What Should an AI Voice Agent Do at a Consumer Lender?

An AI voice agent at a consumer lender should do the work nobody on your branch staff has time to do: call every past customer, ask the same qualifying question the same way every time, call the no-answers again, and write down what happened. Then it should get out of the way. The minute a borrower is interested, a person from their own branch takes over.

That's the whole idea. AI is good at volume. People are good at trust.

The rest of this article is about where that line sits, and what happens when you draw it in the wrong place.

Your branches already know this problem

Picture a branch on a Tuesday afternoon.

Two people at the counter. One is taking a payment. The other is on the phone with someone who's behind. A walk-in is waiting. Nobody is calling the borrower who paid off a loan eight months ago and would probably take another one if someone asked.

It's not that your staff doesn't care. It's that outbound calling always loses to whatever is standing in front of them.

So the list sits there. Past customers who already borrowed once and paid you back. Applications that came in online at 10 PM. Everyone agrees they should be called. Almost nobody calls them all.

That's the job AI is actually good at.

What the AI voice agent should do

Think of it as the most disciplined caller you've ever had. Not the smartest. The most disciplined.

Here's what it does well:

• Calls everyone on the list. Not the first 40 names before the branch gets busy. All of them.
• Asks the same question the same way. "Are you interested in hearing about the offer?" gets asked identically on call one and call ten thousand. No bad days, no shortcuts at 4:45 PM.
• Calls no-answers again. Plenty of people don't pick up the first time. A person gives up after one try because they have a counter full of customers. The agent doesn't.
• Calls only when the branch is open. Because the next step is a live transfer, and a transfer to a closed branch is worse than no call at all.
• Logs every call. Who answered, who didn't, who was interested, who asked not to be called. Every time. Nobody forgets to fill in the notes.
• Hands interested borrowers to a person, live. Not a callback request that sits in someone's queue. The borrower is on the line with their own branch while they're still interested.

Notice what's on that list. Volume, repetition, follow-up, logging. Boring work that needs to happen thousands of times without drifting.

Here's what that looks like in practice. At Sunset Finance, a network of consumer lending branches in Georgia, the agent (called Tommy) placed 13,159 calls in the first 14 days of the pilot. That produced about 100 live transfers across 10 branches. Each one was a past customer who said yes to hearing about a pre-approved loan of up to $1,000 and got handed live to their local branch while it was open.

No branch team was going to make 13,159 calls in two weeks. That was never the question. The question was who would.

What it should never do

This part matters more.

A voice agent at a lender should stay away from four things.

1. Credit decisions. The agent doesn't approve, deny or decide anything about credit. It can mention an offer you already made. It can find out who's interested. The decision belongs to your people and your existing process.

There's a regulatory reason too. The CFPB has said that lenders using complex algorithms still have to give applicants specific, accurate reasons when they take adverse action (Circular 2022-03). You don't want a phone conversation anywhere near that line.

2. Promising approval or terms. "You're approved" or "your rate will be" should never come out of the agent's mouth unless it's exactly what your compliance team already signed off on. If a borrower asks about rates, amounts or terms beyond the approved script, the right answer is: "Let me connect you with someone at your branch."

3. Collections disputes. Someone who says "I don't owe that" or "you charged me twice" needs a person. Those calls are about judgment and exceptions. They're also where a wrong word costs you the most.

4. Pretending to be human. This one surprises people, because the voices are good now. Good enough to fool someone for a while.

Don't.

In February 2024 the FCC confirmed that AI-generated voices count as an "artificial or prerecorded voice" under the Telephone Consumer Protection Act (FCC 24-17). That means the TCPA's consent rules apply, and the FCC's rules require these calls to identify the business responsible at the start. The ruling also says the TCPA doesn't carve out technology that claims to be the equivalent of a live agent.

Rules aside, think about your borrower. They trusted you enough to borrow and pay you back. If they find out halfway through the call that the "person" was software, you haven't gained a transfer. You've lost a customer.

This is educational, not legal advice. Talk to your own compliance counsel about consent and disclosures for your calls.

The simple test for any task

When you're deciding whether a task belongs to the agent or a person, ask one question:

If this goes wrong, do we lose time, or do we lose trust?

If a failure only costs time, like a missed call or a wrong callback hour, the agent can do it. You'll catch it and fix it.

If a failure costs trust, like a borrower who feels misled or a dispute handled badly, a person owns it.

Calling 43,000 past customers? Time. Give it to the agent.

Explaining why someone's application came back different than they hoped? Trust. That's your branch manager.

We go deeper on this split in the Academy's first module.

Why "we tried AI and it didn't work"

You've probably heard this from a peer. Maybe you've said it.

Most of the time, the AI wasn't the problem. It was pointed at the wrong job.

Here are the patterns we see most:

• It was asked to do a trust job. Someone had it negotiate, handle disputes, or talk about terms. Borrowers could tell, and they didn't like it.
• It was pointed at a messy list. Wrong numbers, old numbers, people who'd already asked not to be called. The agent did exactly what it was told, and the results looked terrible.
• Nothing happened after the call. The agent had a great conversation and then... nothing. No CRM update. No transfer. No follow-up. The interested borrower waited for a callback that never came.
• Nobody owned it. If no one on your team is responsible for the calling process today, putting AI on top of it just makes the gaps louder.

That third one is where most pilots fail. And it gets its own article: what has to happen after the AI call.

The voice is the easy part now

Here's the part vendors don't love to say out loud.

A voice agent that sounds good on the phone is a commodity. There are a dozen platforms that will give you one in an afternoon. The voice isn't what you're buying.

What decides whether you get funded loans is everything around the call:

• Cleaning the list before anyone gets dialed.
• A local number and caller ID for each branch, so people pick up.
• Calling only during each branch's hours.
• Routing each interested borrower to their own branch and the right person.
• Calling no-answers again, on a schedule.
• Writing every outcome to your CRM the same day.
• A report by branch, so regional managers can see who's getting transfers and who's missing them.

First the list. Then the call. Finally the handoff and the record. If any of those three breaks, the conversation in the middle doesn't matter.

That's the part we install and run. We don't sell the voice. We sell the plumbing nobody wants to own.

Where this has worked

At Sunset Finance, the starting point was 43,000 past customers who had already borrowed once and were sitting unused. The agent called them about a pre-approved offer, and the interested ones went live to their own local branch.

In the pilot, that came out to about 8 live transfers for every 1,000 calls placed. The pilot started with 10 branches. It's at 28 branches now, with 224 live transfers and counting. The full story is in the Sunset Finance case study.

The agent never decided anyone's loan. It found the people who wanted to talk, and put them on the phone with someone they already knew.

That's the job.

A small next step

The way we start is small on purpose: one branch. Within 14 days every reachable customer on that branch's list gets called, and the interested ones are transferred live. There's no upfront fee. You only pay for results: a booked meeting or a reactivated customer.

If you want to hear what your borrowers would hear first, the AI will call you. Leave your number.