What an AI receptionist really costs, and can do

Prices below are the public list prices on each vendor's own pricing page, checked on 7 November 2026. They change often, they are not quotes, and plans measure different things: minutes, calls or unique customers. We have not tested these vendors and do not recommend one. Many vendors, including some serious ones, publish no prices at all.

The same volume, three ways (illustrative arithmetic)

Take 100 calls a month averaging three minutes, so 300 minutes. This is our arithmetic from the list prices above, not a quote, and a "minute" is counted differently by each vendor.

• Dialzara's $99 plan covers 220 minutes; the extra 80 at $0.45 add $36, so about $135 a month.
• Smith.ai's live-staffed $810 plan covers 90 calls; 10 more at $10.50 add $105, so about $915 a month.
• Ruby's 500-minute plan at $1,725 is the smallest listed that covers 300 minutes. Ruby lists no overage rate, so what happens when you exceed a smaller plan is a question to ask.

The gap is large, but cheap is not the same as right. A cheaper plan that mishandles the calls that matter costs more than it saves. Also count your own setup time, any phone or calendar integration fees and what happens when you go over your allowance.

VeloVoice pricing is agreed on a short call, because it depends on your call volume and the tools you use. We do not publish a figure here.

An AI phone agent listens to the caller, turns speech into text, decides what to say using a language model and speaks the answer back. It can usually do these well when you set it up properly:

• Answer. Pick up at any hour, even when several people call at once.
• Answer common questions. Hours, service areas, prices you have approved, how to prepare for a visit.
• Qualify. Ask your questions: who is calling, what they need, where, how urgent.
• Book. Check a connected calendar, offer a time and confirm it by text.
• Hand off. Transfer the call or pass the details to a person.

And these are the limits we would plan around:

• Accents and dialects. A peer-reviewed 2020 study in PNAS tested five commercial speech-recognition systems and found much higher error rates for Black speakers than white speakers, a gap the authors traced to the underlying models. Newer systems have improved, but the lesson stands: test with your real callers.
• Noise. A caller on a building site, a motorway or a speakerphone is harder to understand. Ask how the agent behaves when it cannot make out a word.
• Emergencies. A gas smell, a flooding ceiling, a dental trauma, a threat of harm. These must reach a person or emergency services quickly. Do not rely on a model to judge it unaided.
• Upset or confused callers. A person who is angry, grieving or frightened wants a human, and an AI that carries on can make things worse.
• Judgment and advice. Quotes that depend on a site visit, medical or legal advice, exceptions to your policy. The agent should say it cannot help and offer a human.
• Making things up. Language models can state wrong facts confidently. Limit it to answers you have given it, and review transcripts.

Most small businesses do not choose one. A person covers office hours, an answering service or an AI agent catches overflow and nights, and every route ends at a person for the calls that matter. If you are not sure how many calls you actually miss, start with our guide to counting missed calls and use your own log to decide.

• 1. What exactly does one unit cost?Is it a minute, a call or a unique customer? What happens at the limit, and are spam calls and wrong numbers counted?
• 2. Can I hear real calls?Ask for recordings from a business like yours, not a polished demo. Better, call the demo line yourself from a noisy place.
• 3. How does it behave when it does not understand?Does it ask again, offer a text, or pass to a person? How many tries?
• 4. How does the hand-off work?Warm transfer with a summary, a message to a named person, or only a voicemail? What if nobody picks up?
• 5. How are emergencies handled?Which words trigger it, what is said to the caller, who is rung and how fast? Ask to test it.
• 6. Does it say it is an AI?And can you set the wording to match your local rules?
• 7. Where do recordings and transcripts live?Who can read them, how long are they kept, and can you delete them? Does it comply with the privacy law that applies to your callers?
• 8. Which tools does it connect to?Your calendar, CRM and phone system. Is each one included, or an extra?
• 9. Who sets it up and who fixes it?Self-serve, or does a person configure and support it? What does a bad call cost you in support time?
• 10. What is the exit?Contract length, notice period, and whether you keep your number.

The hand-off is the part that decides whether an AI receptionist helps or hurts. This list is our own, drawn from the limits above. Settle each point in writing before launch.

• Triggers are written downEmergency words, "speak to a person", anger or distress, two failed attempts to understand, and any topic you never want an AI to handle.
• A named human receives each hand-offWith a phone number that is answered, and a second person if the first does not pick up.
• The human gets a summaryCaller name, number, what they asked and what was already said, so nobody repeats themselves.
• There is an out-of-hours planIf no one is available, the caller hears what will happen next and when, and a message is sent to you straight away.
• The emergency script is testedCall it yourself. Check that it points to emergency services where life or safety is at risk.
• Disclosure and consent are settledThe opening line says it is an automated assistant where required, and recording notices match local law.
• You review real calls weekly at firstListen to a sample, especially the failures, and fix the script. Check accents, noise and the calls that went to voicemail.

If you want to try this with an assistant that is already running, VeloVoice was built for trades, clinics, brokers and law firms. We configure the rules, including what goes to a person, test it with you and support it afterwards. Bring your hand-off checklist to the demo and see how each point would be handled.