Short answer: An AI voice agent's quoted minute rate may combine five different costs: speech-to-text, language-model processing, text-to-speech, telephony and the voice-platform fee. Those components can use different meters and rounding rules. Before comparing vendors, ask which costs are bundled, which are passed through, and how each appears on an itemised invoice.
AI voice agent cost breakdown: why one minute may hide several meters
A voice conversation feels like one service. Behind it, software listens to speech, interprets a request, generates a reply, converts that reply into audio and connects the call over a phone network. A platform may also charge for coordinating those steps.
The important distinction is cost component versus billing unit. One component might depend on audio seconds, another on text tokens and another on rounded telephone minutes. A vendor may bundle all of them into one predictable price or bill some separately.
For a documented example, Bolna's pricing documentation identifies voice AI processing, telephony and platform charges. Its voice AI layer includes STT, LLM and TTS costs. This is one vendor's documented structure, not a universal breakdown or a statement about NimbleBiz's internal costs.
Component 1: Speech-to-text (STT)
Speech-to-text turns the caller's audio into text the agent can process. Depending on the supplier and contract, the meter can be audio duration, streamed usage or another defined unit.
Bolna's published breakdown describes transcriber charges as depending on call duration rounded to seconds. Do not transfer that rule to another provider without checking its documentation.
Ask: Does STT billing include silence, hold time, both speakers or only the audio actually transcribed? Is a minimum duration applied per call or stream?
Worksheet field: STT billable audio seconds and the documented rate or included allowance.
Component 2: Language-model (LLM) processing
The language model decides what to say and may read conversation history, instructions or tool results. Model usage can be measured in tokens rather than telephone minutes. More context and longer replies can change usage even when two calls last the same amount of time.
Bolna's pricing documentation describes its LLM component as depending on model tokens. For your contract, ask whether the invoice distinguishes input and output tokens, cached tokens, model choice and tool-related usage.
Ask: Which model is included? Are model changes, long histories or connected tools charged differently? Can you export token counts per conversation?
Worksheet field: LLM usage units, model and the agreed rate or allowance.
Component 3: Text-to-speech (TTS)
Text-to-speech produces the audio the caller hears. Depending on the supplier, it may be metered by characters, generated audio duration or a package allowance. In Bolna's documented example, the TTS component depends on characters.
A lengthy spoken explanation may therefore cost differently from a brief confirmation even if both calls occupy similar wall-clock time. Voice choice, language and quality settings may matter where the vendor's terms distinguish them.
Ask: What is the TTS unit? Do repeated prompts, retries or pre-generated audio count? Are premium voices billed separately?
Worksheet field: TTS billable units, voice or model identifier and the rate or allowance.
Component 4: Telephony
Telephony connects the AI system to the public phone network. It can involve a carrier, destination-specific rates, number rental or call-leg charges. These are items to check, not fees that every vendor necessarily charges separately.
Bolna's documentation describes telephony as a per-minute charge with duration rounded to minutes in its component model. A transfer may involve more than one call leg, so request precise rules rather than assuming one conversation equals one chargeable leg.
Ask: Are inbound and outbound calls priced differently? What happens with unanswered attempts, transfers, international destinations and number rental? When does the meter start and stop?
Worksheet field: Telephony billable minutes, destination or call type and fixed charges.
Component 5: Platform or orchestration fee
The platform coordinates routing, agent instructions and the underlying services. Some vendors include this in a bundled minute price; others identify a separate platform charge.
Bolna's pricing documentation lists a per-minute platform fee in its documented model. Treat this as an example of component separation, not a guaranteed current quote or evidence that NimbleBiz charges an equivalent fee.
Ask: Is the platform fee included in the advertised minute price? Does it apply to silence, failed calls or transfers? Is usage rounded per call? What happens when you exceed an allowance?
Worksheet field: Platform billable minutes and the contracted rate or allowance.
