Where the problem comes from
A typical voice AI agent works on a simple rule: it hears sound, so the customer must have answered. If they answered, the current step is closed, so it moves on to the next one. On paper, that's logical. In a live conversation, it falls apart at the very first pause.
People make plenty of sounds during a call that aren't answers at all. Sounds like "mm-hmm," "yeah," "I hear you" are backchannel cues, signs that someone is listening, not responding. Then there's coughing, keyboard clatter, a door closing, a line from the TV in the background. For a first-generation AI agent, any of these sounds means the customer is done talking, so it moves on. The script collapses, the customer gets irritated, and the call is lost.
How the listening system works
Recognizes "mm-hmm," "yeah," "I hear you," and similar cues as active listening rather than a completed answer, so the AI agent keeps talking instead of cutting itself off.
Measures the share of its own response the customer actually heard before interrupting, then decides whether to finish the point or follow the customer onto a new thread.
The layer that tells listening from answering
The first layer's job is to catch the difference between an acknowledgment and an actual response. A "mm-hmm" in the middle of an explanation gets tagged as active listening. The AI agent calmly finishes its point, and the customer feels like they're actually being heard out. It's the same principle the best voice assistants rely on: a natural-sounding conversation starts with the ability to recognize backchannel cues.
We keep expanding the list of these sounds and phrases. Everyday spoken language is full of filler words, the equivalents of "uh," "so," "right, right," and every one of them gets added to this layer.
The layer that measures how much the customer actually heard
The second layer solves the opposite problem. The customer cut the AI agent off mid-sentence, in the middle of an important point, so how much of it actually got through?
This layer calculates the share. In one call pulled from our logs, the AI agent's response ran 800 characters, and the customer heard just 91 of them before interrupting, barely any of the point had landed.
What actually lands
In one logged call, the customer interrupted after hearing just 91 of 800 characters, about 11 percent of the point the AI agent was making.
From there, the AI agent decides: finish the interrupted thought, or start a new one, since a customer who interrupts is often changing the subject entirely.
If the customer heard nothing at all, the move to the next step gets blocked outright. The layer registers zero percent, the AI agent finishes its point, and only then continues. That's how an important argument reaches the customer in full, instead of in fragments.
A real person can tell when a thought isn't finished, and waits. Now the AI agent does the same.
What this means for the business
A conversation stops sounding like a survey read off a form. The AI agent hears the customer out, delivers its arguments in full, and never talks over someone mid-thought. Customers stay on the line longer and are more likely to follow through, whether that means booking an appointment, placing an order, or agreeing to a meeting.