AI in Multi-Site Clinical Operations: The Operator's Playbook

We operate Pentus Health — a multi-site clinical business. Before we installed AI across the operation, I spent a lot of time inside the workflows. What I found wasn't what I expected.

The assumption going in was that clinical documentation was the big opportunity. Every provider I'd spoken to complained about charting time. AI scribes were having a moment. It seemed obvious.

The documentation opportunity was real. But it wasn't the biggest one.

Where the actual time goes

When we mapped the full workflow across sites, documentation was the most visible problem but not the most expensive one. The biggest time sinks, in order:

1. Prior authorization management. The process of getting approval from insurance before delivering care. Manual, repetitive, time-consuming — and completely rule-based, which means it's perfectly suited for AI automation.
2. Scheduling and schedule management. Cancellations, late additions, provider schedule changes across multiple sites. Each change cascades through multiple systems and requires manual updates.
3. Billing reconciliation. Matching claims to payments to denials to resubmissions. Most clinical billing systems can do some of this, but the exception management — the things that don't match automatically — still goes to a human.
4. Documentation. The thing everyone talks about. Real opportunity, but not the top of the list.

What AI actually helps with (and what it doesn't)

Prior auth is where I'd start in any clinical operation. The rules are defined. The forms are standardized (or close to it). The status tracking is rule-based. An AI that can handle standard prior auth submissions and flag the exceptions for human review can recover significant provider and admin time.

Scheduling optimization is the second lever. Not patient-facing scheduling — internal schedule management. AI that can look at cancellations in real time, identify available patients who could fill the slot, and send a targeted outreach is straightforwardly valuable.

Documentation genuinely helps, but with a caveat: the AI needs to be trained on the specific clinical context, and the provider still needs to review everything before it goes into the chart. The efficiency gain is real (30–50% reduction in charting time is achievable), but the oversight requirement is non-negotiable.

Where AI doesn't help: anything that requires clinical judgment. Diagnostic decision support tools exist, but they're a different category — regulated, complex, and not what I'm talking about here. Operational AI and clinical AI are different things.

The compliance question

HIPAA is always the first objection. And it's a legitimate one. Any AI tool that touches PHI needs to be evaluated carefully — vendor BAAs, data handling practices, where the processing happens.

The practical answer: most of the operational automation doesn't actually touch PHI if it's architected correctly. Schedule optimization can work with anonymized time slots. Prior auth status tracking can happen at the claim level. The documentation piece requires more careful handling, but the tooling for HIPAA-compliant AI scribes is mature.

Don't let the compliance question stop the evaluation. Let it shape the implementation.

The site-by-site rollout

Multi-site is harder than single-site, not because the AI is different, but because the workflows vary more than you think. The processes at site one look like the processes at site two on paper. In practice, each site has developed its own workarounds, its own handoffs, its own informal rules.

The approach that works: start with one site, get it right, then replicate. Don't try to standardize across sites at the same time you're implementing AI. Pick the most operationally mature site, build the workflow there, prove the results, then bring it to the others with the friction already worked out.