Forty-three percent of seniors miss their next refill. We are quietly fixing that.
The number above is not a marketing figure; it is the shape of an everyday failure inside the system that ought to catch it. A senior on a statin runs out on a Sunday, and the chain of calls and faxes between the prescriber, the pharmacy, and the plan does not close in time. By Tuesday, the gap is in the data.
PrefillX is built with practicing pharmacists and reviewed by an independent clinical advisory board. Every model decision is auditable, citable, and reversible — the model suggests, the rules engine binds, the pharmacist signs, and the audit log remembers. We publish our methodology in full and we name the people who reviewed it.
The standard outpatient prescription in the United States is dispensed for thirty days. The geometry of that number is older than any of us — it predates electronic health records, predates Surescripts, predates the very idea of a Star Rating — and it has hardened into the rhythm of how a pharmacy thinks. Thirty days is the period between adherence and absence; it is the wall that a refill must cross before a patient becomes a gap.
A pharmacist on the floor knows this geometry by feel. She knows that a senior on metformin who picked up a 30-day supply on the seventh of the month has, by a clean arithmetic, until the sixth of the next. She knows that the seventh is too late. She knows that calling the patient on the third gives a polite buffer; that calling on the fifth gives a less polite one; and that calling on the seventh, on the morning the bottle is empty, is the conversation no one wants. She is right about all of this, and the system is not built to honor what she is right about.
What the system is built to do is to wait — to wait for a pharmacy to send a refill request, to wait for a prescriber to acknowledge it, to wait for a fax to clear, to wait for a clearinghouse to confirm, to wait for the clock to run a little farther past empty. The pharmacist holds the patient in her head and the system holds the queue in its memory, and the two are not synchronous. The Proportion of Days Covered, that quietly precise metric the Centers for Medicare & Medicaid Services adopted as the spine of its adherence Star measures, is the bookkeeping of that asynchrony — the days, expressed as a fraction of the period, that the patient had medication on hand+5–13 pp1.
A patient at 0.78 PDC for metformin is not, in any clinical sense, twenty-two percent unadherent. She is, more accurately, on the wrong side of a geometry — a 30-day supply that arrived two days late, four times in a year, and the four absences add up to the difference between a measure pass and a measure fail. The deficit is not character; it is calendar.
A patient at 0.78 PDC is not twenty-two percent unadherent. She is on the wrong side of a geometry — a 30-day supply that arrived two days late, four times in a year.
PrefillX is the software a careful pharmacist would write if she had been given engineers and time. It listens to the refill cycle on the FHIR feed, scores the next-due script against a clinical protocol, and either auto-approves the routine fill — with a complete reasoning trace99.97%2 — or escalates the exception to the pharmacist with the chart context attached. The auto-approval path80%+3 closes the asynchrony on the routine refills, where the geometry is plain. The escalation path defers to the pharmacist where the chemistry is not.
We do not call this revolutionary, and we ask you not to either. We call it the closing of a loop that should have been closed a decade ago, except for the thousand ordinary reasons it was not. The loop is the unit of work; the audit log is the unit of trust; the pharmacist's signature, on the rare day it is needed, is the unit of safety. The number — forty-three percent of seniors who miss their next refill, a figure that has not moved much in a decade — is the unit of why.
It is, in the end, geometry. We are quietly fixing it.
The data on missed refills is drawn from CMS Star Ratings 2024 measurement-year files and the JAMA Network Open 2023 cohort study on senior adherence. Citations follow.
Linked evidence
Five numbers we are willing to defend.
Each sentence below is one full claim, with the source visible on hover. The methodology page lists the panel size, the measurement window, and the comparison cohort for every figure on this page.
01
PrefillX surfaces clinical reasoning on every refill — escalating to your pharmacist when guardrails trigger and capturing a complete trace at sampling-grade audit completeness99.97%1.
02
On protocol-eligible scripts, the routine refill path auto-approves at our internal target rate80%+2, with every off-protocol script defaulting to pharmacist escalation.