A practical cost worksheet for one call
Create one row per call or call leg. Capture the five components separately so you can reconcile the invoice.
| Component | Usage to capture | Meter to confirm | Cost input |
|---|---|---|---|
| STT | Audio processed | Seconds or other unit | Included or quoted rate |
| LLM | Model usage | Tokens or other unit | Included or quoted rate |
| TTS | Speech generated | Characters or audio units | Included or quoted rate |
| Telephony | Connected call legs | Minutes and rounding | Included or quoted rate |
| Platform | Agent session duration | Minutes and rounding | Included or quoted rate |
For each component, use the vendor's actual billable quantity, not raw call duration unless the contract says they are identical.
Estimated call cost = STT charge + LLM charge + TTS charge + telephony charge + platform charge + applicable fixed charges.
If the vendor sells a bundled plan, do not add these five components again. Record the bundled charge and any exclusions, overages or pass-through charges separately. This prevents double counting.
Why call duration alone is insufficient
Imagine two hypothetical calls lasting 90 seconds. The first contains a short question and a brief reply. The second involves repeated clarifications and a long spoken explanation. Their telephony duration may be identical while model and generated-speech usage differ. Without component meters, you cannot reliably infer which costs more.
For another test, compare calls lasting 59 seconds and 61 seconds. Ask for actual and billable usage for each component. This exposes rounding boundaries without inventing a rupee rate. For more detail, read AI voice agent billing: per-minute rates vs 30-second pulses.
Five checks before you accept a quote
- Request an inclusion matrix. Mark STT, LLM, TTS, telephony and platform as included, separately charged, capped or excluded.
- Request the billing unit. Get the exact unit, rounding increment, minimum charge and reset boundary for each line item.
- Test edge cases. Ask for sample invoices covering a short call, a long call, an unanswered call and a transferred call.
- Check allowances and overages. Ask whether credits roll over, whether components share a wallet and what happens at the limit.
- Compare business outcomes. Measure invoiced cost per qualified lead, appointment or resolved query using real test results and consistent definitions.
Compare quotes only after both vendors have priced the same call mix, destinations, usage assumptions and included services.
How to evaluate a NimbleBiz quote
The official NimbleBiz pricing page describes plans with monthly allowances and usage beyond those allowances. It says voice usage is billed per minute, with exact rates discussed directly. The public page does not establish how NimbleBiz allocates its internal STT, LLM, TTS, telephony and platform costs.
Do not assume NimbleBiz exposes five separate charges or includes every possible cost in one minute. Request current plan terms, included voice minutes, the overage definition and a worked invoice example.
For product context, see how the NimbleBiz AI voice agent handles inbound and outbound calls. That product page is useful context, not evidence of a particular rate or billing increment.
Frequently asked questions
What are the five cost components of an AI voice agent?
The five categories to check are speech-to-text, language-model processing, text-to-speech, telephony and platform or orchestration fees. A vendor may bundle several categories into one price.
Is every AI voice agent charged for all five components separately?
No. A bundled quote may include some or all components. Request an inclusion matrix and add only charges that the contract says are separate.
Does a 60-second call always cost exactly one billed minute?
Not necessarily. Each component may use a different meter or rounding rule. Confirm whether rounding applies per call, call leg, component or billing period.
Why can two calls of the same length cost different amounts?
They may consume different quantities of model tokens or generated speech, or involve different telephony destinations and call legs. Actual differences depend on the contract.
Is telephony included in the advertised AI voice minute price?
It depends on the provider and plan. Ask for written confirmation covering call type, destination, number fees, transfers and failed attempts.
Does NimbleBiz publish a five-line component price breakdown?
The reviewed public NimbleBiz pricing page describes plan allowances and per-minute voice usage, but does not publish a five-line STT, LLM, TTS, telephony and platform breakdown. Confirm your current quote directly.
Go deeper
- AI voice agent billing: per-minute rates vs 30-second pulses
- AI Voice Agent for inbound calls: never miss a lead call again
- One AI agent for WhatsApp chats and phone calls
Ready to compare full call costs?
Bring representative call lengths, expected monthly volume and this five-component checklist to a NimbleBiz pricing discussion. Request a written definition of included voice usage, overages and separately billable items before choosing a plan.
Pricing details were checked against the official Bolna pricing documentation, Bolna plan pricing and NimbleBiz pricing page on 13 October 2026. Confirm current plan terms before buying.