Who answers your phone at 7:40 in the morning, while you're on a job or with a customer? And what happens to the caller when nobody does?
On October 1, BT Business launched an AI receptionist for small firms in the UK. It answers calls 24/7, books appointments into Outlook or Google Calendar and transcribes each conversation. BT's own research claims 67% of businesses miss at least 1 call on an average day that they believe costs them work. That's a phone company's survey, so treat it as a sales figure. Still, you probably know your own number better than BT does.
The same kind of tool is already sold in the US by RingCentral and Zoom, among others, and it's cheap to try. Cheap to try is different from safe to switch on, though. An AI that answers your phone speaks for your business to people you've never met.
This guide is designed to:
- Explain what an AI receptionist does and what it costs at the time of writing.
- Help you work out whether your call volume justifies one.
- Walk you through a 2-week test with real callers.
- Give you a 1-page set of rules before it goes live.
After reading it, you will be able to:
- Count the calls you miss and what they're about.
- Compare plans on minutes, overage and trial length.
- Write the script and the "hand it to a human" rules.
- Score a pilot and decide to keep it, fix it or switch it off.
The basics: what an AI receptionist does
An AI receptionist is a voice assistant that picks up your business line, talks with the caller in plain speech and then does something useful with the call. It might answer a common question, book a slot in your calendar, take a message or transfer the caller to a person. You train it on your website, your FAQs or documents you upload. There's no recorded "press 1 for sales" menu.
The pricing is mostly about minutes. As of October 2026:
| RingCentral AI Receptionist | Zoom AI Receptionist | |
|---|---|---|
| Entry price | From $49 a month standalone, $39 as an add-on to RingEX | From $25 a month, billed annually |
| Minutes included | 100 | 100 |
| Extra minutes | $0.50 a minute pay as you go, or bundles | Add-on minutes, rate not listed |
| Trial | 14 days free | "Try for free", length not stated |
| Works without the vendor's phone system | Standalone plan sold separately from RingEX | Yes, no Zoom Phone needed |
Sources: RingCentral's pricing page and Zoom's AI Receptionist page. Check both before you buy, because vendor prices change often.
RingCentral rounds call time up and bills it in 30-second increments. So a caller who hangs up after 35 seconds costs you a full minute. If your average call runs 2 minutes, 100 minutes covers roughly 50 calls a month. Zoom's page says it handles English, Spanish and 10+ other languages, and that most setups are live within 60 minutes.
OK, so the tools exist and the entry price is low. But should your phone be answered by one?
Step 1: Count the calls you actually miss
Big call centers don't buy software on a hunch. They start with a "call driver" count: why people ring, how often and at what times. Your version takes a week and a notepad (or your phone provider's call log).
For 5 working days, write down every call you couldn't take: the time, the reason (if they left a message) and whether you got the work in the end.
Pause and reflect: Look at the times on your list. Are the missed calls mostly after hours, during lunch, or while you're on site? Each pattern points to a different fix.
Ask yourself:
- How many calls a week go unanswered, and how many of those were new customers?
- Are most calls the same 5 questions (hours, prices, availability, directions)?
- Do callers need to book something, or do they need to talk to you?
- Would a wrong answer be awkward, or would it cost you money or a customer?
If you miss 2 calls a week and they're all existing clients who leave a message, a better voicemail greeting may be enough. If you miss 20 and most are new enquiries asking whether you have a slot on Thursday, an AI receptionist starts to make sense.
Step 2: Decide what the AI receptionist may and may not say
The NIST Generative AI Profile, a US government guide for organizations using this kind of AI, names a risk it calls confabulation: "the production of confidently stated but erroneous or false content". On a phone call that looks like a friendly voice quoting a price you don't charge, or promising a Saturday appointment you don't offer.
So write 2 short lists before you set anything up.
It may answer: your hours, your address and parking, which services you offer, how to book, and the next available slot (only if it's reading your real calendar).
It must hand over to a person: complaints, prices beyond your published list, anything medical, legal or financial, emergencies, and any caller who asks for a human.
Keep the "may answer" list short at first. You can grow it once you've heard how it handles the basics.