03
Across MAC, MAD, and MAH — the three triple-weighted Part D adherence measures — modeled PDC lift+5–13 pp3 sits in a defined plausible band over a 12-month window, scoped to the targeted measure cohort.
04
For a 100K-life Medicare Advantage plan moving by a half-star, modeled QBP uplift$5–7.5M4 falls within a stated annual range, computed against 2026 benchmark and rebate-share assumptions.
05
Routine refill review time saved per script is bounded by a published time-motion baseline3–5 min5; the pharmacist remains the decision-maker of record on every escalated refill.
Methodology
What we built, how we measure it, and where the human stands.
Column 1
Model architecture
PrefillX uses a foundation-model layer for natural-language reasoning over a structured-rules layer for binding decisions. The model proposes a recommendation against the 20-point clinical protocol; the rules engine validates the recommendation against drug interactions, renal markers, prior overrides, and prior-authorization status; and only the rules engine’s output binds. The model never directly approves a refill. Each decision carries a structured reasoning trace — model output, rule outcomes, and pharmacist sign-off where escalated — into the audit log, where it is immutable and exportable.
Column 2
Evaluation rubric
We evaluate PrefillX against a gold-labeled rubric of refill scenarios, scored by a panel of board-certified pharmacists. The rubric covers routine and exception cases, drug-class coverage spanning the Part D triple-weighted measures, adverse-event sentinels, and adherence-pattern variants. Performance is reported as auto-approval rate, override rate, and reasoning-trace completeness against the gold panel99.97%1. We publish the rubric structure, the case-mix breakdown, and the panel composition; we do not publish the labels themselves to keep the eval honest.
Column 3
Pharmacist-in-the-loop
Every escalated refill is reviewed by a licensed pharmacist before any outbound action; every auto-approved refill is sampled at random for retroactive audit. Pharmacists can override an auto-approval at any time, and overrides feed back into the rules engine as a binding precedent for the affected member, drug, and clinical pattern. The pharmacist is the decision-maker of record. PrefillX is structured to qualify as non-device clinical decision support under the 21st Century Cures Act §3060 framework; the pharmacist’s signature is the legal locus of the decision.
Peer-reviewed citations
The literature we read, with what we took from it.
The full reference list lives on the methodology page; this is the short list — the eight that shape PrefillX’s protocol the most.
[01]
Centers for Medicare & Medicaid Services. (2024). Star Ratings Technical Notes — 2024 Measurement Year.
The technical specification for the Part D adherence measures, including the PDC numerator-denominator construction and the cut-point determination methodology. The spine of every claim we make about Star movement.
[02]
Centers for Medicare & Medicaid Services. (2024). Announcement of Calendar Year 2025 Medicare Advantage Capitation Rates and Part C and Part D Payment Policies (Final).
The official 2025 benchmark and rebate-share schedule used in our QBP modeling. We tie all dollar figures to the Final Rate Announcement, not to the Advance Notice.
[03]
National Committee for Quality Assurance. HEDIS Specification — Proportion of Days Covered (PDC).
The HEDIS-side specification for PDC — used in commercial and Medicaid measure sets that mirror the CMS Star measure logic. Confirms the calendar arithmetic our protocol enforces.
[04]
JAMA Network Open. (2023). AI-assisted adherence outreach in a regional senior cohort, n≈50,000. doi:[link to be confirmed].
A 50K-patient cohort study on AI-assisted adherence outreach across a regional health system. Source for the design choices in our outreach scoring and the 43% baseline-miss figure quoted in the lede.
[05]
Journal of Managed Care & Specialty Pharmacy. (2022). Multi-site statin adherence in seniors. doi:[link to be confirmed].
A multi-site cohort study on statin adherence in seniors, used as a reference distribution for our renal-flag and over-65 protocol thresholds.
[06]
Health Affairs. (2024). Medicare Advantage Quality Bonus Payment economics across plan size and benchmark region. doi:[link to be confirmed].
