Key takeaways
- Write the pass marks down before the test starts, so your own calls decide the answer rather than a demo.
- Week one measures your current setup; week two puts an AI receptionist on one slice of calls, usually after-hours or overflow.
- Callbacks recover less than owners expect, because most people let unknown numbers ring[1].
- Value the difference as recovered new-customer work plus staff time; in the plumbing example the work alone is about $8,700 a month.
- It is not worth it if you rarely miss calls, your real bottleneck is capacity, or nobody will own the rules.
Decide what worth it means before you start
You can settle whether an AI receptionist is worth it with two weeks of your own calls and one sheet of paper. Week one measures what happens now. Week two puts the software on a slice of the calls and measures the same things. Write the pass marks down before week two starts and the answer is waiting for you at the end, whichever way it goes.
The usual shortcut is to decide on a demo. Demos are built to go well. Your Tuesday at 12:15, with two people at lunch and a delivery at the back door, is not. The test below takes about three hours of your time spread across two weeks, and most of that is reading transcripts.
Most businesses have not made this call yet either. In the Census Bureau's survey, between 17% and 20% of US businesses used AI in any business function from December 2025 to May 2026, and 20% to 23% expected to within six months[2]. A written test means you neither buy on hype nor rule it out on a hunch.
Week one: measure what happens now
Export the call log
Every call for seven normal days, with time, duration and outcome: answered, missed or voicemail. Avoid holiday weeks.
Tag each missed call
New customer, existing customer, supplier or spam. Play the voicemails and check numbers against your customer list.
Track the callbacks
For each missed new customer, note whether someone called back, how long it took and whether they reached the caller.
Log staff phone time
For two days, whoever answers keeps a tally of calls taken and rough minutes per call, including the notes afterwards.
Count what booked
How many new-customer calls ended in a booked job or appointment, and how many ended in a promise to call back.
The callback column is usually the eye-opener. In Pew's 2020 survey, 80% of US adults said they generally do not pick up when an unknown number calls their cellphone[1], so a missed call returned two hours later often turns into voicemail tag. If some calls come from Local Services Ads, tag those separately: Google's help page counts responsiveness among the ad auction factors and says missed calls may count against it[3]. This tagging exercise is a missed call audit in miniature, and it is worth keeping whatever you decide.
Some customers decide on the first conversation. In Zillow's 2026 buyer survey, 54% of buyers who contacted an agent hired the first one they contacted, up from 47% in 2023[4]. Real estate is one market, but the logic carries to any business whose callers can see three other numbers on the same search page: the first business to hold a real conversation has the advantage.
Source: Zillow Research, 2026[4]
Show the data
| Item | Value |
|---|---|
| 2023 | 47% |
| 2024 | 47% |
| 2025 | 50% |
| 2026 | 54% |
Before you close week one, split the missed calls into two piles: those during opening hours and those after closing. They have different causes. Daytime misses are usually a capacity problem at predictable moments, such as the first hour of the morning or the lunch overlap. After-hours misses are a coverage problem: nobody is there at all. The test in week two should target whichever pile holds more new customers, because that is where a result will show up fastest.
Week two: put it on a slice of calls
Do not hand over the whole phone line on day one. Pick the slice where the gap is largest in your week-one data, usually evenings and weekends, or overflow after four unanswered rings. Leave everything else exactly as it was, so the comparison is fair and your regular customers notice nothing new during business hours.
- Your team
- System
- SI Employee
01Your team
Week 1: baseline log
Tag every call; log callbacks and staff phone time.
02Your team
Write rules and pass marks
Hours, booking rules, transfer list and emergencies; pass marks on paper.
03System
Forward one slice
After-hours and no-answer calls only; the daytime routine stays the same.
04SI Employee
Handles the slice
Answers, books, transfers or takes a message, and logs every call.
05Your team
Daily 15-minute review
Read transcripts, fix instructions, note each error and its cause.
06Your team
Score against pass marks
Same measures as week one; then expand, fix or stop.
