What does an AI phone agent cost? Cost blocks, calculation, comparison

"What does it cost per minute?" is usually the first question companies ask when they consider AI phone agents. The honest answer: the per-minute price matters, but it explains only part of the total cost. On top come integration, operations and the question of what quality you need. This article breaks the costs down into clear blocks, works through an example with openly stated assumptions and shows how to set up a sound comparison with staff costs.

In short: The running costs of an AI phone agent arise per call minute from four components: telephony, speech recognition (STT), the language model (LLM) and speech synthesis (TTS). Based on the list prices of several providers, telephony, speech recognition and speech synthesis together come to a few US cents per minute; the language model depends heavily on the model and the length of the conversation. On top come one-off integration costs and ongoing operations. Ready-made solutions bundle all of this into one per-minute price, anicall.io for example from €0.13 per minute.

The six cost blocks at a glance

A voice agent is a chain of several services working together in real time. How that chain works technically is explained in the guide Voice agents for business. For the cost calculation it is enough to look at each link separately:

Cost block	Billing	What it depends on
Telephony	per minute, plus phone number per month	inbound or outbound, landline or mobile, destination country
Speech recognition (STT)	per minute of audio	model, language, streaming or batch
Language model (LLM)	per million tokens	model size, length of prompt and conversation history
Speech synthesis (TTS)	per character or per minute	voice quality, model, share of time the agent speaks
Integration	one-off	CRM, calendar, knowledge base, phone system
Operations	ongoing, often flat-rate	monitoring, knowledge base upkeep, analysis, support

The first four blocks scale with call duration. The last two are largely fixed and, at low volumes, often decide profitability more than any per-minute price.

Running costs per minute: what the price lists say

The following figures come from the public pricing pages of the providers that also sit inside the anicall pipeline, as retrieved in November 2026. Prices change, so check them yourself before every calculation. All figures are in US dollars, as listed there.

Telephony. On its pricing page for Germany, Twilio lists $0.0100 per minute for incoming calls to a local number, plus $1.35 per month for the number. Outgoing calls cost $0.0283 per minute to landlines and $0.0420 per minute to mobile networks. Forwarding the audio to the AI via Media Streams adds $0.0044 per minute.

Speech recognition. On its pricing page, Deepgram lists streaming speech recognition on the pay-as-you-go plan at $0.0077 per minute for Nova-3 Monolingual and $0.0092 per minute for Nova-3 Multilingual.

Speech synthesis. On its API pricing page, ElevenLabs lists $0.04 per 1,000 characters for the Flash and Turbo models and equates that to roughly one minute of speech. The higher-quality v2 Multilingual and v3 models cost $0.08 per 1,000 characters. If you use the ElevenLabs agent platform instead of individual components, the same page states $0.08 per minute, plus the cost of the language model.

Language model. Here I deliberately give no fixed figure. The cost depends on how many tokens are consumed per call minute, and that is determined by the system prompt, knowledge snippets and the length of the history, which is sent again at every turn of the conversation. Use the formula: tokens per minute times price per token, separately for input and output. Take the price from the model provider's current price list.

Worked example with stated assumptions

An example makes the order of magnitude tangible. All assumptions are chosen freely and should be replaced with your own values.

Assumptions:

• A1: 1,000 incoming calls per month, averaging 3 minutes, i.e. 3,000 call minutes.
• A2: Calls to a German local number, audio sent to the AI via Media Streams.
• A3: Speech recognition with the multilingual model, to be on the safe side.
• A4: The agent speaks for about half of the call time; speech synthesis with Flash/Turbo.
• A5: Language model costs per minute are carried as a placeholder L.

Item	Calculation	Result per month
Inbound telephony	3,000 × $0.0100	$30.00
Phone number	1 × $1.35	$1.35
Media Streams	3,000 × $0.0044	$13.20
Speech recognition	3,000 × $0.0092	$27.60
Speech synthesis	1,500 speaking minutes × $0.04	$60.00
Language model	3,000 × L	depends on L
Total excluding language model		$132.15

In this example, the variable costs of the four components excluding the language model therefore come to around $0.044 per call minute. What this figure leaves out is the larger part of the effort: the language model, orchestrating the real-time pipeline, hosting, monitoring, resilience, updates and content upkeep. If you build it yourself, you carry all of that. A per-minute price from a full-service provider typically includes these services. So never compare the sum of raw components with the final price of a ready-made solution without factoring in development and operations.

