Key takeaways
- Judge each feature by the call it finishes. If it does not change how a call ends, it is decoration.
- Six features are must-haves: your schedule, approved answers, booking or clear requests, a way to a person, a record of every call and live control.
- The way out to a person matters most; regulators have documented callers stuck in chatbot doom loops with no offramp[2].
- Ignore human-sounding voices, logo walls of integrations and sentiment scores until you see them work on your own calls[3].
- In the auto shop example, calls finished without an advisor are worth about 31 hours of desk time a month.
Sort features by the call they finish
Feature pages for AI receptionists run to dozens of lines, and most of them describe the same handful of jobs in different words. The way through is to stop reading the list and look at your calls instead. Every call to a service business ends one of four ways: answered, booked, passed to a person, or turned into a message. A feature matters if it changes which of those endings a caller gets, or how much work the ending leaves for your staff. Everything else is presentation.
Answered
Approved facts, plus a plain "I don't know" when the answer is not on the sheet
Booked
A connected calendar, or a request with a preferred time that staff confirm
Passed to a person
A transfer list with backups and a tested step for when nobody picks up
Taken as a message
A summary, recording and transcript in the right inbox, urgent calls marked
Use the four endings as the columns of your vendor sheet. A feature that moves no call into a better column is presentation, not capability.
Callers arrive with expectations already set. The Consumer Financial Protection Bureau counted over 98 million US users, roughly 37% of the population, who engaged with a bank's chatbot in 2022, with growth projected to 110.9 million by 2026[1]. The same report describes customers caught in doom loops of repetitive answers with no offramp to a human, usually when their problem fell outside what the bot could do[2]. On the phone, that experience ends in a hang-up and a call to your competitor. The features that prevent it are the ones to buy.
Source: Consumer Financial Protection Bureau, 2023[1]
Show the data
| Item | Value |
|---|---|
| 2022 | 98 million users |
| 2026 (projected) | 110.9 million users |
Must-have: six features that change how calls end
| Feature | Why it matters | Five-minute check |
|---|---|---|
| Answers on your schedule | Covers the hours you choose: all day, after hours or only unanswered calls | Forward a test line and call it in each mode |
| Answers from facts you approve | Callers get your hours, services and policies, not a guess | Ask something that is not on your facts sheet |
| Booking or a clear request | The caller leaves with a time, or knows exactly who will confirm | Book a slot, then find it in the calendar; ask for a full one |
| A way to a person, with a fallback | Urgent and upset callers reach staff, and nobody is left hanging | Ask for a person with the transfer phone switched off |
| A record of every call | Recording, transcript and summary in the right inbox, urgent calls flagged | Hang up halfway and check what still arrives |
| Live control | You can listen in, guide the call or take it over | Listen to a test call and take over mid-sentence |
Two of these deserve more attention than vendors give them. The first is the way to a person. Ask what happens at each step when the transfer is not answered: what the caller hears, whether the call comes back, and what reaches you afterwards. Then ask what your staff member hears when they pick up. Some products brief staff before connecting and call that a warm transfer; others simply connect the caller. Either can work, but you need to know which, so your team can open with the right line. Warm transfer scripts has wording for both.
A third detail sits under the transfer feature. When your main line forwards to the receptionist, a transfer cannot ring back on that same line, because it is the line that sent the call away. Put direct lines or mobile numbers on the transfer list, and ask the vendor to show a transfer to each one during the demo. Example: a shop whose only number is the front-desk line needs the advisors' mobiles, or a second line at the counter, before transfers can work at all.
The second is the record. A summary someone can act on without playing the recording is worth more than any dashboard. Check that it names the caller, the callback number, the reason in plain words, what was promised and what is urgent, and that it lands in the inbox of the person who acts on it. Call summary examples shows what good looks like.
One more check belongs with the greeting. In a 2025 Pew survey, 76% of Americans said it is extremely or very important to be able to tell whether something was made by AI or by people[4]. A product that lets you open with a plain disclosure and a recording notice, in words you approve, is doing you a favor.
Nice to have, if your calls need it
These features earn their place for some businesses and do nothing for others. Check each one against your own call log before it sways a decision:
- Other languages. The number of US residents who speak a language other than English at home reached 67.8 million in 2019, and Spanish accounted for 62% of them[5]. If your callers switch languages, test that the conversation follows and that the summary is still in a language your team reads.
- Webhooks and live API actions. A signed webhook can send every call to your CRM, directly or through tools such as Zapier, Make or n8n. An API action can look something up in your own system during the call, such as an order status. Ask exactly which fields are sent and when.
- Several numbers and locations. Separate greetings, facts and transfer lists per location, so the downtown caller is not sent to the suburban office.
- Contacts with an export. A contact list built from calls is useful; one you can download as a spreadsheet is useful for longer.
- Outbound callbacks. Useful for returning missed calls, but the FCC has ruled that calls using AI-generated voices need the called party's prior express consent, absent an emergency[6]. Ask how the product records that consent.
- Texts to callers. Confirmations and missed-call texts can help, but they need consent and carrier registration. Ask whether texting is live on accounts like yours today or still on a roadmap.
Integrations deserve one specific question: is this connection live and tested on an account like mine, and does it work through a webhook, through my own API, or as a one-click connector? Vague answers here are common. If booking into your own scheduling software is the reason you are buying, the guide to an AI receptionist that books appointments covers what to test.
Ignore until proven on your calls
Some features sell well in a demo and change nothing about how your calls end. The FTC has said there is no AI exemption from the laws against deceptive claims, and in 2024 it announced five actions over AI claims in one sweep[3]. Treat these as unproven until you hear them on your own calls:
- Sounds exactly like a human. A clear, well-paced voice matters. Passing for a person does not, and it works against the disclosure most callers want.