Example plan for a small service business. The slice and the pass marks should come from your own week-one log.
Decide in advance what stays with people even inside the slice. Emergencies should go straight to whoever is on call, with the system taking the address and a callback number first. Complaints and billing disputes should be transferred or taken as a detailed message. Example: for a plumber, no hot water since this morning can be booked into the next open slot, while water coming through the kitchen ceiling goes to the on-call phone at once. Write those lines down before the first forwarded call, not after the first bad one.
The daily review is where most of the value comes from. Read every transcript for the first three days. Each error has a cause: the call never arrived, a word was misheard, the request was misread, the business information was wrong, or the reply was badly timed (how an AI receptionist works explains each stage). Fix the cause the same day. By day five the errors should be rarer and different, which is a result in itself. The restaurant guide to testing before going live has scripted test calls worth borrowing for the first morning.
If you are working inside a vendor's trial week rather than your own two weeks, keep the same scorecard and put question 14 of the vendor checklist to them first: can you forward a slice of real calls and get full transcripts back? Either way, call it yourself before your customers do.
What counts as a pass
Write these down before week two starts. The marks below are examples for a typical service business; tighten them if your calls carry more risk, such as medical symptoms or legal deadlines.
| Measure | How to count it | Example pass mark | Warning sign |
|---|---|---|---|
| Answer rate | Calls in the slice answered by the system or a person | Every forwarded call answered | Calls still landing in the old voicemail |
| Correct outcome | Booked, transferred or messaged the way your rules say | At least 9 in 10 handled calls | Wrong appointment types, missing questions |
| Detail accuracy | Names, numbers and addresses right in the summary | No uncaught error in a callback number | Staff ringing callers back to check details |
| Staff time | Calls staff did not have to touch, times minutes per call | Hours back that you can name | Staff rewriting summaries |
| Caller reaction | Hang-ups in the first ten seconds, complaints | No worse than the voicemail baseline | Repeated complaints about not reaching a person |
| New business | New-customer calls in the slice that ended booked | More than in week one | Fewer bookings than before |
Two of these matter more than the rest. Detail accuracy, because a wrong callback number turns a booked job into a lost one with extra steps. And caller reaction, because a setup that answers every call but irritates people is not a pass. Compare early hang-ups with your voicemail baseline rather than with perfection: plenty of callers hung up on your voicemail too.
Turn the difference into money
With both weeks logged you can value the gap. Start with the new-customer calls that went unanswered in week one, because those are what week two recovers. The example below is a fictional three-truck plumbing and drain company whose missed calls cluster at lunch and after 5 pm.
calls in a normal week
rang out or went to voicemail
of those were new customers
would likely have booked if answered
Example figures, rounded from the inputs of the calculator below.
| Input | Example |
|---|---|
| Calls per week | 90 |
| Calls that go unanswered (%) | 18% |
| Unanswered calls from new customers (%) | 45% |
| Share of answered new callers who book | 55% |
| Typical job value | $500 |
| Unanswered calls a month | 70 |
| From new customers | 32 |
| Bookings that would have followed | 17 |
| Monthly value at risk | $8,681 |
How it is worked out: Calls a week × 4.33 weeks × unanswered share × new-customer share × booking rate × average value.
Assumptions
- Typical job value of $500: the American Housing Survey's median homeowner spend for plumbing fixture work, 2021 to 2023.[5]
- No credit is given for callbacks, because most adults let calls from numbers they do not recognize ring out.[1]
- The 90 weekly calls, 18% missed, 45% new customers and 55% booking rate are invented for a three-truck shop; replace them with your week-one log.
Example only: roughly $8,700 a month of work sits behind unanswered new-customer calls in this model. Put your week-one figures in to see your own.
Then add staff time. Example: if week two shows the system handled 12 calls a day that staff would otherwise have taken, at about four minutes each including notes, that is roughly 17 hours a month. At the BLS average of $36.73 an hour in total compensation for office and administrative support workers[6], those hours are worth about $640 of staff time a month, which in a plumbing office goes back into dispatch, quotes and chasing invoices.
