Meridian Health Patient Concierge

Meridian Health's patient support line was carrying call volume it was never staffed for, and roughly seventy percent of those calls were routine: scheduling changes, billing questions, insurance verification. Hold times stretched past twenty minutes during peak hours, and the staff answering them were too busy repeating the same information to focus on patients who needed a real clinical conversation. The constraint that ruled out most off-the-shelf chatbots: patient data cannot leave the network's own infrastructure, and anything touching symptoms, diagnosis, or medication has to reach a human, not a model.

Meridian Health runs a regional network of clinics where the support line is often a patient's first and most frustrating touchpoint. Leadership asked for something that would cut call volume without ever putting patient privacy or clinical judgment at risk. That constraint shaped the entire architecture before a single conversation flow was designed.

Before any message reaches a model, a redaction pass strips identifiers and sensitive details the assistant does not need to answer a scheduling or billing question. What the model sees is deliberately less than what the patient typed, which keeps protected health information out of a request it was never necessary to include.

The assistant answers exclusively from Meridian's own scheduling policies, billing procedures, and insurance documentation. It does not improvise medical guidance, because it was never given the latitude to. When the documentation does not cover a question, it says so instead of guessing.

The instant a question touches symptoms, medication, or anything resembling a diagnosis, the assistant stops and hands the conversation to a live nurse queue, full transcript attached. Nothing clinical is ever answered by the model. That hard boundary was non-negotiable for Meridian's compliance team, and it is enforced in code, not just in a prompt.

The entire system runs inside Meridian's own private cloud under a signed data-handling agreement, which is why it has no public URL and no open demo. That constraint was a requirement of the engagement, not a limitation of the build, and it is the reason this case study describes the system without linking to it.

Within its first quarter live, the concierge deflected 62 percent of incoming calls and cut average hold times by more than half. Nurses got real hours back each week, and every escalation that reaches them now comes with context instead of a cold transfer. Meridian has since extended the same architecture to two additional clinic groups within the network.

The concierge now absorbs the majority of routine calls, and average hold time has come down by more than half. Nurses spend less time on scheduling and billing questions that never needed a clinical license to answer, and every escalation that does reach them arrives with context instead of starting cold. Because the entire system lives inside Meridian's private infrastructure under a signed data-handling agreement, it has no public demo and no external link, which is by design rather than an oversight.

Every Devyst engagement follows a structured process: discovery, architecture, build, and handoff. This project was no different. We aligned on scope, reviewed existing systems, delivered iteratively, and handed off with documentation and runbooks.