026. Under paragraph 1, AI systems that interact directly with people must be designed so that those people are informed they are dealing with an AI, unless that is obvious. The notice must come at the latest at the time of the first interaction (paragraph 5). Under Art. 99(4), violations can be fined up to €15 million or 3 percent of worldwide annual turnover. [11]

Two points are often misrepresented. First, according to the European Commission's questions and answers, the "obvious" exception is to be read narrowly. A natural-sounding voice on the phone is precisely not obviously an AI to the caller. [13] Second, the Digital Omnibus, Regulation (EU) 2026/1744 of July 2026, did not postpone this obligation. Only systems placed on the market before 2 August 2026 get more time, until 2 December 2026, and only for the machine-readable marking of AI-generated content under paragraph 2. [12]

In practice one sentence in the greeting is enough: "Hello, you are speaking with the digital assistant of Dr. Weber's practice." It takes two seconds. By its wording, paragraph 1 is addressed to the provider of the system. Put the notice in your own greeting anyway and the question is settled. This is not legal advice. If in doubt, ask your data protection officer or your lawyer.

Data protection and recording

An AI phone agent processes personal data: names, phone numbers, requests. You therefore need a data processing agreement with the provider, a legal basis and an entry in your privacy policy. Also check whether the provider records the audio or only stores a transcript. In Germany, recording the privately spoken word without authorisation is a criminal offence under Section 201 of the Criminal Code. [14] If calls are stored as audio, say so in the greeting and obtain consent. For medical practices, law firms and tax advisers, professional secrecy rules apply on top. Have the deployment checked there before you start. Our documents on hosting, the data processing agreement and subprocessors are on the GDPR page.

And in the United States?

In the US, AI voices in outbound calls have counted as an "artificial voice" under the Telephone Consumer Protection Act (TCPA) since a ruling by the Federal Communications Commission (FCC) in February 2024. Such calls require the prior express consent of the person called. According to the FCC, the TCPA does not cover technologies used to answer inbound calls. The FCC proposed a duty to disclose AI voices at the start of every call in August 2024 but has not adopted it to date. [15]

Frequently asked questions

How much does an AI phone agent cost per month?

It depends on the pricing model. Developer platforms like Vapi and Retell cost roughly $0.07 to $0.31 per minute by their own figures, plus telephony and development. Goodcall charges $79 to $299 a month with a limit on callers. Synthflow lists enterprise contracts from $30,000 a year. At EchoCall the voice bundles start at €75 per month for 250 minutes, with no setup fee.

Is there a free AI phone agent?

Not permanently for a business, because every minute of conversation costs speech recognition, a language model, a voice and telephony. For testing there is starting credit: EchoCall gives 30 voice minutes and 30 chat conversations for 14 days, no credit card required. Retell gives new accounts $10 in credit according to its pricing page, which covers roughly 30 to 140 minutes depending on voice and model. [2] The call screening on some smartphones is free, but it is meant for a personal phone: it does not know your business and does not book appointments.

What share of calls can an AI phone agent handle?

There is no reliable figure across industries. It depends on your calls, your knowledge and your tools. Gartner predicts that by 2029 agentic AI will resolve 80 percent of common customer service issues without human intervention. [20] That is a forecast across all channels, not a measurement on the phone. Your own figure comes from a pilot that counts real resolutions, not just calls without a transfer.

Is an AI phone agent cheaper than an answering service or a part-time helper?

For lots of routine calls, usually yes. In Germany a mini-job may pay at most €603 a month in 2026, which at the minimum wage of €13.90 is about 43 hours. With the flat-rate employer contributions it costs the employer about €790 by our own calculation. The German answering service Starbüro charges €0.89 per minute plus €0.89 per call on its Flex plan, so a three-minute call costs €3.56 net. At EchoCall three minutes cost €0.90 when the bundle is fully used. [18] A person can do more than an agent, though. More in our article on AI phone assistance for SMEs.

Do I have to tell callers they are talking to an AI?

In the EU, yes. Since 2 August 2026, Article 50 of the AI Act requires that people are informed at the latest at the first interaction that they are dealing with an AI, unless that is obvious. On the phone it rarely is. One sentence in the greeting is enough.

May an AI phone agent make sales calls?