Add the two together and set the total against the quote you get from the vendor. If recovered work plus staff time comfortably beats it in a two-week sample, the case is strong. If it only wins on the most optimistic assumptions, it is not worth it yet. The HVAC peak-season example runs the staff-time half of this arithmetic for a twelve-truck company in a hot summer, and the home services guide to after-hours coverage payback works it through for evening calls.
When it is not worth it
Some tests fail, and some businesses should not bother running one. An AI receptionist is probably not worth it when:
- You already answer nearly every call live. If week one shows almost no missed calls and nobody is stretched, the gain is small.
- Almost every call needs a licensed judgment. If most callers to a specialist clinic or a complex legal practice need a professional, a front end that only takes messages adds a step.
- Callers need to share what you cannot pass to a vendor yet. Before patient information flows through an outside service, HIPAA requires a practice to obtain satisfactory assurances, documented in a written agreement with that business associate[7]. Without one, keep the setup to hours, directions and general questions.
- Your bottleneck is capacity, not calls. If the schedule is booked six weeks out, answering more calls builds a longer waiting list, not more revenue.
- Nobody will own the rules. Without an hour a week to read transcripts and update hours, services and policies, accuracy decays.
Sometimes the cheapest fix wins. If week one shows most missed calls fall in one predictable hour, staggering lunch breaks may recover more than any software. Example: a part-time person covering noon to 1 pm, five days a week, works about 22 hours a month; at the BLS median receptionist wage of $18.27 an hour[8] that is roughly $400 in wages before payroll costs. If that hour is your whole gap, the schedule change is the better buy.
Reading the result
At the end of week two you will land in one of three places. Pass on answering and accuracy: widen the slice to daytime overflow and rerun the scorecard for another week. Pass on answering but fail on accuracy: the cause is usually stale information or missing rules, so fix them and retest before deciding. Fail on caller reaction or on bookings: stop, and keep the call log. You now know your missed-call number, which is useful whatever you do next.
For the trade-offs you will meet after a pass, read AI receptionist pros and cons. If the test goes well and you want the wider picture of what an SI Employee takes on beyond answering, such as booking, follow-up texts and contact records, start with the small business buyer's guide.
Frequently asked questions
How many calls do I need for a fair test?
Enough that the slice produces a few dozen handled calls over the week. Below that, one odd call swings the result, so extend the test rather than deciding on a handful.
Should I tell callers it is a test?
There is no need to mention a test, but do introduce the system honestly and offer a person. Callers judge the experience, not your experiment.
What if my results are borderline?
Run a third week on the same slice after fixing the errors you found. If the numbers do not move, treat that as a no for now and keep the call log for later.
Is two weeks enough for a seasonal business?
Run it in a busy week, not a quiet one, and recheck at the next season change. A February test tells an air conditioning company very little about July.
What is the most common reason a test fails?
Stale or missing business information: old hours, unlisted services, no rule for emergencies. These show up in the transcripts within days and are fixable before you judge the product.
References
Pew Research Center (2020). Most Americans don't answer cellphone calls from unknown numbers (opens in a new tab) ↑
U.S. Census Bureau (2026). Large Firms With at Least 20 Employees Biggest AI Users (opens in a new tab) ↑
Google (2026). About ad rankings (Local Services Ads Help) (opens in a new tab) ↑
Zillow Research (2026). Buyers: Results from the Consumer Housing Trends Report 2026 (opens in a new tab) ↑
U.S. Census Bureau and U.S. Department of Housing and Urban Development (2023). Home Improvements (2023 American Housing Survey infographic) (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) ↑
Legal Information Institute, Cornell Law School (2026). 45 CFR 164.502: Uses and disclosures of protected health information, general rules (opens in a new tab) ↑
U.S. Bureau of Labor Statistics (2025). Receptionists: Occupational Outlook Handbook (opens in a new tab) ↑