The first call went out within a week of the project starting. No vendor contracts, no months-long procurement, no pilot committee. One week.

The Problem

The client operates a deeply managed care model for Medicare Advantage seniors — patients with complex, often multiple chronic conditions who need consistent engagement to stay healthy and out of the emergency room. The existing outreach program relied on a call center full of staff making hundreds of outbound check-in calls per day to the highest-risk patients.

That model had a ceiling. Staff could reach only so many patients. Voicemails piled up. Call quality was inconsistent. And the economics of running a human call center to make wellness calls — most of which were routine — didn’t scale.

When the newly formed innovation team was asked what to build first, patient onboarding was on the table. We looked at it and put it back down. The complexity was real but the wrong kind — vendor integrations, insurance carrier negotiations, device provisioning logistics. Hard for all the wrong reasons. We handed it to IT, and a few days later were asked what we thought about automating outreach.

The Architecture

Targeting. The system doesn’t call everyone. A daily data warehouse query surfaces the 40 highest-risk patients across each care center, ranked by CMS-governed HCC risk scores tied to verified clinical conditions. These are the patients most likely to end up in an emergency room if nobody checks in.

The call. The platform places a voice call, confirms the patient’s identity with a date-of-birth check, and asks two or three open-ended health questions. Nothing scripted or robotic — the questions are conversational, designed for an elderly population who may be suspicious of automated calls.

Scoring. Every patient response is embedded and matched against a vector store of approximately 5,000 catalogued responses, built from prior call transcripts and clinical input. This wasn’t a pre-existing asset — we built it. Each matched response carries metadata encoding clinical signal: healthy, stable, depressed, in pain, acute. The match scores accumulate across the conversation into an overall risk score for the call.

Escalation. If the call score crosses a threshold, the system doesn’t hang up and log a note. It stays on the line, notifies the worker pool, and when a clinically trained staff member answers, conveys the full conversation context and transfers the patient — warm, in real time. The patient never has to repeat themselves.

The staff who used to make hundreds of cold outbound calls per day now receive a curated queue of escalations where a patient is already on the line and already contextualized. That’s a meaningful change in the quality of the work, independent of what it does for outcomes.

The Pivots That Shaped the Design

The original design was SMS-first, using Twilio. Regulatory friction — carrier filtering and delayed delivery that made reliable clinical outreach impossible — killed that approach quickly. We moved to a different telephony provider for voice, and the architecture changed with it.

SMS would have been intent classification on short structured responses. Voice meant free-form natural language, which meant we needed a way to score open-ended speech against expected clinical responses at conversation speed. That constraint is what drove the vector store. The forced pivot produced a more sophisticated and more useful system than the original design would have.

SMS has since been added. The same regulatory problems that blocked it initially were eventually resolved. The platform now reaches patients on both channels.

The Competitive Dynamic

The platform ran alongside an external vendor product targeting a different cohort — appointment scheduling outreach for generally healthier patients. Both teams were aware of each other. Internal stakeholders fed status updates from one side to the other with some regularity.

On raw call metrics — answer rates, voicemail rates, no-response rates — the two products performed at parity. Neither meaningfully outperformed the other on those numbers, which tells you something important: the clinical value of automated outreach doesn’t live in the call completion rate. It lives in what you do with the calls that land.

The internal platform won on cost, on control, and on IP. The vector store — 5,000 responses capturing how Medicare Advantage seniors actually communicate about their health — is a proprietary clinical NLP asset. With a vendor, that walks out the door when the contract ends. The client owns it. The vendor never implemented SMS.

Results

Six weeks into a limited rollout across 44 care centers, at 5% of the eligible patient population:

• 40 emergency department visits averted
• 104 inpatient stays averted
• ~$2 million in documented cost avoidance

These figures are model-based estimates derived from standard ED and inpatient cost benchmarks, tied to escalations that clinical staff have verified were appropriate — patients who genuinely needed intervention. The feedback loop from “escalated” to “hospitalization definitively avoided” is still being tightened; the efficacy percentage is actively improving. Six weeks in, that’s where you should be.

At 5% rollout, the implied ceiling at full deployment is roughly $40 million annualized.

What Comes Next

The architecture is being adapted for a second cohort: end-stage renal disease patients, where medication costs are exceptionally high and the clinical questions are longer and more specific. The escalation path is removed — this isn’t triage, it’s ongoing condition management — but the underlying pattern is the same: structured outreach, embedded response scoring, clinical intelligence at scale.

← All case studies  ·  Engage me on similar work