- Integrates with hundreds of apps. A logo wall usually means one general-purpose connection. Ask about the one system you use.
- Sentiment scores and analytics dashboards. A mood score per call is rarely acted on. A good summary and a weekly read of transcripts usually are.
- Handles every call without staff. Some calls should reach a person, and a product that claims otherwise has not thought about your emergencies.
- Takes card payments on the call. PCI rules prohibit keeping card security codes in recordings that can be searched[7]. A payment link sent afterwards is usually safer.
- Speaks dozens of languages. Test the two your callers use, not the list.
A useful rule in demos: for every feature that makes a difference to your decision, ask the vendor to show it on a live call, with your business facts, while you listen. If it can only be shown in slides, it does not go on your must-have list.
What finishing calls is worth: an auto shop example
Example: an independent repair shop with six bays and two service advisors gets about 45 calls a weekday. Many are routine: opening hours, booking a drop-off, which makes the shop works on, the price of an oil change the owner lists. With the must-have features in place, the shop expects a little over half of those calls to finish without an advisor: answered from the facts sheet, booked into drop-off slots the shop opens, or, where the shop's own system can answer through an API action, told whether the car is ready. Diagnosis questions and unhappy customers still go to an advisor.
| Input | Example |
|---|---|
| Calls per working day | 45 |
| Minutes per call | 3.5 |
| Share of calls finished without an advisor | 55% |
| Employer cost per staff hour | $36.73 |
| Calls a month | 977 |
| Staff hours on these calls | 57 |
| Hours handed off | 31 |
| Monthly value of that staff time | $1,151 |
How it is worked out: Calls a day × 21.7 working days × minutes per call ÷ 60 × share handled = hours; hours × hourly cost.
Assumptions
- Employer cost of $36.73 per hour worked, the BLS June 2026 average for office and administrative support workers including benefits.[8]
- The 45 calls a weekday, three and a half minutes per call and 55% finished without staff are invented for this example; use your own log.
Example only: about 31 hours of advisor time a month, worth roughly $1,150 at the BLS cost per hour. Change the inputs to match your shop.
Thirty-one hours is close to four advisor days a month spent at the counter with customers instead of on the phone. Notice which features produce it: approved answers, booking and, for status calls, a live lookup in the shop's own system. None of the ignore list contributes. The bigger number is often the calls that never got answered at all, because eight in ten Americans say they generally do not pick up calls from unknown numbers[9], so a returned call often goes nowhere. The auto repair service desk example models both halves for a shop with morning and evening rushes.
Score vendors on one sheet
List your ten most common calls
From a week of your log, not from memory. Note how each one should end.
Map each call to the features it needs
A drop-off booking needs booking and a record; a breakdown on the highway needs a way to a person.
See every must-have live
Ask each vendor to show all six on a call with your facts loaded, while you listen.
Test on your own calls
Run scripted calls and a slice of real ones; the seven-day free trial plan sets out the days.
Check the exit
Export of calls and contacts, who controls the phone number and how cancelling works.
For reference, an SI Employee covers the six must-haves today: forwarded calls on your schedule, answers from the facts you approve, booking into a connected calendar or a clear request, transfers with a no-answer fallback, a recording, transcript and summary for every call, and live listening, whispering and take-over. It connects to other systems by webhook or through your own API rather than one-click connectors. For the wider buying decision, see the small business buyer's guide and AI receptionist pros and cons; contractors can also use the home services evaluation checklist.
Frequently asked questions
Which AI receptionist feature matters most?
A reliable way to a person, including what happens when nobody answers the transfer. It decides how your hardest calls end, and callers forgive a lot if they reach someone when it counts.
Do I need integrations with my existing software?
Only if a call has to create or read something in that software to finish. Many businesses start with summaries by email and a webhook to their CRM, and add deeper connections once the basics work.
Is a more human-sounding voice worth paying for?
Clarity and pacing matter; sounding human does not. Most callers want to know when they are talking to an automated system, so a clear voice with an honest greeting is the better buy.
How do I check features without a long trial?
Ask the vendor to show each must-have on a live call with your facts loaded, then place your own test calls. The five-minute checks in the table above cover the six that matter most.
Should an AI receptionist take payments?
Usually not on the call. Card security codes must not be kept in searchable recordings[7], so a payment link sent after the call is the safer setup for most small businesses.
References
Consumer Financial Protection Bureau (2023). Chatbots in consumer finance (opens in a new tab) ↑
Consumer Financial Protection Bureau (2023). Chatbots in consumer finance: Failure to provide meaningful customer assistance (opens in a new tab) ↑
Federal Trade Commission (2024). FTC Announces Crackdown on Deceptive AI Claims and Schemes (opens in a new tab) ↑
Pew Research Center (2025). How Americans View AI and Its Impact on People and Society (opens in a new tab) ↑
U.S. Census Bureau (2022). Nearly 68 Million People Spoke a Language Other Than English at Home in 2019 (opens in a new tab) ↑
Federal Communications Commission (2024). Declaratory Ruling FCC 24-17: Implications of Artificial Intelligence Technologies on Protecting Consumers from Unwanted Robocalls (opens in a new tab) ↑
PCI Security Standards Council (2011). Information Supplement: Protecting Telephone-based Payment Card Data (opens in a new tab) ↑
U.S. Bureau of Labor Statistics (2026). Employer Costs for Employee Compensation, Table 4: private industry workers by occupational and industry group, June 2026 (opens in a new tab) ↑
Pew Research Center (2020). Most Americans don't answer cellphone calls from unknown numbers (opens in a new tab) ↑