Key takeaways
- Do the paperwork before the trial clock starts: the facts sheet, the transfer list and the test scripts, so all seven days go on testing.
- Script twelve test calls that sound like your real callers, including a hard name, background noise and a caller who switches language.
- Score every call on the same five marks and write the pass mark down before the first call.
- From day 3, forward one real slice of calls, usually after hours or overflow, and read every transcript the same day.
- On day 6, break it on purpose; on day 7, export what you need and decide before the trial turns into a bill.
Day 0: prepare before the clock starts
Most free trials are lost before the first call. The account opens, the clock starts, and three of the seven days go on hunting for the holiday hours, the fee you quote for a weekend visit and the mobile number of whoever takes urgent calls. Do that work the week before you sign up, and the trial becomes a test of the product instead of a test of your filing.
Plenty of owners are doing this for the first time. In Census Bureau survey data, AI use among US businesses hovered between 17% and 20% from December 2025 to May 2026[2]. There is no house playbook to borrow, so write your own on one page:
- Facts sheet: hours, holiday closures, address and parking, the services you offer and the ones you do not, and any fees you are happy to state on the phone.
- Booking rules: appointment types, how long each takes, how far ahead people can book and what needs a person to confirm.
- Transfer list: who takes which calls, who backs them up, and what the caller is told if nobody answers.
- Urgent definition: the exact situations that count as urgent for your business, written as a caller would describe them.
- Greeting: your business name, a plain statement that the caller is speaking with an automated assistant, and a recording notice.
- Trial terms: the end date, the billing date if a card is on file, how to cancel and what happens to recordings and the phone number afterwards.
The greeting deserves a minute of thought. In a 2025 Pew survey, 76% of Americans said it is extremely or very important to be able to tell whether content was made by AI or by people[3]. Say so in the first sentence. The recording notice matters for a different reason: federal law permits recording with one party's consent[4], but some states want every party told, and Washington, for example, treats a recorded announcement as consent[5]. Your callers do not all live in your state.
The seven days at a glance
Day 1
Load the facts, rules and greeting; place two calls yourself
Day 2
Run the twelve scripted test calls and score each one
Day 3
Fix what failed, then forward the first slice of real calls
Days 4 and 5
Read every transcript daily and fix causes, not symptoms
Day 6
Break it on purpose: failed transfer, busy hour, late night
Day 7
Score against the pass marks, export, decide before billing
Example schedule. Shift it if the trial starts mid-week, but keep the scripted calls before the real ones.
Day 1 should take an hour or two if day 0 was done properly. Paste in the facts sheet, set the booking rules and the transfer numbers, record or approve the greeting, and call the number twice yourself: once with a simple question and once to book something. If either call goes wrong, fix the information before day 2, because every scripted call after that depends on it. How an AI receptionist works explains which part of the system each kind of mistake comes from, which saves guessing.
Day 2: twelve test calls that sound like your callers
A demo is the vendor calling itself. Your test is people who sound like your customers calling about what your customers call about. Hand each script to a different person if you can, and have them call from their own phones.
| Test call | What the caller says | What a pass looks like |
|---|---|---|
| Routine question | What time do you close on Saturday? | The right hours, said once, then an offer to help with anything else |
| New booking | I'd like a first appointment next week, mornings are best | Booked into an open slot or taken as a clear request, details read back |
| Hard name | It's Siobhan Achterberg, A-C-H-T-E-R-B-E-R-G | Spelled correctly in the summary |
| Corrected number | My number is 555-0142, sorry, 555-0124 | Only the corrected number in the summary |
| Noisy line | Any request, made from a car or next to a television | Asks the caller to repeat rather than guessing |
| Slow, older caller | Long pauses, repeats the question | Waits, does not talk over, keeps the answers short |
| Language switch | Starts in English, switches to Spanish | Carries on in Spanish; the summary is still usable by your team |
| Off the list | Do you also do [something you don't]? | Says no or says it does not know, then offers a message |
| Wants a person | Can I just talk to someone? | Transfers or takes a message by your rules; no loop |
| Urgent | Your own urgent case, in a caller's words | Follows the urgent path you wrote down |
| Price question | How much do you charge for [service]? | Only the fees you approved; otherwise a message |
| Sales pitch | A vendor selling advertising | Polite, short, not booked and not transferred |
Brief testers lightly. Give each one the opening line and the outcome they want, not a full script, because real callers wander. Ask them to note the time of the call, then fill in the scoring card from the summary before you listen to the recording. That order matters: the summary is what your team will work from on a busy morning, so judge it first and use the recording only to find out why something went wrong.
The mix of voices matters as much as the mix of requests. Speech recognition does not hear everyone equally well: in one study of five commercial systems, the average word error rate was 0.35 for Black speakers against 0.19 for white speakers[6]. Language is the same story. Among people who speak Spanish at home, 39% speak English less than very well, and the share is higher for some other languages[1]. If your callers include people who would rather not do business in English, at least one tester should be one of them.
Older callers belong in the test too. About 15% of American adults report some trouble hearing, and disabling hearing loss rises to 22% of people aged 65 to 74[7]. Ask the slow-caller tester to say whether they could follow the replies, not just whether the booking was right.
Score every call the same way
Use one scoring card for every call, scripted or real, and write the pass mark down before day 2. Give each call 0, 1 or 2 points on five marks:
Outcome
Did the call end the way your rules say: answered, booked, transferred or messaged?
Details
Are the name, callback number, address and request right in the summary?
Honesty
Did it say plainly when it did not know, and never invent a price, a slot or a promise?