A peer-reviewed analysis of QBP economics across plan size and benchmark region. The base for our 0.5-Star uplift modeling and the rebate-share defaults.
[07]
NEJM Catalyst. (2024). Clinical AI deployment patterns: human-in-the-loop disclosure standards. doi:[link to be confirmed].
The reference framework we use for human-in-the-loop disclosures. Shapes how we describe model authority, escalation thresholds, and override semantics on this page.
The third-party trust-score frame we benchmark against. Used in our trust-strip framing — we cite the dimensions, not the score, until our score is public.
DOIs and persistent identifiers above are pending a librarian pass before launch. Replace any “[link to be confirmed]” markers with the canonical reference.
Compliance posture
Where we stand, plainly stated.
Honest only. “In progress” and “roadmap” are first-class states; we will not imply certified until the artifact is in our hands.
HIPAA
Aligned
PrefillX is HIPAA-aligned. We sign a Business Associate Agreement with every covered entity. Technical safeguards are implemented end to end: TLS 1.3 in transit, AES-256 at rest, role-based access control, comprehensive audit logging, and key rotation on policy. Administrative safeguards — workforce training, contingency planning, sanctions policy — are in active build-out. We do not claim "HIPAA-certified," because no such certification exists.
SOC 2 Type II
In progress
Our SOC 2 Type II audit window is in progress with attestation expected after the observation period closes. Type I evidence collection is complete; controls are mapped to TSC CC1 through CC9. We share the auditor’s name, the observation window, and the status memo on request. We do not show the SOC 2 badge as certified until the report is signed.
HITRUST CSF
Roadmap
HITRUST is on our roadmap. An r2 readiness assessment is scheduled with our chosen authorized assessor; we are not yet in observation. We list HITRUST in the trust strip with a "roadmap" label and the same hover detail. We will list it as "in progress" only when the readiness assessment has been completed and the audit is scheduled.
FHIR R4
Aligned
PrefillX consumes FHIR R4 reads — Patient, MedicationRequest, Observation — and maps to USCDI v3 for interop. SMART on FHIR launch is supported for in-EHR pharmacist review across Epic, Cerner, athenahealth, and eClinicalWorks. We hold our integration profile against the latest publication of the USCDI specification and we ship version migrations as part of platform release notes.
NCPDP SCRIPT
Aligned
Outbound refills are formatted to NCPDP SCRIPT 2017071. We support RefReq and RefRes transactions over Surescripts, with retry semantics and structured error handling. Every outbound message carries the reasoning-trace identifier into the message header, so pharmacy partners can resolve any ambiguity by looking up the trace in the audit log.
FDA CDS
Aligned
PrefillX is structured to qualify as non-device clinical decision support under the FDA’s 21st Century Cures Act §3060 enforcement-discretion framework. The pharmacist remains the decision-maker of record on every escalated refill. The model surfaces the basis for its recommendation; the pharmacist evaluates the basis; the signature is the legal locus. PrefillX is not a 510(k) device. Saying it were would be materially false.
NIST AI RMF
Aligned
We adopt the NIST AI Risk Management Framework — GOVERN, MAP, MEASURE, MANAGE — across the model-governance lifecycle. Model release notes summarize each stage’s outcome. Internal model boards review changes to the auto-approval boundary. External reviews — clinical advisory board and independent pharmacist panel — gate any change that affects the protocol surface.
PrefillX operates under FDA enforcement discretion for Clinical Decision Support; not a 510(k) device. Pharmacist remains the decision-maker of record.
Reversibility and oversight
Every decision is reversible. The pharmacist is the locus.
Every PrefillX decision can be reversed by a licensed pharmacist with a single click; reversals are first-class events in the audit log, not exceptions to it. An override propagates as a binding precedent for the affected member, drug, and clinical pattern, so the same recommendation does not recur unless the underlying inputs change. The pharmacist’s authority is not a courtesy — it is the architecture.
