How much does an AI voice agent cost per minute in 2026?
How much does an AI voice agent cost per minute in 2026?
A production AI voice agent costs a build fee of $17,500–56,000 (₹11.2–38.4 lakh) plus usage of roughly $0.05–0.15 per minute at scale, made up of telephony, speech recognition, the model and speech synthesis. Here is the line-by-line breakdown and how to cut it.
Voice is the one AI channel where the running cost is visible on every call, so it deserves a clearer answer than "it depends". A voice agent has two costs: a build fee, which covers the integration, dialogue design, language benchmarking, tools and evaluation, and a per-minute usage cost, which is the sum of telephony, speech-to-text, the language model and text-to-speech. This guide gives real ranges for both, a worked monthly example, the levers that move the per-minute figure, and how it compares with a human-staffed line.
Build fee
At Eazyware a multilingual voice agent runs $17,500–56,000 (₹11.2–38.4 lakh) to build, depending on the number of intents, languages, the systems it acts in and the evaluation depth. A single-site booking agent in two languages sits at the low end; a multi-site, multi-language agent with collections, payments and CRM actions sits at the top. Current figures are on the pricing page and the voice agent page.
Usage cost per minute: the four lines
| Line | Typical range (USD / min) | Notes |
|---|---|---|
| Telephony (inbound/outbound minutes, number rental) | $0.005–0.03 | Varies by country and provider; India is at the low end for inbound |
| Speech-to-text (streaming) | $0.005–0.03 | Indian-language engines and streaming modes cost more than batch English |
| Language model (per turn, tokens) | $0.01–0.05 | Routing simple turns to small models is the biggest lever |
| Text-to-speech | $0.01–0.04 | Natural voices cost more; caching repeated phrases helps |
| Total | ≈ $0.05–0.15 per minute | At scale, after routing and caching |
Provider prices change often; the figures are the ranges we see across deployments in 2026 rather than any single vendor's list. Always build the model from your actual traffic and revisit it quarterly.
Worked example: a clinic chain
| Assumption | Value |
|---|---|
| Calls handled by the agent per month | 18,000 |
| Average handled duration | 2.4 minutes |
| Agent minutes per month | 43,200 |
| Blended usage cost | $0.09 per minute |
| Monthly usage | ≈ $3,900 |
| Care Plan with AI add-on (evals, routing, dictionary support) | $1,750 (₹1.08 lakh) |
| Total monthly running cost | ≈ $5,650 (₹3.5 lakh) |
Against that, the same 18,000 calls at roughly four minutes each of staff time, including hold and wrap-up, is about 1,200 staff hours a month, before accounting for missed calls after hours. The point is not that people are replaced; it is that the front desk stops losing its mornings to routine calls, and the cost per handled call becomes a known number.
What moves the per-minute cost
- Routing by turn: greetings, confirmations and yes/no turns go to a small model; only reasoning turns go to a larger one
- Caching synthesis of repeated phrases (greetings, confirmations, menus of options)
- Shorter turns: dialogue designed to confirm efficiently cuts minutes as well as tokens
- Language choice of engines: the best engine per language is often not the most expensive
- Telephony provider and region: inbound Indian minutes are cheap; international outbound is not
- Barge-in and endpointing tuned so the agent stops talking when the caller starts, which cuts dead air
Outbound versus inbound
Outbound campaigns add answering-machine detection, retry logic and compliance checks, and the cost per *connected* minute is what matters. Reminder and confirmation calls are short and efficient; surveys and qualification calls run longer. Budget outbound at the top of the range and measure connect rate separately. See Outbound AI calling.
Comparing vendor quotes
Ask every vendor for the same four lines and the assumptions behind them, and whether provider accounts are yours (they should be, so the usage cost is transparent and yours to optimise). A quoted "$0.03 per minute" that excludes telephony or synthesis is not comparable with a fully loaded number. Ask also what is in the build fee: language benchmarking, integrations, evaluation and hand-off design are the parts that get dropped from low quotes, and they are the parts that decide whether the agent works.
When voice is not the right channel
If most of your contacts could be a WhatsApp message, the WhatsApp agent is cheaper per contact and easier to run. Voice earns its cost when callers prefer calling (older demographics, urgent needs, low literacy in the written language), when calls are the existing channel, and when the task benefits from the immediacy of a live confirmation.
