We Replaced a Recruiting Call Center With an AI Agent. Here's What It Actually Does.

A national door-to-door sales company hires a lot of people for an entry-level sales role.

People apply online by the thousands, and every one of those applications used to roll into a call center. Then a human agent would text the applicant, walk through the same prerecorded script, ask the same questions, and if the person qualified, try to get them onto an interview calendar. Thousands of conversations, but really it was one conversation over and over again.

Now an AI agent does that job.

Not a little chatbot glued onto a form. I mean a system that makes first contact, qualifies the recruit, and gets them scheduled for an interview — at a local office if there’s one nearby, or remote if there isn’t.

I built it with my team. I didn’t plan the project, and some infrastructure choices were already in place before I showed up, but I ran the implementation, built the tools, and built the system that’s doing the work now. After 5,000 recruits, here’s what it does, what the numbers look like, and the main thing I’d want someone to notice if they run a business with this kind of conversation hiding somewhere inside it.

The job, before

The call center’s job had three parts:

1. Make contact. Someone applies, and someone has to text them quickly enough that they still remember applying.
2. Qualify. Ask a fixed set of questions. Are they eligible, available, and in a place where we can actually place them.
3. Book the interview. If they pass, get them onto the right calendar for their area.

There wasn’t much relationship in that loop. The script existed because the conversation was basically identical every time. That ends up mattering.

The job, now

The agent runs that same loop from start to finish:

• First contact goes out over SMS, same as before.
• Qualification happens in conversation. The agent asks, reads the answer, writes what it learned back to the recruit’s record, and moves to the next question — or ends the conversation politely if the person is disqualified. Part of that process is location: if we don’t have a ZIP code, the agent looks it up, then checks whether the client has a local office there.
• Scheduling also happens in conversation, same as the human process used to. A tool connects to the client’s calendar, so the agent can book the interview itself — local office or remote — and handle reschedules when plans change.

Every step writes back to the recruit’s record, so the humans downstream can see exactly where someone is in the funnel: contacted, qualified, scheduled, no-show. Same funnel as before. The call center just isn’t stamping it by hand anymore.

At this point it’s automated and hands-off. We’re still watching it closely, but right now it’s running at about a 0.5% error rate while we keep polishing the machine.

The numbers, after 5,000 recruits

Here are the numbers I can actually own, and which ones I can’t.

• Book ratio: about 74%. Out of the recruits the agent handles, roughly 74% end up scheduled. That’s 7 percentage points better than the human team using the same script. The sample is 5,000 recruits. That’s enough for me to take it seriously, but not enough for me to say the agent is “better than people.” It’s 7 points.
• Cost per conversation: about $0.60 at launch, about $0.40 now. By conversation, I mean the full thread with one recruit from start to finish. How we got that cost down is its own story, and I’ll write it up separately.
• What I don’t have: the call center’s cost. That wasn’t my department, and it’s a client situation. What I can say is the work that used to go into those conversations is likely being reallocated, because the agent is doing that job now. I’m not going to slap a savings number on this if I can’t back it up.

The point isn’t that the robot won

This is the part I actually care about.

The agent didn’t outperform the call center because it’s somehow wiser or more human than the people doing the work. It did better because the job had already been squeezed down into a script.

When a conversation is formulaic — same questions, same order, no real relationship being built — the person running it is doing work a machine can do. And usually they know it. That’s exactly the kind of conversation an agent is built for.

What that creates for the company isn’t automatically fewer people. It’s better use of people. That can help the bottom line, sure. It can also mean smaller training cohorts or better care once someone is hired. The human recruiters — the ones who actually influence whether a strong candidate joins — get to spend less time on the front-end script and more time with qualified recruits. The repeatable part gets cheaper. The human part gets more humans.

I’d put it like this: agents aren’t here to replace human connection. They’re here to create more of it, if we want, by taking over the conversations that were never really connection to begin with.

If you’ve got one of these conversations

You probably do.

Look for the spot in your business where someone is basically working from a script: intake, qualification, scheduling, that first reply every inbound lead gets. My test is simple: if the conversation would play out mostly the same no matter who handled it, it’s a candidate.

From this build, here’s what it took to hand that work to an agent:

• Write the script down honestly. The agent is only as good as the questions, rules, and disqualifiers you can clearly define. In our case, the call center’s script became the spec.
• Give it somewhere to write. Every answer has to land in a record the humans can see. The agent’s job stops at “scheduled.” Somebody else’s job starts there, and they need context.
• Give it the tools a human would use. Ours can look up a ZIP code, check for a nearby office, and book directly on the client’s calendar. Without that, it’s just a chatbot that politely stops right before the useful part.
• Watch it while it’s new. Ours is hands-off now, but we still monitor it. It’s sitting around a 0.5% error rate today, and we’re still polishing it.
• Measure it the same way you measured the humans. For us, that was interviews booked. If you don’t know the human baseline, you can’t really say the agent is better. You can only say it functions.

That’s why I gave you 74% and 7 points, and not some dramatic savings claim. That’s the number I can stand behind.