Westpark Clinic is a multi-specialty practice. Twelve providers, eight specialties, one shared front desk that absorbs every administrative shock the day can produce. The clinical part is fine. The clinical part is what the clinicians went to school for. The back office is where the days disappear.

The work, in order of when it shows up: patient intake, insurance verification, prior authorization, appointment reminders, refill requests, referral coordination, claims follow-up, denials, appeals. None of it is hard on its own. All of it is constant.

"We don't lose money on the medicine. We lose it on the forms."

Why local mode.

Patient data is governed. The clinic could not, and would not, send protected health information to a cloud LLM. Their lawyers were clear. Their patients would have been clearer.

LeedAB runs in local mode here. The ABs, their brain, and the model itself run on the clinic's own GPUs in the back office. Nothing gets sent to a model provider. Integrating with an enterprise provider was on the table; the clinic chose the local route. Patient data is read, reasoned over, and written without a single token leaving the building. For non-PHI tasks, the system can optionally route to a cloud model. Anything that touches a chart, a claim, or a name stays local. This is the configuration the lawyers signed.

The team we deployed.

Five ABs. One job each. They share a brain that knows the practice.

  • Intake AB. Greets new patients via the clinic's booking link, walks them through the intake forms, verifies insurance eligibility against the payer's portal, flags coverage gaps before the visit instead of after.
  • Pre-auth AB. Drafts prior authorization requests from the chart note and the procedure code, submits to the payer, follows up on the schedule the payer respects (not the one the payer publishes), escalates denials with the supporting documentation already attached.
  • Reminder AB. Sends appointment confirmations and prep instructions in the patient's preferred channel, recovers no-shows by offering same-week slots that match the provider's schedule.
  • Refill AB. Handles medication refill requests, checks the chart for the prescribing provider's protocol, routes to the right clinician for approval, files the refill once granted.
  • Claims AB. Tracks every submitted claim, drafts appeals on denied ones, watches aging buckets, escalates the claims that are about to fall off the timely-filing window.

Each AB has its own role and tone. Intake AB is warm. Pre-auth AB is patient. Claims AB is firm.

What it looks like when a patient calls.

10:42 am. A patient calls to confirm Tuesday's MRI. Reminder AB has already sent the prep instructions, but the patient has a question about contrast dye. Reminder AB pages a human.

10:43 am. While the human handles the call, Pre-auth AB notices that the MRI authorization has not come back. It checks the payer portal, sees it is still pending, and pings the payer through the appropriate channel.

10:51 am. The payer responds with an additional documentation request. Pre-auth AB pulls the chart note, attaches the relevant clinical justification, resubmits. Without paging a human, because the chart note already had everything it needed.

11:09 am. Approval comes back. Pre-auth AB updates the chart, notifies Reminder AB, which sends the patient the green-light text.

The front desk was on the phone the whole time. None of this required their attention.

What surprised us.

First, performance. We had assumed local-mode models would feel sluggish. With the AB architecture, the model isn't doing dramatic work most of the time. It is reasoning over context the brain has already organized. Latency is fine. The bottleneck was network calls to insurance portals, not the model.

Second, staff. The two admin staff were nervous about the deployment. They had heard the word "AI" and assumed it meant their jobs were on the timer. Three weeks in, both of them lobbied to expand the pilot. The work the ABs do is the work they had always hated. The work they liked, the human conversations, the difficult cases, the moments that need warmth, is still theirs.

What's next.

Westpark is rolling the ABs out to a second location later this quarter. Same brain, same memory, same local-mode constraint. The clinic owns its data. The ABs do the work.

This is the configuration we expect most regulated industries will end up with. Clinical practices, law firms, financial advisors, anyone who handles records they cannot afford to leak. The cloud is not the future of work for them. The cloud is a vendor they cannot accept. ABs running locally, with a brain they own, is.

If you run a practice that handles records you cannot send to the cloud. Get access. We'll deploy your first AB on your hardware.