A lender with branches has an asset most companies do not: a list of people who already borrowed and paid it back. They have a history you can see, they know the brand, and nobody is calling them, because branch staff are serving whoever walked in today.
That is where AI calling earns its place in this industry.
Which lists, in order
1. Paid off borrowers. Clearest reason to call, best history to qualify on, warmest reception.
2. Dormant accounts. People who were active and drifted. The reason to call is what changed since.
3. Aged declines. Circumstances change. This list needs the most care in how it is approached and the tightest rules about what the agent may say.
What the agent does, and what it never does
Does: confirms it is speaking to the right person, explains why it is calling in one sentence, asks whether there is a current need, and transfers to the branch when there is.
Never does: quotes terms, implies approval, or discusses account specifics it cannot verify it is entitled to discuss. Those belong to a person. Putting them in a script is how a compliance problem begins.
The agent's job is a warm conversation with the branch, not a sale.
The three things that decide whether it works
Calling windows per branch. Hours that connect in one market do not connect in another. Set them from each branch's own connect data, not centrally.
Somebody answers the transfer. This is the one that damages relationships when it is skipped. A branch that spent years building trust does not recover it quickly from a machine that called and then dropped the person into a ringing line. Confirm coverage before a branch goes live.
The outcome reaches the record. The branch rep should open the record and know what was said, what the person wanted, and what was promised. Otherwise they call back asking questions the borrower already answered, and the borrower concludes nobody is paying attention.
Compliance sequencing
In the United States, application to person messaging is gated by brand and campaign registration with the carriers, and texting is blocked until it is approved. Voice is generally not gated by the same process.
The practical order: start the registration early, design the SMS follow up, launch voice while it is pending, and enable messaging when it clears. Rollouts that treat this as a detail discover it the week before launch.
Reporting the branch can use
A company wide average across thirty branches is a number a branch manager cannot act on. Each branch should see its own dials, connects, transfers and outcomes, next to what was expected.
That changes the conversation from "how is the AI doing" to "why is this branch at half the transfer rate of that one", which has an answer.
Scaling past the pilot
A branch has to be a configuration, not a build: its own number, its own window, its own script variables, its own transfer destination, on one shared flow. If adding branch two means rebuilding, you will have thirty systems that drift apart and a fix applied once that silently does not apply everywhere.
Who does this work
MitHub publishes open profiles of engineers who have taken exactly this from a pilot to a live branch network, with the deployments written out.