Refill medication with AI: how an agentic refill runs from request to doorstep

Refill, renewal, and why the difference matters

In GLP-1, hair-loss, hormone, and sexual-health programs the medication runs on a subscription and the dose may step over time, so the “refill” a patient asks for is usually a renewal — and a renewal is a prescribing act. That is not a limitation of AI; it is the reason the workflow has a clinician in it. The agent’s job is to make the clinician’s decision fast and well-informed. Definitions: AI refills.

Why refills come first for agents

• Volume. Refill requests are the single largest category of patient contact in a running program.
• Structure. The request has a known shape — patient, active care plan, catalog treatment — and a protocol that says what to check.
• The relationship exists. Consent, identity, and history are on file; the agent is continuing care, not starting it.
• Retention. Patients rarely churn because a refill was easy. They churn when it was not.

Regulators have noticed the same thing. The first state-sanctioned test of AI in prescribing — Utah’s 2026 sandbox pilot with Doctronic — is a renewals program: the AI may renew existing prescriptions for roughly 200 maintenance medications, never write a new one or change a dose, never touch controlled substances or injectables, and in the current phase a licensed physician reviews each renewal before the pharmacy sees it. Early data showed the AI escalating about a quarter of requests to a physician and physicians agreeing with its renewal recommendations roughly nine times in ten. It is one company in one state on a one-year term, but it confirms the shape of the workflow below: collect well, route well, keep the decision human. More in Can AI prescribe medication?

The agentic refill, step by step

1. Intent. The patient says “I’m almost out” — in your app, by message, or to an assistant acting for them. The agent recognizes a refill intent and identifies the medication.
2. Match. It finds the patient and the active care plan; a new patient would need intake first.
3. Eligibility. Fills remaining, date of last clinician review, the state’s re-evaluation interval, whether a protocol check-in is due, and whether the drug is controlled.
4. Follow-up. The protocol’s questions — weight, side effects, adherence, new medications or conditions — collected conversationally and stored as structured data.
5. Submit. A refill encounter with the follow-up attached and a proposed next step (“continue at current dose”, “step to the next dose per protocol”).
6. Authorize. A licensed clinician in the patient’s state reviews and approves, changes, or declines. The prescription carries their NPI and signature; for controlled substances they sign with EPCS two-factor authentication personally.
7. Fill and ship. The pharmacy fills and ships — cold-chain for injectables. Webhooks report completion, shipment, and delivery; the agent relays them.

The checks that protect the patient

The follow-up is not friction; it is what makes an agent-requested refill safe. A GLP-1 dose should not step up without the check-in that catches severe nausea or abdominal pain. A relationship that has lapsed past the state’s re-evaluation interval needs a new evaluation rather than a refill. A medication the patient started elsewhere may interact. Protocol-defined red flags — chest pain, signs of pancreatitis, an allergic reaction — escalate immediately, outside the queue, with the agent stepping out of the conversation. See Where AI stops.

Infrastructure should enforce these checks, not the agent’s prompt: a request that fails eligibility or state modality rules never reaches a clinician as a refill; it routes to a check-in, a live visit, or a decline with a reason.

A GLP-1 example

A patient on compounded semaglutide, four weeks into the current dose, messages that they have one dose left. The agent matches the care plan, sees the titration protocol calls for a check-in before the next step, and runs it: weight, side effects, adherence, anything new. Expected nausea, good adherence, down four pounds. It submits a renewal encounter with the follow-up attached and the protocol’s proposed step. A clinician reviews the structured case, agrees, and signs. The pharmacy ships cold-chain. The agent tells the patient the new dose is approved and arriving Thursday, and schedules the next check-in. The clinician’s time went to the decision; everything else happened around it.

What the patient sees

• “Got it — a few quick questions before your next shipment.”
• “Thanks. I’ve sent your renewal to a clinician licensed in your state for review.”
• “Approved. Your next dose is 0.5 mg; the pharmacy is preparing it.”
• “Shipped — arriving Thursday. I’ll check in with you in four weeks.”

Every message is true because each one is triggered by an event in the pipeline rather than a guess: the encounter completing, the order shipping, the delivery scan.

Measuring it

The metrics that tell you the loop is working: time from request to authorization, refill completion rate (requests that end in a delivered order), support tickets per hundred refills, and lapse rate — patients who run out before the next shipment. An agent-run refill workflow should move all four in the same month it ships. Watch the decline rate too: a rising share of declines usually means the eligibility checks are letting ineligible requests through to clinicians.

Building it

On Lithos a refill is the same three-call loop as a new treatment: match the patient, use the existing care plan, submit an encounter with refill as the request type and the follow-up attached. A licensed clinician in the patient’s state reviews and signs; encounter.completed, order.shipped, and order.delivered arrive as webhooks. Your agent stays in the conversation; Lithos runs the clinician, the pharmacy, and the paperwork.

Frequently asked questions