We put our own AI on our phone line. Here’s what it took.

By Samir Shukri, Founder · 5 min read

We sell AI that talks to customers. So it felt wrong that our own phone number went to a person’s cell and, after hours, to voicemail. This week we changed that: the Dendrites AI front desk, (641) 384-4222, is now answered by AxiomAI Voice, the same AI receptionist we set up for our customers.

Then we called it. A lot. On a handset, on speakerphone, in a hurry, in Arabic, halfway through another call. Demos are polite; real calls are not. Here is what the first day taught us, and what we changed.

1. It kept talking after you interrupted it

The AI produces speech faster than it is played, so several seconds of its answer were already queued on the phone line. When a caller cut in, the AI stopped thinking, but the caller kept hearing the old sentence. Now the queued audio is dropped the instant the caller speaks, and the AI is told exactly how much of its answer the caller actually heard, so it doesn’t assume they heard the rest.

2. A silent phone line is never silent

Phone systems send audio every twenty milliseconds, even when nobody is talking, so our “caller went quiet” timer never fired. We now measure real speech on both sides. After about fifteen seconds of nobody talking, the receptionist asks, “Are you still with me?” That line happened on a test call when we stepped away to take another call, which is exactly when you want it.

3. On speakerphone, it heard itself

One transcript showed the caller saying, “Thank you for calling. How can I assist you today?” That was the receptionist’s own greeting, coming back through a speakerphone. An AI that hears itself interrupts itself and loses its place. We switched the line to noise handling built for speakerphones and car systems, and made it a little more patient before deciding a caller has finished. Plenty of tradespeople call from a truck.

4. Names and emails are where calls go wrong

A caller’s name came through in the wrong alphabet. An email address was confirmed with a “yes” that was really a “can you hear me?”. It even offered to call back a made-up number, because nobody had told it the caller ID. Now it has the caller ID; it asks callers to spell email addresses, reads them back one character at a time, reads digits one by one (“two, zero, one, eight”, not “two thousand eighteen”), and when a caller corrects it, it asks again slowly instead of repeating its own version. The summary email uses the version the caller confirmed.

5. “Thursday” is not a date

Asked to set up a callback “Thursday at four”, it cheerfully confirmed Thursday, September 30. September 30 was a Wednesday. Language models are unreliable at calendar arithmetic, so we stopped asking them to do it: every call now starts with the next eight days written out, weekday and date, and the receptionist is told to use those and nothing else.

6. It must never say more than the business would

This is the one that matters most. An older instruction set on the line quoted plan prices that were out of date and described a compliance option we do not offer. Nobody meant it to; the words were simply in its instructions, and the AI repeated them confidently to a caller who said they ran a clinic.

Rewriting the script fixed that call. It did not fix the category. So the rules are now built into every call, whatever a business’s instructions say: no prices that aren’t written down, no certifications or compliance claims, no invented appointment times, discounts or guarantees. When the receptionist doesn’t know, it says so and offers a callback. A receptionist who is slightly less impressive and never wrong about your business is worth far more than the reverse.

What we didn’t change

Every call still opens with a plain notice: you are speaking with an AI assistant, and the call is recorded. It’s tempting to make an AI receptionist sound as human as possible and say nothing. We think callers deserve to know who, or what, they are talking to, and a business’s reputation is better served by being upfront.

Why we’re building this for the trades first

A plumber under a sink, an electrician on a ladder, a roofer on a roof: the phone rings at the worst moment, and a homeowner with a problem calls the next company that picks up. That is the job AxiomAI Voice is for. Answer, take the job down accurately, and put a clean summary in the owner’s inbox for the callback. Nothing clever beyond that, and nothing it wasn’t told to say.