AI & Automation

Case Study: From a 10 Branch Pilot to 28 Live Branches

A consumer finance network put an AI agent on its dormant borrower list. What was built, the numbers it produced, and the four things that made the rollout possible.

Mauricio Esparza By ·Published ·2 min read
mithub.club
Short answer

A US consumer finance company with branches had a borrower list nobody was calling. An AI agent was deployed per branch to qualify and transfer live to the branch line, with every outcome written back to the CRM. The rollout produced roughly 29,500 dials at a 76% connect rate, 346 live transfers, 93 sales handoffs and 33 booked meetings, tied to $193K in qualified pipeline and 5 signed deals. It started with no system at all and reached 28 live branches, averaging a new branch live every three days once the first one existed.

This is a deployment write up, anonymised. The company is a consumer finance business in the United States with a branch network. Names are withheld because publishing a client name requires their approval in writing.

The situation

The company had something valuable and idle: a list of borrowers who had paid off or gone dormant. Every branch was serving whoever walked in that day, so the list stayed untouched.

The constraint was not willingness. It was arithmetic. Nobody in a branch has hours to make hundreds of calls while the lobby has people in it.

What was built

One flow, one configuration per branch. Each branch got its own number, its own calling window in local hours, its own script variables and its own transfer destination. Adding a branch became a settings change.

An agent that qualifies and transfers. It confirms who it is speaking to, explains the reason for the call in one sentence, asks whether there is a current need, and transfers live to the branch when there is. It never quotes terms and never implies approval.

Write back into the CRM. Every call ended with the outcome, a short description and the next step on the borrower's record, so the branch rep opened a file rather than a blank.

Branch level reporting. Calls, outcomes and transfers pulled from the calling platform's API into a report per branch, next to what was expected, because a company wide average is not something a branch manager can act on.

The numbers

Across the rollout:

  • ~29,500 dials at a 76% connect rate
  • 346 live transfers to branch staff
  • 93 sales handoffs
  • 33 booked meetings
  • $193K in qualified pipeline and 5 signed deals
  • From a 10 branch pilot to 28 live branches

The four decisions that made the rollout possible

1. A branch became a configuration. This is the one that decides whether a pilot becomes a network. Rebuilding per branch would have stalled the rollout at three or four.

2. Calling windows per branch. Set from each branch's own connect data rather than centrally. This moved the connect rate more than any script change.

3. Outcomes written where the rep already looks. Without it, reps call back asking what the borrower already answered, and the borrower concludes nobody is paying attention.

4. Coverage confirmed before launch. Somebody at the branch answers transfers during the calling window. A branch that went live without that would have damaged a relationship it spent years building.

The constraint that shaped the sequence

Application to person messaging in the United States requires brand and campaign registration with the carriers, and texting is blocked until it is approved. Voice is not gated by the same process.

So voice shipped first and the SMS follow up was designed and held. In hindsight the registration should have started weeks earlier.

What this transfers to

The structure works wherever there is a list nobody is calling and a local team that owns the relationship: clinics, tax practices, insurance books, service companies. The scripts and the compliance work do not transfer. The configuration model does.

Frequently asked questions

Why is the company not named?

Because naming a client requires their written approval, and we do not publish names without it. Everything else here is described exactly as it happened. A fully published MitHub case study also requires the work to be audited and the client to confirm they are satisfied.

What was the single biggest lever?

Calling windows set per branch from local connect data. It moved connect rate more than any change to the script, and it cost nothing but attention.

What would be done differently?

Start the messaging registration earlier. It was the constraint that shaped the sequencing, and beginning it weeks sooner would have allowed the SMS follow up to launch alongside voice instead of after it.

Can this be repeated in another industry?

The structure transfers wherever there is a list nobody is calling and a local team that owns the relationship. Clinics, tax practices and service companies have the same shape. The scripts and the compliance work do not transfer.

Sources

  1. A2P 10DLC registration · Twilio Docs (accessed 2026-09-22)
  2. Building effective agents · Anthropic (accessed 2026-09-22)
  3. Handle errors gracefully · n8n Docs (accessed 2026-09-22)
AI & AutomationCase StudyMulti-location
Mauricio Esparza
Mauricio EsparzaGTM Systems Lead · Revenue Engineer · Founder of MitHub. Designs and runs revenue systems for multi-location businesses: AI voice campaigns, enrichment, CRM automation and attribution. Founded MitHub to teach the method in the open.

Part of AI & Automation on MitHub.

Keep going

AI & Automation

AI Calling for Consumer Lenders with Branches

How a multi branch lender puts AI on the phone without breaking compliance or the branch relationship: which lists to call, what the agent should and should not do, and how each branch sees its own numbers.

Read · 2 min →mithub.club
AI & Automation

Rolling Out AI Calling Across a Branch Network

What changes when an AI calling pilot has to work in thirty locations: treating a branch as a configuration, the compliance gate, branch level reporting and the QA that keeps it honest.

Read · 3 min →mithub.club
AI & Automation

The Pilot to Production Gap in AI Projects

The pilot worked and the rollout did not. The six failures that sit between a demo and a production system, and what has to exist before the second unit goes live.

Read · 3 min →mithub.club
Revenue Engineering

How to Run a Database Reactivation Campaign

The revenue sitting in your existing customer list, how to segment it, what to say, which channel to use first and how to prove the result.

Read · 2 min →mithub.club
AI & Automation

AI Voice Compliance Checklist Before You Dial

What to have in place before an AI agent calls or texts your customers: registration, consent, disclosure, recording, calling windows and the records that prove it.

Read · 2 min →mithub.club
AI & Automation

When to Hire a Forward Deployed Engineer

The buyer's view of the forward deployed engineer: the problems that need someone embedded in your systems, how the engagement is scoped, and how to tell a real one from a consultant with slides.

Read · 3 min →mithub.club