An AI calling pilot in one location is a solved problem. Thirty locations is a different exercise, and the failures are predictable enough to plan around.
Treat a branch as a configuration
This is the whole game. If adding a location means building a workflow, you do not have a rollout, you have a series of projects that drift apart.
One shared flow, and per branch a configuration that holds: the phone number, the calling window in local hours, the script variables that differ, the transfer destination, the owner, and the compliance status of that number.
The test before you scale: write down what changes between branch one and branch two. If everything on the list is a value, you are ready. If anything on it is logic, fix that first, because the cost of not fixing it multiplies by every location.
The compliance gate comes first
In the United States, application to person messaging is registered with the carriers through brand and campaign registration. Until that is approved, texting is blocked, and an unapproved or rejected registration will stop a launch that was otherwise ready.
Voice generally does not wait on the same gate. That makes the sequencing obvious: ship voice, keep the SMS follow up designed but dark, and turn it on when registration clears. Rollouts that treat messaging as a detail discover it the week before launch.
The calling window belongs to the branch
Hours that connect in one market do not connect in another. A window set centrally spends dials to learn nothing, and worse, it teaches the team that the agent does not work.
Set the window per location, from that location's own connect data once you have a few weeks of it. This single setting usually moves connect rate more than any change to the script.
One number per branch
Shared numbers create a shared reputation, which means one branch's bad week degrades everyone's answer rate. Separate numbers isolate the damage and make the reporting honest, because you can see which market is burning and why.
Report where someone can act
A company wide average across thirty branches is a number nobody owns. The branch manager cannot act on it and the executive cannot diagnose from it.
Pull calls, outcomes and transfers per location, and put each branch's own numbers in front of the person accountable for that branch, next to what was expected. The conversation changes from "how is the AI doing" to "why is this location half the transfer rate of that one", which is a question with an answer.
Write the outcome where humans work
The agent learns things: what the person wants, what was promised, whether the contact details were wrong. If that stays inside the calling platform, the human calls back and asks what the customer already answered.
Every call should end with the outcome, a short summary and the next step written onto the record the rep already opens. This is the difference between an AI layer that helps a team and one the team works around.
Somebody listens, every week
The dashboard reports what the agent believes happened. The recording reports what happened. When those disagree, the script is wrong, and you only find out by listening.
A weekly sample, scored against what the record says, catches the expensive failures: an objection handled badly and marked as a good outcome, a transfer to a line nobody answers, a promise the business cannot keep. Automation platforms make the same point about workflows generally: you cannot reason about non deterministic systems theoretically, you measure them by running real data through them.
And do not switch it on without a human at the other end
An agent that qualifies and transfers into an unanswered line destroys more trust than not calling at all. Before a branch goes live, confirm somebody picks up during the calling window. It is the cheapest check on this list and the one most often skipped.
The people who have done this
MitHub publishes open profiles of engineers who have taken this work from a pilot to a live network, with each deployment written out: the problem, the approach, what was built and what it produced.
See the engineers. Reading the work and contacting them is free.