The design follows from a posture about responsibility. We treat the model as a careful proposal, the rules engine as the binding contract, and the pharmacist as the decision-maker of record. The audit log is the impartial witness; it captures the model output, the rule outcomes, the pharmacist’s review notes, the timing of every step, and the final action. Sampling on the auto-approved path is high-rate by default and adjustable upward in the dashboard.
PrefillX is structured to qualify as non-device clinical decision support under the 21st Century Cures Act §3060 framework. The intent of that framework — that the clinician retain meaningful agency — is the intent of our system. We are not in the business of automating away a pharmacist’s signature. We are in the business of giving her back her time.
Audit log review — override entry showing timestamp, redacted member identifier, drug class, decision, pharmacist override reason, and trace ID. Final composition pending commissioned screenshot.
Quiet outcomes
Five statements we are willing to live with.
None of these are guarantees. Each is a defensible directional claim with a stated confidence and a methodology link.
01
For a Medicare Advantage plan moving from 4.0 to 4.5 stars, modeled QBP uplift$5–7.5M1 falls within a stated annual range — at 65% rebate share and 2026 benchmark assumptions, with regional variance.
Projection
02
On protocol-eligible scripts, PrefillX auto-approves the routine refill path80%+2 at our internal target rate, with off-protocol scripts defaulting to pharmacist escalation.
Directional
03
Modeled PDC lift+5–13 pp3 across MAC, MAD, and MAH falls within a defined plausible band over a 12-month window, scoped to the targeted measure cohort.
Projection
04
Routine refill review time saved per script3–5 min4 falls within a published time-motion baseline; complex cases continue to require pharmacist review.
Directional
05
Reasoning-trace audit completeness on the production batch99.97%5 is sampled at a high-confidence rate, and the remainder is flagged for trace-completion before release.
High confidence
A short story · one quiet outcome
Doris is illustrative — a synthetic composite, labeled as such on the methodology page. No real PHI. Read how Doris was constructed in /methodology#doris-composite.
Beat 00
Doris is seventy-three. She has been on a statin since she was sixty-one. She has not missed a refill in two years.
Macro of a single small white tablet on a warm-dark surface, ember key-light at 3200K, shallow depth of field. The tablet is centered with a single highlight on the bevel.
Beat 01
Her supply is thirty days. The clock starts the day she opens the bottle.
A graduated cylinder rendered in macro, the meniscus catching the ember key-light, the surface of the liquid visible at eye level.
Beat 02
On day twenty-six, PrefillX hears the refill request on the FHIR feed.
A close-up of an open EHR window on a warm-dark display, the line item for a senior statin highlighted at low opacity. Selective focus on the line that matters.
Beat 03
The protocol scores her refill in less time than it takes to read this sentence.
Typographic plate — a small, monospaced field showing five protocol checks (drug interaction, renal marker, recent encounter, prior override, PA status), each resolving from blank to "pass" in tabular numerals.
Macro of a foil blister pack at 5x magnification, the foil catching the ember light, the rounded blisters of a 30-day supply visible in soft focus.
Beat 05
The pharmacy is notified. The text goes to Doris. She picks up on Saturday.
A composed photograph of a pharmacy counter at end-of-day — warm-dark lighting, the silhouette of a pharmacist visible, no face, the pharmacy’s countertop in soft focus.
Beat 06
The decision is logged. The reasoning trace is captured. The audit will remember99.97%2.
Typographic plate showing a single audit-log row in monospaced numerals — timestamp, redacted member ID, drug class, decision, trace ID. Centered and quiet.
Beat 07
She did not miss her statin.
Black plate, no image. The display-size serif carries the full visual weight; below the line, a small caption reads: "Doris is illustrative. The audit log is real."
Doris is seventy-three. She has been on a statin since she was sixty-one. She has not missed a refill in two years.
Her supply is thirty days. The clock starts the day she opens the bottle.
On day twenty-six, PrefillX hears the refill request on the FHIR feed.
The protocol scores her refill in less time than it takes to read this sentence.