Centralized or Decentralized Back Office? The Answer Changes Once You Add AI

There are two ways to run the back office of a multi branch home care agency, and we have watched both of them work.

In the centralized model, billing, intake, eligibility, payroll, and AR live in one team that serves every location. In the decentralized model, each branch or region runs its own back office, close to its own caregivers and its own clients. Both can produce a healthy agency. We have clients thriving on each side. So this is not an argument that one is right and the other is wrong.

It is an argument that the question is changing.

Because the moment an agency starts moving toward real automation, and then toward agentic AI doing the work end to end, the two models stop being equal. Centralized gets easier to automate. Decentralized gets harder. Not impossible, harder. And the gap widens the further you push.

Here is why.

Why decentralization made sense in the first place

Decentralization usually isn't a strategy. It is an accumulation.

An agency opens a second location. The new branch needs to bill, so it bills the way the people there know how to bill. A third location opens, inherits some of the second one's habits, invents a few of its own. Ten years later there are six branches and six slightly different ways of running eligibility, six slightly different conventions for entering authorizations into AlayaCare, six people who each privately consider their way the correct one.

This is not a failure. Local teams build local knowledge. The branch that runs a particular Medicaid waiver every day knows its quirks better than any central team would. Decisions get made by people close to the work. For a long time that proximity is worth more than the consistency you give up.

The cost is invisible until you try to automate. Then it becomes the whole problem.

Automation rewards sameness. Decentralization produces difference.

Automation is, at its core, the encoding of a process into something repeatable. You define the steps, the rules, the exceptions, and then a system runs them the same way every time.

That only works cleanly when there is one version of the process to encode.

When eligibility runs one way in the north region and a different way in the south, you are not building one automation. You are building two, or you are building one and forcing two teams to converge first. When authorizations get entered into AlayaCare with different conventions at each branch, the parser that reads them has to account for every convention, or the data underneath stops being trustworthy. Every local variation is a branch in the logic. Every branch in the logic is another thing to build, test, and maintain.

The decentralized agency that wants to automate ends up paying for its history. The variation that was a feature when humans ran the work becomes a tax the moment you try to hand the work to software.

Agentic AI raises the stakes again

Standard automation follows rules you wrote. Agentic AI is different. You give it a goal, the systems it can act in, and the boundaries it has to stay inside, and it works through the steps itself. Check eligibility for this client, surface the problem, open the task, follow up until it closes.

For an agent to do that reliably, the environment around it has to be coherent. One definition of what a clean authorization looks like. One source of truth for client status. One place the agent reports back to. One set of rules it is operating under.

A centralized back office is already that environment. The definitions are shared, the data conventions are consistent, the system of record is used the same way by everyone. Dropping an agent into that environment is dropping it into a world that makes sense.

A decentralized back office is six environments wearing one logo. An agent that works perfectly in the north region hits the south region and finds different conventions, different assumptions, a different idea of what done means. You are no longer deploying one agent. You are deploying and supervising several, each tuned to a local reality, each drifting on its own. The thing that made agentic AI attractive, that it runs the work without a person steering every step, is exactly the thing that breaks when the ground underneath it keeps shifting.

This is a data problem before it is an org chart problem

It is tempting to read all of this as an argument about reporting lines. It isn't, really. You can centralize the org chart on paper and still have six branches doing the work six ways underneath it. The reorg doesn't fix the automation problem by itself.

What automation actually needs is consistency in the layer where the work happens. The same fields filled the same way. The same definition of a valid authorization. The same eligibility cadence. The same understanding of when a visit is billable. That consistency is easier to achieve when one team owns the process, which is why centralization helps. But the thing you are really buying is a coherent data and process layer. Centralization is the common path to it, not the point of it.

We have said for a long time that you redesign the workflow before you automate it. This is the same idea at the level of the whole back office. The redesign step for a multi branch agency is converging on one way of doing the work that everyone can live with. Skip that, automate the variation as it stands, and you encode six versions of the truth and ask the software to reconcile them forever.

What this does not mean

It does not mean centralize everything tomorrow. For a lot of agencies, a full back office consolidation is a large, disruptive project, and local knowledge is genuinely worth protecting. Ripping it out to chase an automation roadmap can cost more than it returns.

It does mean that if automation and AI are anywhere on your horizon, the structure of your back office is now part of that decision, not separate from it. The two questions used to live in different meetings. They don't anymore.

A practical middle path exists. Keep local judgment where judgment matters, the conversations with families, the calls that need someone who knows the case. Centralize the rule based, data moving, repetitive work, the eligibility checks, the authorization ingestion, the reconciliation, into one consistent layer. That is also, not coincidentally, exactly the work that automates well. The split between what to centralize and what to leave local turns out to be close to the split between what to automate and what to leave human.

Where this leaves the decision

Both models still work. Agencies will keep succeeding on each. But the calculus has shifted.

A decentralized back office optimizes for proximity. A centralized one optimizes for consistency. For most of the history of home care, proximity won often enough that decentralization was a reasonable default. Automation, and especially agentic AI, pays out on consistency. The more of your operation you intend to hand to software, the more that single coherent environment is worth, and the more the old default deserves a second look.

The agencies that will pull ahead are not the ones with the most tools. They are the ones whose back office is coherent enough that the tools have something solid to stand on.

If your agency runs on multiple branches and you are starting to think about automation, the structure question is worth having out loud before you buy anything. Whoever owns your back office already knows which branches do it differently. That conversation is where this starts.

Less noise. More performance. Measured by what shows up on the P&L.