A missed call in a service business is usually a lost job, because the caller rings the next number on the list within ninety seconds.
TL;DR Disclose that it is automated. Route emergencies to a human immediately, never to a bot. Capture and confirm rather than trying to fully resolve. Test it by calling it yourself, angrily.
Disclose it, always
Start here because it is both a trust issue and, increasingly, a legal one.
In the US, the FCC treats AI-generated voices as artificial or prerecorded for the purposes of the TCPA, which brings consent requirements, particularly for outbound calls. Rules differ by jurisdiction and are moving. Check what applies where you operate.
Practical position
- Say it is automated in the opening line. “Hi, this is the automated assistant for Sicc Heating.”
- Never impersonate a named person.
- Offer a human immediately. “Say ‘person’ at any point and I will put you through.”
- Inbound is far less fraught than outbound. Someone who called you is a different consent situation from someone you called.
People mind an automated system far less than they mind being deceived by one.
What it should actually do
Capture and route. Not resolve.
Good use
- Take the name, number, address and nature of the problem.
- Classify urgency.
- Confirm what it heard, back to the caller.
- Route or escalate.
- Text a confirmation to the caller.
- Log the whole thing with a transcript.
Bad use
- Quoting prices.
- Diagnosing faults.
- Handling complaints.
- Negotiating anything.
- Making promises about timing.
The test: could this response be wrong in a way that costs money or trust? If yes, a human handles it.
Emergency routing
This is where a badly configured system does real damage.
Build an explicit emergency path
- Keyword and phrase detection. Leak, flood, no heat, smell of gas, sparking, no power, elderly, vulnerable, baby.
- On detection, stop the script. Do not continue collecting structured information.
- Transfer to a live number immediately, or take the number and trigger an urgent alert.
- If nobody answers, say so honestly and give the emergency alternative.
That sounds urgent. I am putting you through to someone now.
Default to escalating. A non-emergency escalated costs you a phone call. An emergency handled as routine costs you far more than that.
Anything gas-related or electrical-safety-related should escalate unconditionally.
Call routing rules
Beyond emergencies, keep the tree shallow.
| Caller type | Route |
|---|---|
| Emergency | Human, immediately |
| Existing customer with an issue | Human during hours, urgent callback outside |
| New enquiry | Capture full details, book if possible, confirm by text |
| Existing customer, routine booking | Book directly, confirm by text |
| Supplier or admin | Message, normal callback |
| Sales calls | Politely ended |
Three options maximum at any prompt. Long menus are why people hate phone systems, and the assistant does not need them if it can understand plain speech.
Booking through voice
Works well for routine, standardised work. Poorly for anything requiring assessment.
If you enable booking
- Confirm every detail back before writing it in.
- Send a text confirmation immediately. Voice-only confirmation gets misheard and misremembered.
- Hold rather than commit for anything unusual, then have a human confirm.
- Do not let it book emergencies. Those need a person deciding.
Out of hours
This is where the return is clearest.
Overnight and weekend calls fall into three groups
- Genuine emergencies. Escalate to the on-call number.
- Non-urgent, will book in the morning. Capture and confirm. This is the majority.
- Wrong number and sales calls. Filter.
Group two is the win. Those callers currently ring you, get voicemail, and ring somebody else. Capturing them properly with a same-morning callback promise recovers a meaningful share.
Test it properly
Call it yourself, repeatedly, in the ways real customers call.
- Calmly, with a straightforward request.
- Angrily, mid-complaint. It should recognise this and escalate.
- Panicked, with an emergency. Time how long to a human.
- With background noise. A road, a busy house.
- With a strong local accent and local place names. Address recognition is where these systems most often fail.
- Interrupting it.
- Saying something completely off-script.
Then have somebody unfamiliar with it call. You know what it expects. Customers do not.
Review transcripts weekly for the first month. You will find failure patterns you did not anticipate, particularly around street names.
The honest limitations
- Accents and noise degrade accuracy meaningfully.
- It cannot read tone reliably. Frustration often reads as neutral.
- Local knowledge is absent unless you supply it explicitly.
- It will occasionally invent something. Constrain it to reading from your data rather than generating.
- Some customers will hang up, particularly older ones. That is a real cost against the calls you save.
Compare it against your actual current state, which is voicemail, not against a perfect human receptionist.
Measure it
- Missed calls before and after. The headline.
- Calls handled entirely without a human.
- Escalation rate, and whether escalations were appropriate.
- Booking rate from automated capture versus human answer.
- Hang-up rate in the first fifteen seconds. Rising means the opening is wrong.
- Emergencies missed. Must be zero. Audit this specifically.
Set it up for out-of-hours only to begin with, then call it yourself six times in six different moods before it takes a single real call.
Need a pro to configure it? [BOOK A CALL]