Quick answer (published 25 September 2026): The cost of a voice AI agent decomposes into four metered components — speech-to-text (~$0.005–$0.02 per minute), the language model (per-token and the least predictable line), text-to-speech (~$0.01–$0.30 per minute of speech depending on the voice tier), and the telephony minute (~$0.004–$0.02 in the US, more to hard-to-reach destinations) — plus a platform fee. The ROI of that agent is decided not by its per-minute cost but by its cost per resolved conversation: multiply the calls it completes by average minutes by the gap between a fully-loaded human talk-minute (~$0.50–$1.50+) and the agent's all-in rate, then scale by the share of calls it finishes without escalating. Both halves — the breakdown and the ROI — are covered with formulas here.
The AI voice agent pricing guide introduced the four components; this post teaches them as a discipline — how to build a cost breakdown you can audit against a vendor invoice, and how to run the ROI arithmetic that a blended quote hides from you. Where the pricing guide asks "what does it cost?", this one asks "where does the bill move, and is the return real?"
1. Why a cost breakdown beats a blended quote
Vendors who quote a single per-minute rate have folded four services into one number. That is convenient for a demo and disastrous for an audit: the moment your configuration diverges from the default — a premium voice, a frontier-tier language model, international destinations — the folded rate stops describing your bill. A breakdown assigns each stage its own meter so the variance is attributable:
- Speech-to-text (STT) transcribes caller audio in real time, billed per minute of audio. It is the cheapest and most stable line on the invoice.
- Language model (LLM) decides the reply, billed per token in and out. It is simultaneously the largest and the most sensitive to agent design — prompt length, tool calls, context reuse.
- Text-to-speech (TTS) speaks the reply, billed per character or per minute of speech. A cloned or premium voice costs 10–30× a standard one; the spread is a choice, not an accident.
- Telephony carries the call itself, billed per minute per destination. It is the line most buyers forget to budget and the one most protective when you control the carrier path.
- Platform / orchestration is the fee the vendor layers on top — for pipeline, tools, dashboards, and SLAs.
Once each stage has its own meter, you can audit the invoice line-by-line. When your audit of a provider bill lands outside your recomputed band by even a modest multiple — and the published true cost of voice AI and SMS explainer collects independent analyses showing a 3–8× headline-vs-real gap on one popular unbundled stack — the breakdown locates the inflated line before you accept it as unavoidable. The component TCO math post works the same four-box grid on a per-stage formula; this post takes the discipline the final mile, into ROI.
2. The ROI framework: from cost per minute to cost per resolved conversation
Cost per minute is an input. The return on the investment is cost per resolved conversation, measured against the fully-loaded human minute the agent displaces. The framework has four steps and the discipline is in measuring two things most buyers skip: the resolution rate, and the loaded human rate.
- Fully-loaded human cost per minute — wages plus benefits, tooling, management, and idle time, divided by talk-minutes actually productive. In most support markets this lands above $0.50 per talk-minute and often reaches $0.75–$1.50+ once overhead is counted.
- All-in AI cost per minute — the four components above summed, plus any platform fee, on the configuration you actually run.
- Resolution (deflection) rate — the share of conversations the agent finishes with no human involved. A 60% resolution rate on a $0.12 all-in minute is a different return from 20% on the same rate.
- The math —
Monthly saving ≈ (AI-handled calls) × (avg minutes) × (human $/min − AI $/min), minus one-time build and tuning time.
The sensitivity in that formula is almost entirely in the resolution rate and the human baseline, which is why a pilot pins both before extrapolation. The explicit guard against over-claiming: the savings line above turns on exactly two free variables, so extrapolate only after both are measured — not the day a demo exists.
3. Worked ROI example, end to end
Pick a line taking 10,000 calls a month at 4 minutes each. The agent resolves 60% without a human; the human-handled minute costs $0.90 fully loaded; the all-in AI rate is $0.12. Then:
10,000 × 0.60 × 4 × ($0.90 − $0.12) ≈ $18,700 per month before build time
Treat the arithmetic literally — this is the same four-step framework above, not a forecast promise — and the exposed variables map back to the two knobs that move the result. The interactive voice-AI cost calculator runs the same staged arithmetic on your own numbers instead of doing it by hand. For the break-even-across-three-tiers view, the AI vs human TCO comparison prices this same formula against FAQs (85% resolved), triage (70%), and complex (40%) call mixes.
4. Where the ROI discipline breaks
The framework fails in predictable places, and naming them here is part of teaching it:
- Measuring the deflection rate as a pilot artifact. A demo deployed to a short list of FAQ intents deflects everything; production with the broader call mix often deflects a fraction. Measure resolution against the mix it will face, not the mix you escaped while testing.
