Comparing AI voice agent pricing for small business is harder than it looks, because a vendor page shows one number and your bill contains several. The headline is usually a per-minute rate. It rarely covers everything a caller sets in motion, so your first job is to see which layers sit inside that rate and which do not.
This guide breaks the bill into its parts, shows why the per-minute figure misleads, and gives you a better unit for comparing quotes. It does not quote market prices, because they shift quickly and depend on your call volume. Instead, you get the questions that make any quote comparable.
A voice agent is a chain of services, not a single product. OpenAI's voice agents documentation describes a chained pipeline with three stages: speech-to-text, the agent workflow itself, and text-to-speech. Add the phone line that connects callers, and you have four cost sources before any business logic runs. Many platforms bundle these into one rate, which is convenient but hides where the money goes.
TechCrunch's report on Retell AI notes that the company's customers pay per minute of voice agent conversation. It also describes the agents as a mix of language models tuned for customer service and a speech model that voices their text. So a single per-minute rate typically stands in for several underlying components, and you should ask which ones.
| Cost layer | What it covers | What to ask the vendor |
|---|---|---|
| Telephony | Phone numbers, call routing and transfers | Are numbers and carrier usage included, or billed separately? |
| Speech recognition | Turns the caller's voice into text | Is it bundled in the rate, and does it handle your callers' accents and languages? |
| Reasoning model | Decides what to say and which tool to call | Is usage metered by tokens, or hidden inside the rate? |
| Speech output | Gives the text a voice | Does a premium voice change the rate? |
| Integrations | Calendar, CRM, payments, ticketing | Which connectors come included, and who builds the rest? |
| Testing and monitoring | Recordings, transcripts, evaluation | How long are recordings stored, and at what cost? |
| Human handoff | Passing hard calls to your staff | Do minutes spent before the transfer still count? |
A per-minute rate distorts any budget, because call length varies far more than the rate does. Short, clean calls look cheap, while calls with interruptions, corrections or repeated questions cost much more than the average suggests. Retell's co-founder told TechCrunch that the team focuses on reducing latency and handling edge cases such as interruptions, which hints at why call length is hard to predict.
Model pricing adds another moving part. Anthropic's page for Claude Haiku 5.5 says pricing on the Claude Platform depends on prompt length, and lists $0.10 per million input tokens for prompts up to 100K tokens. A voice agent that carries a long script and the whole conversation into every turn sends more text each time. As a result, a longer prompt can lift the cost per call even when the call itself stays short.
Consider a hypothetical clinic where one call in five needs a human transfer. The vendor quote covers the agent's minutes, and then your receptionist spends time finishing the job. The second cost never appears on the invoice, but it belongs in your comparison.
The comparison that holds up is cost per completed task. Divide your total monthly spend by the number of calls that ended with the job done and no human rescue. A completed task might be a booked appointment, a qualified lead or a rescheduled delivery. This single figure pulls every layer into one number, including the ones outside the quote.
To calculate it, add platform fees, telephony, model usage, integrations and the staff time spent on transfers and corrections. Then divide by completed tasks, not by calls answered. A cheap rate with a low completion share can cost more than a pricier agent that finishes most jobs.
OpenAI's guidance supports this view. It tells teams to test both the conversation and the completed task, for example by listening to a booking confirmation and checking that the correct appointment was saved. It also recommends comparing median and 95th percentile response times across similar calls. Ask every vendor for completion results on scenarios that resemble your own calls.
Payments deserve their own line. In its Y Combinator launch post, the startup Maven argues that voice agents can qualify leads and book appointments but stall at the transaction. Its list of obstacles includes PCI-compliant infrastructure, handling card data, failed payments mid-call and support for gateways such as Stripe, Authorize.net and Adyen. If your callers pay by phone, budget for that work separately.
A low per-minute rate tells you what a call costs. It does not tell you what a finished job costs.
Search results for this topic mix two categories. Google's Workspace page lists Google Voice business plans starting at $10, and those plans price phone numbers and users. An AI agent answers calls on its own, so vendors tend to bill it by usage rather than by seat. Mixing the two gives you a budget that is wrong in both directions.
Off-the-shelf platforms keep upfront cost low and move spend into usage. A custom build raises the upfront cost and gives you control over each stage, which is the reason OpenAI gives for the chained pipeline: you can inspect, transform and replace every component. Neither route is cheaper by default, because the answer depends on call volume and how unusual your workflows are.
For a narrow task such as booking or call triage, a configurable platform is usually the sensible start. Custom work earns its cost when you must connect the agent to internal systems, run policy checks before it answers, or avoid dependence on one vendor's rate. Our Xerx AI development and workflow automation service covers this kind of scoping, and the same logic applies to any provider you consider.
If a voice agent is one candidate among several, read where business process automation AI really pays off first. Then run the checks for any automation business process to confirm your call volume justifies the effort. Many vendors compete for this work, since TechSci Research projects the global contact center AI market to reach nearly $3 billion in 2028, up from $2.4 billion in 2022. That competition works in your favour when you ask for a paid pilot.
Good AI voice agent pricing for small business is a pilot result, not a brochure figure. Costs and suitability vary by company, and recording calls raises privacy duties, so confirm your obligations with the vendor and a qualified adviser before launch.
Some do. TechCrunch reported that Retell AI's customers pay per minute of voice agent conversation. Other vendors may bill by seat, by task or by token usage, so ask which units apply to your quote.
A voice agent combines telephony, speech recognition, a reasoning model, speech output, integrations and monitoring. A single rate may bundle only some of them, and transfers to your staff add cost outside the invoice.
No. Google Voice is a business phone number and calling system priced by plan, with Google Workspace listing plans starting at $10. An AI voice agent answers and handles calls itself and is usually priced by usage.
Run the same call scenarios through each one and calculate cost per completed task. Include platform fees, usage, integrations and the staff time spent on transfers and corrections.
Usually not at first. A configurable platform suits narrow tasks such as booking. Custom work makes sense when you need deep system integration, policy checks before answers or independence from one vendor. Costs vary, so price a pilot before committing.
Keep reading
Visual automation builders work well for low-risk handoffs, but ownership, data and failure risk decide how far they should go. Here is a way to sort your processes.
Business process automation AI rewards stable, repeatable work and punishes unclear processes. Here is how to pick candidates, plan for exceptions and judge cost.
Open source AI coding agents cost nothing to download, but running them does not. Here is where the real costs sit and when a pilot makes sense.