Step 3: Pick the plan by minutes and limits
With your Step 1 count in hand, you can estimate minutes. Calls a month, multiplied by the average length, rounded up. Add a margin, because the AI picks up calls that used to go to voicemail and those calls now count.
Then look at what happens when the minutes run out. This varies a lot. On BT's product page, overage is off by default, admins get an email at 80% and 100% of their minutes, and the service turns off at 100% unless you buy more or switch overage on with a spend cap. That's a sensible default. Ask any US vendor what theirs is.
Ask yourself:
- What happens to a caller when the minutes are used up? Do they reach voicemail, or nothing?
- Can you set a monthly spend cap?
- Can you cancel monthly, or are you locked into an annual bill?
- Does it connect to the calendar and CRM you already use? RingCentral's lists Salesforce, HubSpot and Zoho, and Zoom's lists Google, Microsoft and Calendly calendars.
- Does the vendor use your call data to train its AI? BT says it doesn't. If the page is silent, ask in writing.
Pause and reflect: If the AI receptionist went quiet at 3pm on a Friday because the minutes ran out, who would notice first? Make sure it's you, not a customer.
Step 4: Run a 2-week pilot on a slice of your calls
Larger organizations never move every call to a new system on day 1. They pilot with 1 queue, agree on success measures in advance and listen to recordings before expanding. You can do the same with the free trial.
Start small. Point only after-hours calls, or only the overflow when nobody picks up in 4 rings, at the AI receptionist. Before any customer hears it, make 10 test calls yourself and ask a friend to make 5 more, including at least 1 rude caller and 1 question it shouldn't answer.
Then score the 2 weeks:
| Measure | Before (Step 1) | After 2 weeks |
|---|---|---|
| Calls missed or sent to voicemail | ||
| Bookings made without you | ||
| Wrong or invented answers found in transcripts | ||
| Callers who asked for a human | ||
| Minutes used |
Read at least 10 transcripts a week. You're looking for the confident wrong answer from Step 2, and for callers who sounded frustrated.
Step 5: Write the rules on 1 page
Big companies keep a risk register: each risk, who owns it and what happens when it occurs. Here's the small-business version, adapted from the NIST Generative AI Profile, which lists confabulation, data privacy and over-reliance on AI among the risks to manage.
| Risk | Your rule | Who owns it |
|---|---|---|
| Wrong answers | AI answers only the "may answer" list. Transcripts checked weekly. | Office manager |
| Personal data | No card numbers or health details taken by the AI. | Owner |
| Over-reliance | A person returns every handed-over call the same day. | Whoever is on shift |
| Cost | Spend cap set, alerts go to 2 people. | Owner |
2 more points belong on that page. Tell callers they're speaking with an automated assistant, in the greeting. And recording and consent rules differ by state and by industry, so check with your lawyer before you switch recording on. Zoom says it offers a Business Associate Agreement for healthcare customers, but whether you need one is a question for your own adviser.
Step 6: Go live, then review every month
If the pilot scored well, widen it 1 step at a time: after-hours first, then lunch, then overflow all day. Keep your Step 4 scorecard and fill it in once a month. Update the AI's training whenever your prices, hours or services change, because it only knows what you've told it.
At the end of each month you have the same 3 choices as in the pilot. Keep it, fix it or switch it off.
Questions owners ask
Will callers know they're talking to an AI?
Tell them in the greeting either way. If you're worried that puts callers off, the pilot transcripts will show whether it does.
Do I have to change my phone number?
Usually not. BT says businesses keep their existing number and phone system, and Zoom says its standalone plan works with any phone system. Ask any other vendor the same question before the trial starts.
Can it take payments?
Keep payments with a person or a secure payment link for now. Card details over an AI call are a data risk you don't need in a first pilot.
Concluding thoughts: start with the count
An AI receptionist is easy to buy and quick to switch on. The 2 weeks of counting, scripting and listening are what make it safe for your customers. Go back to Step 1 and start your missed-call list this week.
If you'd like help, our setup service can configure and connect the receptionist to your calendar and CRM, and an audit can check what your current phone plan already includes before you add another bill.