One-off and fixed costs: integration and operations

Integration. A voice agent only creates value once it writes appointments into your calendar, files call notes in the CRM and answers questions from your knowledge base. The effort depends on whether your systems have interfaces, how clean the data is and how many call scenarios are covered. Plan integration as a project in its own right with a fixed scope, not as a side task.

Operations. This covers analysing calls, refining prompts, maintaining the knowledge base and resolving cases the agent could not handle. Internal time is what gets underestimated most: someone on the team has to close knowledge gaps and grant approvals.

Legal and data protection. Data processing agreements, notices to callers and, where necessary, a data protection impact assessment take time before the first minute runs. The rules are summarised in the article Using AI telephony in line with the GDPR (in German).

Comparing with staff costs: doing the maths fairly

The obvious comparison is staffing the phone with people. There is a solid starting figure: according to the Federal Statistical Office (Destatis) (in German), companies in Germany's industry and services sectors paid an average of €45.00 per hour worked in 2025. That is an average across many sectors, including non-wage labour costs. Costs in your company or at an external call centre can differ considerably.

From this you can derive a comparison figure per call minute, provided you know the utilisation:

Staff cost per call minute = cost per hour worked ÷ (60 × share of time spent on calls)

Suppose a person actually spends half of their working time on calls, with the rest going to follow-up work, waiting and gaps between calls. With the Destatis average, that gives €45.00 ÷ 30 = €1.50 per call minute. At 3,000 minutes a month, that would come to €4,500 on paper.

Three points make the comparison fair:

• Availability. People cost money even when the phone isn't ringing; the AI only costs money during calls. If you want to be reachable at night and at weekends, staff cost you for every hour covered.
• Call quality. AI does not replace every call. Complex, emotional or sensitive calls still belong with people. Only count the share of calls the AI can genuinely resolve.
• Value, not just cost. The bigger effect often lies not in minutes saved but in calls that used to be lost. How to turn such effects into a robust business case is shown in the article Building an AI business case (in German).

How latency and quality drive the price

The cheapest component is not automatically the best. Three trade-offs shape every calculation:

Latency versus model size. Larger language models often give more considered answers, but take longer and cost more per token. On the phone, every pause sounds unnatural. A fast, smaller model with a good knowledge base is often the better choice for standard calls.

Voice quality versus price. According to the ElevenLabs price list, the higher-quality models cost twice as much per 1,000 characters as Flash and Turbo. Whether callers notice the difference is something to test with real calls, not in a demo room.

Context length versus token costs. The more knowledge and conversation history sent at each turn, the more expensive every minute becomes. A targeted knowledge base search that returns only the relevant sections cuts costs and usually improves the answers too.

The role that reachability and answer quality play in the customer experience is described in the article AI in customer experience (in German).

From my own products

With anicall.io I approached the cost question from latency, not from price. The real-time pipeline of Twilio, Deepgram, Groq with Llama 4 and ElevenLabs streams end to end. The lesson from building it: every slower component makes the conversation worse, and a worse conversation costs more than any tenth of a minute saved, because callers hang up or ask for a human after all. Another cost factor is that knowledge gaps land with the team as an approval task instead of the agent improvising. That saves follow-up work. For customers we bundle all components into one per-minute price from €0.13 per minute. How this architecture came about is described in the field report Building a voice agent SaaS (in German).

If you would like your own calculation reviewed, you will find my approach under Voice AI.

Frequently asked questions

What does an AI phone agent cost per minute?

That depends on telephony, speech recognition, the language model and speech synthesis. Based on public list prices, telephony, speech recognition and speech synthesis together come to a few US cents per minute, with the language model on top. Complete solutions that also cover operations and integration are usually billed at a flat per-minute rate, anicall.io for example from €0.13 per minute.

Is it cheaper to build a voice agent yourself?

The raw components are cheap; developing and operating a stable real-time pipeline is not. Building your own mainly pays off with high volumes, an in-house development team and special requirements. For most companies, a ready-made solution with customisation is faster and more predictable.

How do I compare an AI phone agent with staff costs?

Divide the cost per hour worked by 60 and by the share of working time actually spent on calls. That gives you the staff cost per call minute. Also take into account availability outside business hours and the share of calls the AI can genuinely resolve.

What hidden costs come with AI phone agents?

Commonly underestimated are integration with CRM and calendar, knowledge base upkeep, call analysis and internal time for approvals. On top comes data protection work, such as data processing agreements and notices to callers.

Does better quality make an AI phone agent more expensive?

Often yes, but not linearly. Higher-quality voices and larger language models cost more per minute and can lengthen response times. A well-maintained knowledge base and short, targeted prompts, on the other hand, improve quality without raising the per-minute cost.