12 Checks Before Any AI Talks to Your Customers

Why this matters

Air Canada: its website chatbot invented a refund policy. A tribunal made the airline pay the difference (2024). If your bot promises it, you own it.
A plumber's AI receptionist: a homeowner in the middle of a remodel called to book his plumber. The AI couldn't understand his address and had no way to reach a human. He hung up and hired another plumber. The first plumber lost a five-figure job and found out from the call log three weeks later. (One homeowner's account, posted online.)
DPD, a UK parcel company: a customer trying to track a missing parcel got its chatbot to swear and call DPD "the worst delivery firm in the world" (2024). DPD switched the AI part off.

AI can help you answer faster. Set it up wrong, and it loses jobs quietly.

It says it's AI, and that the call is recorded. "Hi, I'm the virtual assistant for [Company]. This call is recorded." Hiding it is what makes people angry, and some US states need everyone's consent to record.
A human is one step away. "Press 0" or "reply HUMAN" reaches a person, or a callback within a set time.
It never quotes a price. At most, it gives the range you wrote.
It never promises dates, refunds or warranties.
Emergencies go straight to a person (leak, no heat, no AC, safety).
It knows your service area, hours and services, and says no politely to jobs you don't do.
It reads back the name, phone number and address before it saves them.
Every conversation lands in your CRM or booking software, with a summary.
Someone reads the log every day for the first 30 days, then weekly.
It can spot spam calls so they don't eat your time or your minutes.
Start with after-hours only. Add daytime overflow once it's proven.
Measure it against voicemail, not against perfection: calls answered, leads captured and jobs booked, before vs after.

"Show me results from a named rollout, not the headline number." One AI sales tool advertises 36% higher close rates. Another tool, Rilla, rolled out across Neighborly's 500+ home-service locations and reported +1.9% close rate and +13% average ticket (the two companies' own figures, 2025). Different tools, different companies. But notice the gap.
"Can I hear 5 real calls, including one that went wrong?"
"How does a caller reach a human, and how fast?"
"What happens when it doesn't understand an address or an accent?"
"How long is the contract, and how do I cancel? Can the price go up without notice?"
"Who owns the recordings and the customer data?"
"Does it tell callers it's AI?" Several US states, including California, Utah and Maine, require bots to say they're AI in some situations, and more are debating it. Disclose anyway.
"Does it make outbound calls?" In the US, AI voices count as "artificial" voices under the phone rules (FCC, 2024). So outbound AI calls need the person's prior consent, and written consent if it's marketing. Texts have their own consent rules too.

Let AI draft, sort, answer the simple stuff and chase. A person handles money, promises and anything with your name on it.

Facts and rules checked October 2026. This isn't legal advice: texting, recording and review rules differ by state and country.