Handoff
Did urgent calls and requests for a person follow your transfer list, including the no-answer step?
Summary
Could someone on your team act on the summary without listening to the recording?
Example pass mark: twelve scripted calls at 10 points each is 120 points; a pass might be 100 or more, with no zero on any callback number and no zero on the urgent call. Two zeros on details is a stop sign whatever the total, because a wrong number turns a booked customer into a lost one. Call summary examples shows what a summary worth 2 points looks like.
Days 3 to 5: forward a real slice
Scripted calls find the obvious problems. Real callers find the rest. On day 3 forward one slice of your line: evenings and weekends, or calls nobody picks up within four rings. Leave everything else as it was, so regular customers in business hours notice nothing new. The real estate team example models a four-ring overflow setup for agents who are out at showings.
Watch the first few live. With an SI Employee, an owner can listen to a call in progress, whisper guidance the caller does not hear, or take over, which turns the first real calls into supervised practice rather than blind trust. After that, read every transcript the same day. Most early errors trace to one of four causes: a fact missing from the sheet, a rule that was never written down, a misheard word, or a wrong transfer number. Fix the cause, not the single call.
Keep a short log next to the transcripts: the date, what went wrong, the cause and the fix. By day 5 it should show fewer new problems and no repeats. A problem that comes back after you fixed the facts sheet usually means the rule was never clear, even to your own staff. Rewrite it in the words a caller would use.
One detail catches many owners: a forwarded call does not tell the system which ad, listing or sign the caller saw. If you track where calls come from, add a question about it to the script and check the summaries carry the answer.
Day 6: break it on purpose
The calls that hurt in month three are rare ones, so stage them now:
- Switch off the on-call phone and place an urgent test call. Check what the caller hears and what reaches you.
- Ring from two phones at once during your busiest hour and note what the second caller experiences.
- Change one fact, such as a holiday closure, and call an hour later to see it used.
- Call at 11 pm with a booking request and check where it lands the next morning.
- Hang up halfway through a call and see whether a summary still arrives.
- Caller
- SI Employee
- Your team
- System
01Caller
Caller reports urgent case
Uses your own urgent wording, such as water coming through a ceiling.
02SI Employee
Transfers to on-call phone
Follows the transfer list you wrote on day 0.
03Your team
Nobody answers
You switched the on-call phone off for the test.
04SI Employee
Caller is told plainly
No pretending someone is on the way; the details are confirmed.
05System
Urgent message saved
Flagged as urgent, with recording, transcript and summary for the team.
Example. If any step fails, fix the transfer list or the urgent rules and rerun the test before day 7.
Day 7: decide before the trial bills
Add up the scores, compare them with the pass mark you wrote on day 1, and decide on paper. Before you close the account or carry on, run through this list:
- Export anything you want to keep: summaries, transcripts and the contact list, which should come out as a spreadsheet.
- If you are not continuing, switch off call forwarding first. A line still forwarding to a closed account is worse than voicemail.
- If you are continuing, widen the slice one step at a time, such as adding lunch-hour overflow, and keep the scoring card.
- If the plan includes the system calling customers back, check consent first: the FCC has ruled that calls using AI-generated voices need the called party's prior express consent, absent an emergency[8].
- Note the date the trial converts to a paid plan, even if you have already decided.
The scores usually point one of three ways. Strong on details and handoffs but weak on outcomes or honesty: the information is thin, so fill the gaps and extend the test. Strong everywhere except languages or noisy lines: decide whether those callers are a big enough share of your business to matter. Zeros on callback numbers or the urgent call: stop, whatever the rest of the card says.
A failed trial is still useful if you kept the scores: you now know which calls your business gets and what a good answer sounds like. Before the next vendor, read the questions to ask an answering service and the 15 questions for choosing an AI receptionist. If the trial passed and you want a longer, money-based verdict, is an AI receptionist worth it sets out a two-week comparison against your current setup, and trades can borrow scenarios from the home services test-call list.
Frequently asked questions
Is seven days enough to judge an AI receptionist?
It is enough to find setup problems and judge accuracy on your own calls. It is not enough for a seasonal business to judge a busy month, so treat a pass as permission to keep testing a wider slice.
Should I forward all my calls during the trial?
No. Forward one slice, such as after hours or unanswered calls, so a bad day affects a small share of callers and you can compare against the hours you still answer yourself.
Who should make the test calls?
People who sound like your customers: different ages, accents and speaking speeds, and a speaker of any language your callers use. The owner calling alone tests very little.
What should I do with the recordings after the trial?
Export the ones you want before the account closes and ask the vendor in writing how long it keeps the rest and how to have them deleted.
What if the trial needs a card up front?
Write the conversion date and the cancellation steps on your day 0 sheet, and set a reminder for the day before, whatever you expect to decide.
References
U.S. Census Bureau (2022). Nearly 68 Million People Spoke a Language Other Than English at Home in 2019 (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) ↑
Pew Research Center (2025). How Americans View AI and Its Impact on People and Society (opens in a new tab) ↑
Legal Information Institute, Cornell Law School (2026). 18 U.S. Code 2511: Interception and disclosure of wire, oral, or electronic communications prohibited (opens in a new tab) ↑
Washington State Legislature (2026). RCW 9.73.030: Intercepting, recording, or divulging private communication; consent required (opens in a new tab) ↑
Proceedings of the National Academy of Sciences, via PubMed Central (2020). Racial disparities in automated speech recognition (PNAS 117(14)) (opens in a new tab) ↑
National Institute on Deafness and Other Communication Disorders (NIH) (2024). Quick Statistics About Hearing, Balance, & Dizziness (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) ↑