Some People Tell an AI Things They Would Never Tell a Person
What it actually takes to build a voice agent that sounds like it cares
Most people have a story about calling a company and feeling worse after the call than before it. The hold music, the menu trees, the agent reading from a script, the transfer to another department, the request to repeat the same information for the fourth time. The cumulative message is clear: you, the caller, are a ticket in a queue.
Now imagine that call is the hardest call of your life. You are the mother of a 22 year old who has been using for six years, and tonight is the night he finally said he would go. You have one window, maybe two hours, before he changes his mind. You dial the number on the website. You get a recording. You leave a voicemail. Nobody calls back.
This is the baseline that voice AI in behavioral health is being measured against, and it is a terrible baseline. The bar is so low that simply answering the phone in under two seconds feels like a miracle. But answering the phone is not the goal. The goal is to make the person on the other end feel like they reached the right place, talked to someone who heard them, and got what they needed without having to fight for it.
That is a much harder problem than it sounds.
The Comcast problem
Most voice AI on the market today, even the well-funded stuff, sounds like the worst version of the customer service we already hate. It interrupts. It misunderstands. It loops back to the same question when you give it an answer it did not expect. It says things like “I am sorry, I did not catch that” four times in a row until you start yelling into your phone. It is built by engineers who have never sat with a caller and watched the call happen, and it shows.
The problem is not the underlying technology. The models that power modern voice AI are remarkable. They can hold a natural conversation, change tone mid-sentence, recover from interruptions, and handle dozens of edge cases that would have been impossible three years ago. The problem is that almost nobody is taking the time to shape that technology into something that actually feels human on a hard call.
Building an agent that sounds like it cares is not a prompting exercise. It is a thousand small decisions about how the agent listens, how long it waits before responding, what it does when someone starts crying, how it handles a caller who is angry, how it knows when to stop talking and just let the person speak. It is the difference between an agent that says “I understand” reflexively and one that knows when silence is the right answer.
The choices that matter
When we built the first version of our admissions agent, the temptation was to make it sound smart. To prove the technology. To have it rattle off program details and verify insurance in one breath. We pulled almost all of that back. A good admissions counselor on her best day does not lead with a sales pitch. She asks how you are doing. She listens. She lets the caller set the pace.
So the agent does the same. It opens warmly. It asks one question at a time. It does not interrupt. When the caller pauses, the agent pauses too, instead of barreling ahead with the next prompt. When the caller goes off script and starts telling the story of how they got here, the agent does not redirect. It listens, because the story is the call. The information the admissions team actually needs comes out naturally inside the story, if you let it.
Get Samuel Staples’s stories in your inbox
Join Medium for free to get updates from this writer.
We also made a less obvious choice, which is that the agent never pretends to be a human. It introduces itself for what it is, an AI assistant for the admissions team, here to help you get to the right person or get you started right now. We thought this might create friction. It does the opposite. People relax. The expectation resets. They are not trying to figure out whether they are talking to a robot, and they are not embarrassed when they have to ask a clarifying question.
The thing no one expected
The most surprising finding from the last year of running this in production is that some callers are more honest with the agent than they would be with a human.
We did not design for this. We noticed it in transcripts. Callers disclosing substance use histories they had clearly never said out loud before. Callers admitting they had relapsed last week, that they had been using more than they told their family, that they were scared. Callers asking questions about confidentiality and detox and what withdrawal feels like, the questions you do not ask a stranger if you are worried about being judged.
There is something about talking to a non-human voice that lowers the cost of telling the truth. The agent is not going to call your mom. The agent is not going to remember you next week and look at you differently. The agent is not a person in a chair at a desk in a building somewhere who might be tired, or distracted, or making a face you cannot see. The agent is just a voice, asking questions, listening.
This is not a marketing line. It is a real shift in how the first contact with the treatment system can work, and it has implications we are still working out. A caller who tells the truth on the first call gets matched to the right level of care faster. A caller who feels heard on the first call is more likely to actually show up for intake. A caller who is not judged on the first call may make a second call when they otherwise would not have.
None of this happens by accident. It happens because someone made a series of deliberate choices about how the agent should behave, choices grounded in what it actually feels like to be the person calling.
What we owe the caller
The thing I keep coming back to is that the people calling an admissions line are not customers. They are not leads. They are people in one of the worst moments of their lives, often calling on behalf of someone they love, often after years of trying everything else first. The call they make is sometimes the only call they have in them that day.
Whatever voice answers that call has a responsibility that goes beyond business metrics. Voice AI in this space is going to be everywhere within three years. The question is not whether it gets deployed. The question is whether the people deploying it understand what they are deploying, and whether they take the time to build it right.
Building it right looks like sitting with admissions teams and listening to real calls. It looks like rejecting a feature because it would make the agent feel less human, even if it would make the demo more impressive. It looks like obsessing over the moment the caller starts crying, and what the agent should do in the next three seconds, and the three seconds after that.
The technology to answer every call in under two seconds, twenty four hours a day, finally exists. The harder work is making sure the call that gets answered is the kind of call you would want your own family to make.
That is what we built DIAL3D to do, and we are here for the long haul, to consult, to guide, to tune, and to keep showing up every day for the operators we work with. The mission is the same on both sides of the call: lift the bottom line for the facility, and get more people the help they came looking for.
Samuel Staples is Co-Founder and CTO of DIAL3D, a HIPAA-aligned voice AI platform built for behavioral health and addiction treatment. dial3d.ai
