What do voice AI agents for customer service cost?

The five line items every bill consists of

1. Telephony minutes

   The surface where every other cost hangs: the PSTN/SIP trunk minutes that carry the call and the carrier termination fee behind them. Per-minute models bundle this into the metered rate; per-seat models bill it on top as carrier pass-through; the honest number is always the destination-specific per-minute rate it resolves to.

2. Speech recognition (ASR)

   Converting the caller's audio to text. Priced per second of audio in most stacks, bundled into the agent's per-minute rate in integrated ones. The line item to audit when a provider quotes a low 'agent minute' but bills recognition separately.

3. Language model (LLM)

   The reasoning layer that decides what the agent says each turn — billed per token on the model's own price sheet in DIY stacks, or bundled inside the agent's per-minute rate by providers that absorb it. The swingiest line: a small-model default and a large-model escalation differ by an order of magnitude.

4. Speech synthesis (TTS)

   Voicing the reply back to the caller. Per character on speech-vendor sheets, folded into the agent's per-minute rate in integrated platforms. Voice quality is the real decision — the cheap line item that buys nothing if callers hang up before the sentence ends.

5. Orchestration

   The real-time layer that keeps ASR, the LLM, and TTS in step inside ~1 turn of latency — a platform service in integrated stacks, or an engineering function you staff in DIY ones. The line item the per-seat and per-call models sweep under 'platform fee'; the one the per-minute model has to price explicitly, so it is the fairest thing to ask for a line-item rate.

Audit a voice AI quote in five checks

1. 1

   Identify the pricing model first

   Before any number, name the model the quote runs on: per-minute (pay-as-you-go), per-call, hybrid platform-plus-usage, or seat-licensed. The same 'voice AI agent' price means a different thing in each model — until the model is named, no total can be compared.

2. 2

   Break the quote into the five line items

   Ask for the bill split into telephony minutes, speech recognition, the language model, speech synthesis, and orchestration. A quote that cannot break into at least these five cannot be audited; the bundled per-minute model collapses them into one visible rate precisely so it can be.

3. 3

   Audit rounding on metered minutes

   In per-minute and per-second models, ask how seconds roll into a billable minute — per-second billing is the auditable shape; per-minute-with-60-second-rounding turns half-minute calls into full-minute bills, which is where short customer-service calls overpay.

4. 4

   Weigh the hidden platform or seat fee

   Hybrid and seat models meter minutes on top of a fixed monthly fee. That fee is real money even at zero minutes — divide the fee by your expected volume to see the effective per-minute adder, then compare it against the provider's own published per-minute rate.

5. 5

   Reconcile against a published rate card

   The cleanest audit is one you can run yourself: a published per-destination rate card, readable before you sign up. If the provider cannot point you at a public page that quotes the number, the layered quote is doing the hiding the page described.

Voice AI agent cost — frequently asked

What do voice AI agents for customer service cost?

The cost breaks into five line items: telephony minutes (the call's transport), speech recognition (transcription), the language model (the reply decision), speech synthesis (voicing the reply), and the orchestration layer that keeps them in real time. Per-minute pricing bundles all five into one metered rate; per-call pricing charges per completed conversation; hybrid pricing pairs a monthly platform fee with metered usage; seat-licensed pricing charges per agent-month like a human seat.

Is per-minute or per-call pricing cheaper for voice AI agents?

It depends on your call-length distribution. Per-minute wins when calls are short and their volume is low (you pay only for what you meter); per-call wins when calls are long enough that the flat per-call fee amortizes under the metered total. The only honest comparison is your own call-length distribution against BOTH rate cards — which is why published rates beat layered quotes.

What are the five cost components of running a voice AI agent?

Telephony minutes (PSTN/SIP transport + termination), speech recognition (audio to text), the language model (token-billed reasoning, or bundled), speech synthesis (per character, or bundled), and orchestration (the real-time ASR-LLM-TTS layer every stack needs). Any quote that cannot decompose into at least these five cannot be audited.

Why is a monthly platform fee a hidden cost?

Because the fee is real money at zero usage. Hybrid and seat-licensed models meter minutes on top of a fixed fee, so at low volumes the fee-to-volume ratio dominates the per-minute adder. Divide the monthly fee by expected minutes to get the true per-minute cost; a 'low per-minute' rate on top of a platform fee is rarely lower than a pay-as-you-go published rate.

How does Orbit price voice AI agents?

Orbit by Devotel prices voice AI agents as pay-as-you-go minutes with no monthly platform fee. The per-destination voice rate is published on the pricing page, the native agents ride the same minutes, and outbound calls terminate on Devotel's own wholesale softswitch — so the five line items (minutes, recognition, the model, synthesis, orchestration) are bundled into one auditable per-minute number you can read before you sign up.

Explore more

• Voice AI agents
• Voice pricing by destination
• Programmable voice API
• Pricing-model head-to-head
• How the platform works
• Pay-as-you-go pricing