- Underestimating the human minute. When wages are quoted but benefits, queue idle time, and supervision are omitted, the human baseline compresses and the ROI inflates. Load it before you compare.
- Counting the p50 where the p95 drives cost. An agent averaging 700 ms per turn with a spiky tool-call tail burns minutes per conversation beyond the median budget; the all-in rate should come from the p95 row, not the p50.
- Double-counting soft returns. After-hours coverage, faster time-to-answer, and consistent handling are real returns; they rarely appear on a spreadsheet, and double-counting them against the per-minute saving is how the discipline unravels. Keep them as upside, not as loaded line items.
5. How Orbit prices the breakdown it teaches
Orbit by Devotel bills AI voice agents pay-as-you-go on one account, so the four components are not four invoices you reconcile against each other. As of August 2026, the published pay-as-you-go rate is $0.014 per voice minute an AI agent is on a call, $0.007 per AI agent message, and an optional $0.005 per resolved outcome if you opt into outcome-based billing — with no platform fee, no seat fee, and no monthly minimum. The staged rate card is on the pricing page, with per-country voice rates on the voice pricing page.
Two design choices keep the cost discipline honest. The pipeline runs on Orbit's own streamed speech-to-text → language model → text-to-speech stack, with per-agent model tiering, prompt caching, and speculative execution against a published ~1.1-second p50 turn target— the per-stage methodology is open on the latency benchmark page. And the telephony line rides Devotel's own wholesale softswitch — which connects to 500+ global carriers directly — so the AI agent's call terminates over the same softswitch rather than a resold aggregator hop. That is one fewer intermediary marking up one of your four component lines; on a breakdown-and-ROI view of the bill, that is a material line. See the AI voice agents overview, the CPaaS overview, and the cloud phone system.
Frequently asked questions
How do I break down the cost of a voice AI agent?
Assign each pipeline stage its own meter: speech-to-text per audio minute, the language model per token, text-to-speech per character or speech-minute, and telephony per minute per destination. Sum the four, add any platform fee, and the result is an auditable per-minute rate you can compare to an invoice — unlike a blended quote folded across all four.
What is the ROI of an AI voice agent?
ROI is the gap between cost per resolved conversation for the AI agent and the fully-loaded cost of a human minute it replaces. Estimate it as (AI-handled calls) × (average minutes) × (human $/min − AI $/min), minus build time, and let the resolution rate and the human baseline be measured on a pilot before extrapolation. The published four-step framework above is the formula to reuse.
Why does the per-minute cost vary so much between configurations?
Because a marketed per-minute rate usually assumes a standard voice, a mid-tier model, a short prompt, and a domestic destination. Premium or cloned voices, frontier-tier models, longer prompts, and international destinations each move the bill independently — which is why a folded blended quote misleads and a component-by-component breakdown does not.
What makes a cost-per-minute quote misleading as an ROI proxy?
Cost per minute is an input, not a result: a low per-minute rate with a low deflection rate loses to a higher per-minute rate with a higher deflection rate the moment a human still handles escalated conversation at the human rate. The metric that decides the investment is cost per resolved conversation — so measure the resolution share, not just the staged rate.
How does Orbit bill a voice AI agent differently from an unbundled stack?
Unbundled vendors pass the STT, language model, text-to-speech, and telephony stages through from separate providers, which means four invoices, three markups, and one reconciliation problem. Orbit bills on one account at one staged pay-as-you-go rate, with no platform fee, no seat fee, and no monthly minimum — and telephony rides Devotel's wholesale softswitch rather than an aggregator hop.
Where do I find the follow-on cost and ROI math?
The AI voice agent pricing guide covers the four components in the base case; the component TCO math post prices the per-stage formulas; the AI vs human TCO comparison breaks even across three resolution tiers; and the interactive voice-AI cost calculator runs the arithmetic on your own configuration.
The takeaway
The cost breakdown of a voice AI agent is a discipline, not a vendor concession: separate STT, LLM, TTS, and telephony into their own meters, add the platform line, and audit the invoice against the band you computed. The ROI of the same agent is the cost per resolved conversation — scored against a fully-loaded human minute, with the deflection rate measured and the two soft-side returns (coverage, speed) kept as upside rather than loaded into the ledger. Do both and the blended quote stops pricing the demo and starts pricing the deployment.
Published 25 September 2026 — an educational deep dive into the cost breakdown and ROI discipline for voice AI agents, building on our AI voice agent pricing guide and closing the gap between blended quotes and auditable invoices.