Technically yes, legally only within narrow limits. In Germany, an advertising call to a consumer needs their prior express consent, and a call to a business needs at least presumed consent. Advertising with an automatic calling machine always needs prior express consent. [21] Austria requires prior consent for calls to businesses too. [22] What applies in the United States is covered above under "And in the United States?". What is allowed in cold calling with AI is described in our guide to AI cold calling.

Can I keep my existing phone number?

Yes, with most providers. You set up call forwarding with your phone company to the agent's number, for example only when the line is busy or after a few seconds without an answer. Phone systems can alternatively be connected through a SIP trunk. EchoCall supports both.

Which languages does an AI phone agent speak?

That varies by provider and by the model in use. Ask for a list and test the languages your callers actually speak. EchoCall understands and speaks more than 30 languages on the phone and recognises the caller's language.

What happens when the agent does not know the answer?

You decide. The usual setup is a transfer to a person and, if nobody picks up there, a callback request with the caller's concern. What matters is that the agent does not make anything up. Whatever is not in its knowledge, it should say openly and pass on.

Is an AI phone agent GDPR compliant?

No product is GDPR compliant on its own. It depends on how you use it. You need a data processing agreement, a legal basis, clear retention periods and transparency towards callers. Check where the provider operates its platform, which subprocessors are involved and whether you can set the retention period yourself.

Sources

1. Vapi, pricing and pricing calculator: vapi.ai/pricing
2. Retell AI, pricing: retellai.com/pricing
3. Goodcall, pricing: goodcall.com/pricing
4. Synthflow, pricing: synthflow.ai/pricing
5. Parloa, pricing page: parloa.com/pricing
6. Twilio, voice pricing for Germany, current as of August 2026: twilio.com
7. OpenAI, GPT-Live 1 model page and API changelog: developers.openai.com, changelog
8. Google, Gemini API pricing and changelog: ai.google.dev, changelog
9. xAI, models and pricing: docs.x.ai
10. Deepgram, Voice Agent API of 16 June 2025 and changelog of 2 October 2025 on Flux: deepgram.com, developers.deepgram.com
11. Regulation (EU) 2024/1689 on artificial intelligence, Articles 50, 99 and 113 and Annex III: eur-lex.europa.eu
12. Regulation (EU) 2026/1744, Digital Omnibus on AI: eur-lex.europa.eu
13. European Commission, questions and answers on the transparency obligations under Article 50: digital-strategy.ec.europa.eu
14. Section 201 of the German Criminal Code, violation of the privacy of the spoken word, in German: gesetze-im-internet.de
15. FCC, Declaratory Ruling FCC 24-17 of 2 February 2024 and proposal FCC 24-84 of 7 August 2024: docs.fcc.gov, docs.fcc.gov
16. Gartner, survey of 3,566 customers, press release of 4 August 2026, quoted by IT-Online: it-online.co.za
17. Forrester, Predictions 2026 for B2C marketing and customer experience, 28 October 2025: forrester.com
18. Starbüro, pricing; Minijob-Zentrale on the earnings limit and contributions in 2026; German Minimum Wage Ordinance, all in German: starbuero.de, minijob-zentrale.de, minijob-zentrale.de, gesetze-im-internet.de
19. EchoCall, pricing and GDPR: echocall.de/en/pricing, echocall.de/en/gdpr
20. Gartner, press release of 5 March 2025 on its forecast for agentic AI by 2029: gartner.com, quoted by cmswire.com
21. Section 7 of the German Act against Unfair Competition (UWG), unacceptable nuisance, in German: gesetze-im-internet.de
22. Section 174 of the Austrian Telecommunications Act 2021, Federal Legal Information System, in German: ris.bka.gv.at

All links were checked on 6 October 2026, the Goodcall pricing page again on 9 October 2026, the Retell starting credit and sources 20 to 22 on 21 October 2026. Third-party prices may change. The provider's current price list applies.

Read next

• AI Receptionist 2026: How It Works, What It Costs and the Best Niches
• Call Center or AI Voice Agent? The 2026 Cost Comparison
• Answering the Phone in the Trades 2026: Take Jobs, Run the Emergency Line, Set Up Call Forwarding
• Automating Appointment Booking in Medical Practices 2026
• Vapi Alternative 2026: EU Region, GDPR and Real Costs