AI Receptionist: Our Own Phone Line

The problem

Seemab Akbar leads Nametech Canada's client work, and her phone rings while she is in meetings, on site or on a stage. A missed call from a business owner is usually a lost lead. An answering service can take a message, but it cannot book a meeting, it knows nothing about the business, and it cannot tell a returning client from a stranger. We wanted a receptionist that answers the way she would, books straight into her real calendar, and is honest with every caller about what it is.

What we built

We built the AI receptionist into NameCRM, the client system we run the agency on. When Seemab cannot pick up, it answers in a voice she approved, says it is her AI assistant and that the call is recorded, recognises callers who are already clients, books into her real calendar, suggests a meeting spot near the caller, and texts a recap the moment the call ends. Every call is filed to the client's record with a transcript and a list of follow-ups. It also places calls for the business, such as reservations and price checks, and reports back.

How a call goes

The line rings Seemab first. When nobody picks up within about 20 seconds, the receptionist answers, day or night, and it can hold up to five calls at once. If she taps Decline, the caller goes to the assistant instead of a voicemail robot.

Before it says a word, it reads the caller's record: open deals, recent texts, upcoming meetings and past calls. The greeting says what it is: "you've reached Seemab's AI assistant at Nametech Canada", and that the call is recorded.

A booking happens while the caller waits. The time is checked in code against her open hours before it is spoken, and a taken slot gets the next open one in the same breath. The meeting goes into her calendar with a Google Meet link and an invite, or, in person, at a vetted cafe near the caller.

After the call, each change request becomes one ticket and each promise said out loud becomes a task. A client who asks about their account gets their invoices and plan texted to the number on file, built by code so the AI never sees a figure.

Every call is an exam

Before a change reaches the phone line, it has to pass a practice caller: an AI that phones the line and plays a scripted situation, from an angry client chasing late website changes to a site hacked late at night. A judge scores each call out of 10 and cites the transcript line behind each finding.

When we counted on September 26, 2026, it had taken 257 practice calls across 46 situations since September 9, and failed 148 of them, because the standard is strict. Each failure became a written rule, 101 of them, most now enforced in code rather than hoped for in a prompt. More than 470 automated checks run before any change goes live.

What it will not do

It never quotes a price it was not given, never claims to have read an email it cannot see, and never promises a callback time that has already passed. Code checks each of these before a sentence is spoken, so the rule does not depend on the model remembering it.

It does not interrupt Seemab for sales calls, press or routine questions; it takes a message. It tries to get her on the line when a caller's site, store or email is down or they are losing orders.

Asked whether it is a real person, it says plainly that it is an AI assistant and offers a call back from her. On calls it places, it introduces itself the same way: Seemab's AI assistant, calling from Nametech Canada on a recorded line.

How we built it

Calls run through NameCRM's own audio relay, which links the phone line, speech recognition, Claude and the voice, and starts speaking from the first clause so replies begin in under a second. The voice is a professional clone of Seemab's own, approved by ear.

Order numbers, codes and emails are read one character at a time, and addresses are read back in parts and confirmed once, because the cloned voice garbled an email address on an early call. It understands callers in French, Hindi, Urdu and Punjabi, and replies aloud in French or Hindi.

In use

It has answered Nametech's line since September 5, 2026. Every call it takes is recorded on two channels, stored privately and summarized onto the client's timeline.

It also places calls for the business: it dials, gets through phone menus, waits on hold, never pays or commits to anything, and reports back with a verdict. We have sent it to book restaurant tables, to check whether a print order was ready for pickup and to ask shops for their prices.

Its data lives in NameCRM, hosted in Canada. The receptionist is sold only as part of NameCRM, because it needs the calendar, the client records and the tasks; the NameCRM product page covers the rest.