The short answer: test it on your own calls, then check eight things
To choose an AI voice agent, judge it on your real calls, not the vendor's demo. Ask eight questions: what the true cost per minute is, how fast it responds, what it does when it gets stuck, how it connects to your calendar or software, how it handles consent and recording, who owns the call data, what the contract locks you into, and how success is measured.
On cost, published rates in 2026 run from roughly $0.05 to $1.00 per minute, according to Aircall's pricing guide. Managed, all-in-one platforms typically land between $0.25 and $0.50 per minute. For a small business taking 300 minutes of calls a month, that is about $75 to $150 in usage, before any setup or platform fees.
The rest of this guide walks through each question, what a good answer sounds like, and when an AI voice agent is the wrong purchase.
Question 1: What will I actually pay per month, all in?
The per-minute rate on a pricing page is rarely the number on your invoice. Cheap "infrastructure" rates, as low as $0.05 to $0.15 per minute according to Aircall, usually assume you separately buy the speech recognition, the language model, the voice and the phone line. Pricing comparisons from Autocalls (read from vendor pages in October 2026) show the same pattern: platform fees quoted before telephony and voice costs are added.
Ask the vendor for a written monthly estimate based on your call volume. Make them list:
- Setup or onboarding fee
- Monthly platform or seat fee
- Included minutes and the overage rate
- Phone number and telephony charges
- Charges for integrations, extra languages or premium voices
Some vendors sell minute bundles. JustCall, for example, lists a $99 monthly plan with 100 minutes and $0.99 per minute beyond that. A bundle that looks cheap can cost far more if you run over.
Also work out your own number first. Check your phone system for monthly inbound minutes, then multiply. If you take 1,000 minutes a month, the difference between $0.15 and $0.50 per minute is $350 a month.
Question 2: How fast does it respond, and can I hear it on a real phone?
Delay is what makes callers hang up. Telnyx notes that people expect a reply in about 500 milliseconds and that pauses past a second feel unnatural. SignalWire puts acceptable total delay at roughly 800 to 1,200 milliseconds, with problems showing past about 1,300 and calls feeling broken beyond 2,000.
SignalWire also warns that claims of under 200 milliseconds are not realistic for generative voice AI. Treat any vendor quoting a number like that with caution.
Do not rely on a web demo. Call the demo number from your mobile, from a car, and from a noisy room. Listen for three things:
- A long gap after you finish speaking
- The agent talking over you, or failing when you interrupt it
- Mishearing names, street addresses and phone numbers
Ask the vendor to show you a recorded call from a business like yours, not a scripted showcase.
Question 3: What happens when it cannot help?
This is the most important question for most owners. Every AI voice agent will hit a call it cannot handle: an angry customer, an emergency, an odd request, or a heavy accent on a bad line.
A good vendor can answer these specifically:
- Transfer: Can it hand the live call to a named person or group, during business hours, with a short summary of what the caller wanted?
- After hours: If nobody picks up, does it take a message and text or email you straight away?
- Triggers: Can you set words or situations that always go to a human, such as "emergency", "complaint" or "cancel"?
- Fallback: If the system goes down, do calls ring through to your cell or voicemail?
If the answer to any of these is vague, move on. A voice agent that traps a frustrated caller costs you more than a missed call does.
Question 4: Does it connect to the tools I already use?
An agent that only takes messages is a more expensive answering machine. The value comes when it books into your calendar, creates a customer record, or sends a text confirmation.
List the software you use and ask about each by name: your scheduling tool, CRM, practice or job management system, and your phone provider. Then ask whether the connection is built in or needs custom work, and who pays for that work.
Be wary of "we can integrate with anything" without a named example. Ask for a customer using the same software as you. If the vendor cannot offer one, assume you will be the test case.
Also check whether the agent can read live availability. An agent that offers times without seeing your calendar creates double bookings, which are worse than no booking.
Question 5: How does it handle consent, recording and the law?
Legal exposure sits mostly with you as the business, not the vendor. Two areas matter.
Outbound calls. The FCC ruled in February 2024 that AI-generated voices count as "artificial" under the Telephone Consumer Protection Act, so prior express consent is required for AI voice calls to mobile and residential lines. Statutory damages can reach $1,500 per willful violation, as summarised by Resemble AI. If you plan to have an AI call leads or past customers, ask how the vendor captures and logs consent. If you only want it to answer inbound calls, this risk is much smaller.
