Quick answer: An AI voice assistant can answer your business phone or website calls around the clock, take several calls at once, answer questions from your own information, book appointments and hand a caller to a person when it should. It works well on a small, clear set of call types and badly on everything you did not prepare it for. Before you buy, check five things: whose information it answers from, how it hands off to a person, how it says it is an AI, how long call records are kept, and what it costs per month at your real call volume. You can call our live demo to hear one.
This guide is for small businesses deciding whether a voice assistant is worth it and how to choose one. It covers how the technology works in plain terms, the features that matter and the ones that do not, how to work out the cost on your own numbers, how to compare vendors, a realistic setup plan, and the mistakes we have made ourselves running assistants for our own sites and for clients. We do not quote industry statistics we cannot source, and we do not promise you an automation rate before seeing your calls.
How an AI voice assistant actually works
Every voice assistant, whoever sells it, is the same chain of four parts. Knowing them helps you ask vendors the right questions.
- Speech to text. The caller's voice is turned into text in real time. Quality depends on the language, the accent, background noise and the phone line.
- A language model with your knowledge. The text goes to a large language model that decides what to say. On its own the model knows nothing about your business, so a good assistant answers from your documents: prices, hours, services, policies. This is usually called retrieval-augmented generation (RAG): the system looks up the relevant part of your information and the model answers from it, instead of guessing.
- Text to speech. The answer is spoken back in a synthetic voice. Providers such as ElevenLabs, Azure and OpenAI offer voices in many languages; they differ in how natural they sound and, as we found, in how loud they are on a phone line.
- Actions and handoff. The useful assistants do something: check a booking calendar, create an order, send a summary by email, or transfer the call to a person.
The whole chain has to run in about a second per turn or the conversation feels broken. A slow step anywhere (a slow model, a slow lookup, a slow calendar) shows up as awkward silence, and no amount of prompt writing fixes a timing problem.
What it can and cannot do for a small business
What it does well
- Never a busy signal. Each caller gets their own conversation, so a rush or a lunch break does not turn into missed calls.
- After-hours calls. Evening and weekend callers reach an answer instead of voicemail.
- Repetitive questions. Opening hours, prices, delivery areas, what is included in a service: answered the same way every time, from your information.
- Simple bookings and orders, read back to the caller before they are confirmed.
- A written record. Every call can produce a transcript and a summary, which shows you what people actually ask.
Where it stops
- Anything you did not give it. If the answer is not in its information, a well-built assistant says so and offers a person. A badly built one invents an answer.
- Judgement calls. Complaints, refunds, negotiations and anything unusual should go to a person.
- Names, emails and numbers spoken aloud. Speech recognition mishears them. The assistant should ask the caller to spell them and read them back.
- Bad lines and noise. A caller on speaker in a car is harder to understand, for an assistant as for a person.
Features that matter, and nice-to-haves
Must-haves
- Answers only from your information, with a clear "I don't know, let me pass you to someone" instead of a guess.
- A handoff to a person: a transfer to a phone number, or a message taken and emailed to you straight away.
- AI disclosure at the start of the call. Under Article 50 of the EU AI Act, people must be told when they are talking to an AI system, from 2 August 2026. When we checked seven of our own live assistants against it, four needed fixes: one called itself a "virtual assistant" without saying AI, two only admitted being AI when asked, and one needed the fix in two separate codebases.
- Transcripts and summaries you can export. Some platforms keep call history only for a limited time. One we use keeps 14 days on its standard plan, so anything not exported in that window is gone.
- A voice that is clear on a phone. Test it on a real phone line, not only in a browser. We once replaced a voice that sounded fine on a laptop but was too quiet on phones.
- Your languages, tested with real callers, not only a language list on a sales page.
Nice-to-haves
- Calendar or booking integration, if bookings are a large share of your calls.
- CRM or order-system integration, once the basic assistant works.
- A website voice widget in addition to the phone number.
- Outbound calls (reminders, follow-ups). Start with inbound; outbound brings consent and calling-rule questions.
What it costs: work it out on your own numbers
Prices vary a lot between vendors and change often, so instead of quoting a market range we cannot source, here is how to compare any offer.
A voice assistant usually has three cost parts:
- Setup: writing the knowledge base, configuring the call flows, integrations and testing. Often a one-time fee.
- A monthly fee: hosting, maintenance and updates.
- Usage: a per-minute charge for the phone line, speech recognition, the language model and the voice. Ask for the all-in cost per minute.
To see whether it pays off, write down your own figures:
- How many calls a month go unanswered today (your phone system or provider can usually tell you).
- What share of those callers would have become a booking or a sale, and what that is worth to you on average.
- How many minutes a month the assistant would handle, times the all-in cost per minute, plus the monthly fee.