Team and timeline
Six to ten weeks to a pilot with a conversational AI engineer, a real-time media engineer, an integration engineer and a delivery lead. The cost model is built in week one from your call volumes and revisited at the end of the pilot with real per-language data, so the monthly figure you plan with is measured rather than estimated.
Before you start: a checklist
- Monthly call volumes by intent and by language
- Current average handling time from your phone system
- Your telephony provider and whether numbers can be routed over SIP
- Which provider accounts you will hold directly
- The share of calls that could move to messaging instead
- The success metric: resolution without a human, cost per handled call
INR budgeting for Indian deployments
For Indian businesses the usage line lands in rupees through provider accounts you hold: inbound minutes are inexpensive, speech engines are billed in USD or INR depending on vendor, and model costs are USD. A practical monthly model for the clinic example above is roughly ₹3.2–3.6 lakh in usage plus the Care Plan, before GST on the plan. Build fees are invoiced in INR with GST for Indian clients; the pricing page shows both currencies with one toggle so the same numbers appear everywhere.
A simple formula you can reuse
Monthly usage = handled minutes × blended per-minute rate. Handled minutes = calls the agent takes × average handled duration. Blended rate = telephony + speech-to-text + model + synthesis, after routing and caching. Add the Care Plan and, for outbound, divide by connect rate to get cost per connected minute. Put it in a spreadsheet with your own volumes, and update the rate quarterly from the dashboard. Vendors who cannot fill in those four lines for you are not quoting a production system.
Glossary
- Connected minute: a minute of an answered call; the unit that matters for outbound
- Turn: one exchange between caller and agent; models are billed per turn's tokens
- Streaming rate: the per-minute price for real-time recognition, higher than batch
- Synthesis cache: reusing generated audio for repeated phrases
- Routing: sending simple turns to small models and reasoning turns to large ones
- Care Plan: the fixed monthly line for evals, tuning and support
Mistakes we see
Where budgets go wrong: quoting only the model cost; forgetting telephony for outbound and international; letting a vendor hold the provider accounts so usage is opaque; benchmarking speech on accuracy alone and paying for latency in abandoned calls; and treating the pilot's per-minute number as final instead of tuning routing and caching after the first month.
Questions clients ask
- Does per-minute cost fall with volume? Yes; provider tiers, caching and routing tuning typically bring the second-quarter figure below the pilot's.
- Is a human cheaper? Per minute of talk, sometimes; per handled call including hold, wrap-up and after-hours coverage, rarely.
- What about after-hours calls? They cost the same per minute and are the ones you currently miss entirely.
- Do we pay when the agent transfers? Only for the agent's minutes; the human minutes are your existing cost.
- Can we mix inbound and outbound on one budget? Yes, with separate dashboards, because outbound connect rates change the maths.
What good looks like after 90 days
A ninety-day review compares the pilot's blended per-minute cost with the tuned figure, per language, and the cost per handled call with the pre-agent baseline. Most deployments show the per-minute figure falling and the resolved-call count rising, which is the shape finance expects.
Related reading
What is an AI voice agent? for the architecture, How much does an AI agent cost? for the build side across agent types, and Voice AI compliance in India before any outbound campaign. Telephony pricing references: Twilio and Exotel.
Budget both numbers, hold the provider accounts yourself, and measure per minute per language from the first week of the pilot. A voice agent whose cost you can see is one you can improve.
Finally, treat the per-minute figure as a managed number rather than a fixed one. Providers reprice, models improve, routing rules can be tuned and caches grow. The deployments with the lowest cost a year in are the ones where someone reviews the four lines every quarter, which is exactly what the Care Plan's cost review exists to do.
Frequently asked questions
Is there a minimum monthly usage?
▾
No. Usage is pay-as-you-go through your own provider accounts; the Care Plan is the only fixed monthly line.
Why is Indian-language speech more expensive than English?
▾
Fewer engines compete, streaming modes cost more, and routing per language sometimes uses two engines; the difference is usually a cent or two per minute.
Can we cap spending?
▾
Yes. Per-agent budgets and rate limits are configured from day one, and the dashboard shows cost per minute per language.