A Real-Time AI Call Agent That Sounds Like a Person
Read the full call (timed transcript)
- 0:00CALLER with CC Docs. Am I speaking with the owner?
- 0:04AI Sam here.
- 0:11CALLER I will be quick. How many agents do you have on the floor right now?
- 0:16AI Uh, we've got five agents. Look, just send me the info.
- 0:24CALLER Got it. On a scale of 1 to 10, how happy are you with your floor's numbers?
- 0:30AI Eh, they're alright. Look, just send me the info. I'm in the middle of something here.
- 0:37CALLER What would it take to make that a 10?
- 0:41AI Look, how did you get my number anyway? Money's tight right now. Just send me the info and we'll see.
- 0:48CALLER That makes sense. We are not selling software. We bet we can lift your floor, and we prove it before you pay.
- 0:57AI Sounds interesting, but how does it work, and what do I have to do?
- 1:04CALLER Totally fair. This is not about more agents. It is about more sales from the agents you already have.
- 1:11AI Okay, but what exactly do I have to do for you to prove it works?
- 1:19CALLER Happy to send an email too. Can we also put 15 minutes on the calendar so I can walk you through it? Would tomorrow at 12 or 3 work better?
- 1:28AI Three works. Send the info first, and we'll talk then. Yeah. Okay. Bye.
Most people can tell a voice bot within a few seconds. It waits too long, it talks over you, or it sounds like it is reading. We wanted to show how close an AI call agent gets to a real phone call, so we recorded a test call and drew it out lane by lane.
What you are looking at
The player above has two lanes. The top lane is the caller: in this recording it is our own test caller, a robot we use to test the line, not a person. We cut one sentence of its sales pitch out of the recording; nothing else is edited. The bottom lane is our AI agent, speaking in a female voice. As the call plays, each lane lights up when that side talks and the words fill in as they are said. Under the player, each chip is one answer from the AI, with how long it waited after the caller stopped.
What we measured
In this test call the AI answered 7 times. The median wait was 1.3 seconds, the fastest was 1.2 seconds and the slowest was 2.1 seconds. We measured it from the sound itself: each side of the call is on its own channel, so the wait is the silence from the moment the caller’s channel goes quiet to the moment the AI’s channel starts speaking.
A separate measurement across 13 practice calls, timed from speech-to-text word times, gave a median of 1.76 seconds over 39 AI replies. One clean test call is a best case; a day of real calls is the number to judge.
Why the wait matters
On a phone call, a long pause after you speak is the first sign that nobody, or nothing, is really listening. Taking time out of that pause is what makes an AI call sound like a conversation instead of a recording.
The rules for AI calls
Under the FCC’s ruling on AI-generated voices, an outbound call that uses an AI voice needs the called person’s prior express consent. Any AI agent on outbound calls has to follow that rule.
Want to hear it on your calls?
If you run outbound or inbound calls and want an AI agent on them, contact us and we will walk you through it.