Recording. California generally requires all parties to consent to recording a confidential call. Calls handled by AI bots have been the subject of litigation. Ask whether the agent announces recording at the start of the call and whether that wording can be changed.
If you handle health information, financial data or legal matters, you need more: signed agreements, restricted storage and clear retention rules. Have your own attorney review the setup. This article is not legal advice.
Question 6: Who owns the call data, and where does it go?
Recordings and transcripts contain names, phone numbers, addresses and sometimes payment or health details. Ask in writing:
- Where recordings and transcripts are stored, and for how long
- Whether your calls are used to train the vendor's or a third party's models
- Whether you can export all your data, and in what format
- Whether you can delete recordings on request, and how fast
- Which outside companies (voice, language model, phone carrier) touch the audio
The last point matters because many agencies and platforms resell a stack built from several suppliers. Each supplier is another place your customers' data sits. A reputable vendor can name them.
Question 7: What am I locked into?
Contract terms are where small buyers get caught. Some enterprise-oriented platforms publish no public rate at all. Autocalls reports one major platform listing plans from $30,000 a year with no per-minute rate, which is far outside what a single-location business needs.
Before signing, ask:
- Is there a month-to-month option, or is it a 12-month minimum?
- What happens to unused minutes?
- Can prices rise during the term, and with what notice?
- Who owns the phone number? You should be able to take it with you.
- Who owns the script and the prompts built for you?
- How do you cancel, and is there a fee?
Prefer a paid trial of 30 days or less over a long contract. If a vendor insists on an annual term before you have heard a single live call, that tells you something.
Question 8: How will we know it is working?
Agree on measures before launch. Useful ones for an owner are:
- Calls answered versus calls missed, compared with your current month
- Bookings or qualified leads created from those calls
- Transfer rate: the share of calls passed to a person, and why
- Caller drop-off: how many hang up in the first 30 seconds
- Your own review: listen to 20 recorded calls a week for the first month
Ask the vendor how they report this and who reviews failed calls and fixes the script. Improvement after launch is normal. A vendor with no review process is selling software, not a working phone solution.
If you want a sense of what a managed setup includes, Zapplon describes its AI voice agents and AI receptionists service, which covers script, integration and ongoing call review. Compare any provider, including that one, against the eight questions above.
When an AI voice agent is a poor fit
Be honest about whether you need one. It is likely a poor fit if:
- You get fewer than about 20 calls a week and a person or voicemail already handles them
- Most calls are complex, emotional or high stakes, such as sensitive legal or medical conversations
- Your callers are mostly older or speak languages the agent does not handle well
- Your best revenue comes from a personal relationship on the phone
In those cases, a live answering service, a part-time receptionist or a better voicemail-to-text setup may serve you better and cost less. It is a good fit when you lose calls to after-hours, lunch breaks or busy periods, and the calls are repetitive: booking, pricing, directions, basic questions.
A two-week test before you commit
Run a short trial with a clear scorecard:
- Days 1 to 3: Place 15 test calls yourself, including a rude caller, a mumbler, an interruption and a request the agent should not handle.
- Days 4 to 7: Go live for after-hours or overflow calls only. Keep your normal line for business hours.
- Days 8 to 14: Listen to every recorded call. Count wrong answers, failed transfers and wrong bookings.
- Day 14: Compare the cost per answered call with your current cost of missed calls. If an average new customer is worth $300 and the agent rescued five that you would have missed, the maths is easy to check.
Only extend to all calls if wrong answers are rare and every transfer worked.
FAQ
How much does an AI voice agent cost for a small business?
Expect about $0.25 to $0.50 per minute on managed platforms, with cheaper build-it-yourself options from about $0.05 per minute before extras. Most single-location businesses pay roughly $100 to a few hundred dollars a month in usage, plus any setup fee. Your exact bill depends on call minutes, integrations and add-ons.
How long does it take to set up an AI voice agent?
A simple inbound agent that answers questions and takes messages can be live in days. Adding calendar booking, CRM updates and custom call rules usually takes a few weeks, mostly because of testing and script changes. Treat any promise of "five minutes" as a demo, not a working deployment.
Can callers tell they are talking to AI?
Often yes, especially when there is a delay of more than a second or the caller interrupts. Many callers do not mind if the call is quick and ends with a result. Decide in advance whether the agent will say it is an AI, and make sure a person is reachable on request.