- How many staff hours on the phone it would free, and what those hours are worth to you.
If (1) times (2) plus (4) is clearly larger than (3), plus the setup spread over a year, it is worth testing. If the numbers are close, start smaller: after-hours calls only, or one call type only. For a structured version of this, see our worksheet on pricing the wait before pricing the project.
A vendor scorecard
Ask every vendor the same questions, in writing, and compare the answers side by side:
| Question | What a good answer looks like |
|---|---|
| Where do the answers come from? | Only from your documents, with a defined "I don't know" response |
| How does it hand off to a person? | Transfer or immediate message, tested in the demo |
| How does it disclose it is an AI? | In the first sentence of the call, in every language you use |
| How long are recordings and transcripts kept, and where? | A stated period, with export, and the hosting region named |
| What is the all-in cost per minute at my volume? | One number, in writing |
| Can I test it on my own questions before paying? | A test on your real call types |
| What happens if the platform changes or shuts down? | Your knowledge base and transcripts are yours to take |
A vendor who will not answer one of these in writing is telling you something.
A realistic setup plan
- List your calls. Take two weeks of calls (or your best estimate) and group them: the five or six call types that make up most of the volume are where to start.
- Write the knowledge base for those call types only: the facts, prices, policies and the exact cases where a person must take over.
- Build and test internally. Call it yourself, with a noisy background, with a strong accent, spelling a name, asking something it does not know. Fix what breaks.
- Go live on a limited slice: after hours, or overflow when nobody picks up.
- Read the transcripts every week for the first month. Every question it could not answer is either a gap in the knowledge base or a call type that should go to a person.
- Widen it only when the first slice works.
Mistakes we have made, so you do not have to
- Assuming the widget is still there. After a client's website theme was updated, our assistant's widget silently disappeared from the site, and it took about two weeks of zero calls before we noticed. Check that it is still on the page after every site change, and watch the call count: zero calls is a signal, not a quiet week.
- Testing only in the browser. The voice that sounded best on a laptop was too quiet on mobile phones.
- Trying to fix slowness with instructions. A long pause comes from a slow step in the chain, not from the prompt.
- Not asking callers to spell. Emails and surnames heard by a machine are often wrong unless they are spelled and read back.
- Forgetting to export. Call history on some plans disappears after a fixed number of days.
Security, privacy and compliance
- Disclosure. Tell callers at the start that they are speaking with an AI, and how to reach a person (EU AI Act, Article 50).
- Personal data. Calls contain personal data, so the usual data-protection rules apply: tell callers what is recorded and why, keep it only as long as you need it, and know which provider stores it and where.
- Access. Limit who can read transcripts and recordings, and remove access when people leave.
- Payments. Do not take card numbers over the voice assistant; send a secure payment link instead.
This is technical guidance, not legal advice. For your specific obligations, ask your lawyer or data protection adviser.
Frequently Asked Questions
What is the best AI voice assistant for a small business? The best one is the one that answers only from your information, hands off to a person cleanly, tells callers it is an AI, and has an all-in cost you can afford at your call volume. Compare vendors on those points in writing, and test each one on your own real questions before paying.
How much does an AI voice assistant cost for a small business? Usually a setup fee, a monthly fee and a per-minute usage charge. Ask each vendor for the all-in cost per minute at your volume, and compare it with what your missed calls and phone time cost you today.
Can an AI voice assistant replace my receptionist? It can take the repetitive and after-hours calls, and every call at once during a rush. It should not handle complaints, refunds or unusual requests on its own, so most small businesses use it alongside a person, not instead of one.
Do I have to tell callers they are talking to an AI? In the EU, yes: Article 50 of the AI Act requires people to be informed that they are interacting with an AI system, from 2 August 2026. It is also simply good practice everywhere.
How long does it take to set up? A first version for a handful of call types can be built and tested in a few weeks. Plan time after launch to read transcripts and fill the gaps; that is where most of the improvement comes from.
What does RAG mean for a voice assistant? Retrieval-augmented generation: the assistant looks up the relevant part of your own documents and answers from them, instead of relying only on what the language model learned in training. It is what keeps the answers specific to your business.
Working with us
We build voice assistants for websites and phone lines around your actual call types: a knowledge base from your own documents, AI disclosure in the first sentence, a handoff to a person, spelled-and-read-back details, transcripts you own, and weekly review in the first month. We do not publish success rates we cannot show you, and we do not promise an automation rate before we have seen your calls.
Pricing is a fixed price for the setup, agreed in writing, plus a monthly fee. Hear one in the live demo, read about the Voice AI service, or send us your call types in writing and we will reply with a written scope and price. We work in writing, no call needed.