Quit Before It Worked: Automating the Early-Weeks Drop-Off in GLP-1 Programs

Almost every revenue automation I have written up on this site recovers something that was lost at a single moment. A call nobody answered. A quote that went quiet. A form abandoned at the third field. A policy that expired without a conversation. The loss has a timestamp, and the automation reaches back to that timestamp and tries again.

Medical weight-loss programs do not work like that. Nobody cancels. There is no moment. A patient starts a GLP-1 program, comes to the first monthly visit, maybe the second, and then simply does not book the third. The card stops getting charged, or the refill request never arrives, and four weeks later a staff member notices a name they have not seen in a while. By then the patient has usually told themselves a story about why it did not work, and that story is very hard to reverse with a text message.

I am Gideon Wafula, an AI Automation Engineer based in Seoul, and I build these systems for clinics and local service businesses. This is the vertical where I have been asked the most questions in 2026, and it is also the one where I turn down the most requests, because a lot of what clinics want automated here is clinical judgement wearing a marketing costume. So this teardown is as much about the guardrails as the build.

Why this is not the same as plan-of-care drop-off

I wrote previously about the visit-three cliff in physical therapy and chiropractic, and on the surface this looks like the same automation with different branding. It is not, for three structural reasons.

First, the churn reason is pharmacological, not motivational. PT patients tend to drift when the pain stops. GLP-1 patients tend to quit during dose escalation, when the side effects are at their worst and the scale has not moved much yet. The published persistence literature and clinic-side reporting through 2026 both converge on the same uncomfortable shape: a large share of people discontinue before the medication has had time to do what it does, and one-year persistence, while improving in recent cohorts, still leaves a very significant fraction of starters gone. The window that matters is weeks two through twelve, not month six.

Second, the revenue is recurring and the unit is a refill, not a visit. A clinic can hold a perfect appointment show rate and still lose the patient at the pharmacy step, because the prescription lapsed, the coverage changed, or the compounded supply ran out on a Thursday and nobody noticed. Your appointment metrics will look healthy while the program quietly empties. This is closer to failed-payment recovery in a membership business than it is to clinical no-shows.

Third, the regulatory ceiling is much lower. In PT you can send a fairly warm, fairly specific check-in. Here, every sentence that touches dose, symptoms, expected results or safety is a clinical communication, and a language model has no business generating it. That single constraint shapes the entire architecture, and it is why most of the off-the-shelf "GLP-1 retention" sequences being sold to clinics right now are a liability rather than an asset.

The afternoon export that tells you whether you have this problem

Before building anything, pull four numbers out of your EHR or practice management system. This takes an afternoon and it is the only honest way to size the opportunity for your clinic rather than borrowing someone else's percentage.

• Cohort survival by month. Take every patient who started the program twelve or more months ago, and count how many were still filling at month one, month three, month six and month twelve. Not an average retention rate — a curve. The shape tells you where to spend. Most clinics I have looked at find a steeper cliff between month one and month three than they expected, and a much flatter curve after month six.
• Silent gap length. For patients who dropped, how many days passed between their last clinical contact and the point where someone at the clinic noticed. If that number is over thirty days, the automation's main job is detection, not persuasion.
• Reason coding, where it exists. Whatever your staff wrote in the notes: side effects, cost, coverage change, supply, results, moved, no reason logged. If eighty percent of them say nothing, that is itself the finding, and it means your first build is a structured exit question rather than a retention sequence.
• Program revenue per retained patient at twelve months versus three months. Include the visits, the medication margin if you dispense, and the adjacent services that attach later. This is the number that justifies the build, and it is usually a multiple rather than a percentage difference.

Then do the thing that no software will do for you: call twenty patients who dropped in the last six months and ask them why. Twenty calls will tell you whether your constraint is side-effect support, cost, expectations, or a competitor with a cheaper cash program down the road. Automation cannot fix the last one.

The build

I use n8n for the orchestration, self-hosted when the clinic needs the data to stay inside a specific jurisdiction, and I treat the whole thing as a scheduling and detection system that happens to have a language model doing the copywriting. Six steps.

1. A permissioned, staged patient table

Read program start date, current program week, last clinical contact date, next scheduled visit, refill due date and refill status into a controlled table — by API where the EHR supports it, by scheduled export where it does not. Leave nulls explicit. A missing refill date must never be inferred, because an inferred date drives a message to a patient who may have stopped for a clinical reason your workflow cannot see. Everything downstream reads from this table and nothing reads from the EHR directly.

2. A suppression list evaluated at send time

This is longer here than in any commercial vertical, and it must be checked immediately before sending rather than when the audience was selected, because a patient's status can change in the intervening days. Suppress: anyone whose clinician has paused or discontinued treatment, anyone with an open clinical issue or adverse event logged, anyone pregnant or with a pregnancy status change recorded, anyone contacted by a human in the last seventy-two hours, anyone with a billing dispute, anyone who has opted out, anyone who has completed or graduated the program, and anyone flagged by a clinician for any reason at all. The flag needs to be one click for staff, and it needs to be permanent by default.

3. A cadence anchored to program week, not calendar date

The whole point is that risk is concentrated in the escalation weeks, so the sequence follows the patient's own program clock. In practice that means a light touch in the first fortnight while side effects typically peak, a second around the first refill decision, one before the month-three visit, and then a much sparser rhythm afterwards. Each touch is short, and each one does exactly one thing: invite the patient into a channel where a human can actually help them. The message does not manage their symptoms. It asks whether they would like a call, and gives them a one-tap way to book it.

4. A hard clinical boundary on the model

This is the part I will not compromise on, and it is the reason I have walked away from a couple of these projects. The language model personalises tone and drafts scheduling copy. It is deterministically blocked from producing any dose, titration step, injection instruction, side-effect management advice, drug interaction comment, safety reassurance, coverage or eligibility statement, or any figure describing expected weight loss. Not "discouraged from" — blocked, with a validation step that rejects the draft if it contains any of those tokens and falls back to an approved template. Anything a patient asks that touches those areas gets a human, same day.

5. Symptom-language escalation that bypasses everything

The reply handler needs a deterministic first pass, before any model sees the text, that scans for language indicating a serious symptom or distress and routes it straight to a clinician with an alert rather than into the automation's normal queue. Slow is acceptable in a quote follow-up. It is not acceptable here. I also stop the entire sequence for that patient on any inbound reply, in every case, without exception — the automation's job is to start conversations, not to keep talking over one that has already started.

6. Refill-lapse detection as a separate workflow

Run this nightly and keep it apart from the retention sequence, because it is a different signal with a different owner. When a refill passes its due date without being filled or requested, that is an operational alert to a named staff member — not a text to the patient. Half the time the cause is a pharmacy, supply or prior-authorization problem that the patient cannot solve and does not know about yet, and a chirpy "we miss you" message in that situation is how a clinic loses someone permanently.

Compliance, stated plainly

In the US this build sits under HIPAA and TCPA simultaneously. You need documented patient authorization for SMS contact, immediate and deterministic opt-out handling, contact restricted to reasonable local hours, and message bodies that carry no clinical detail — the message invites, the secure channel discusses. Minimum necessary applies to what the automation reads, not just what it sends, so do not pull the full chart into a workflow table when six fields will do. If you dispense or partner with a compounding pharmacy, the promotional line moves closer and the copy needs to stay on the servicing side of it.

In the UK and EU, servicing an active course of treatment and marketing an additional service are different lawful bases, and health data is special category data, which means the data protection impact assessment is not optional paperwork you can skip because the workflow is small. Wherever you are, get the clinic's own counsel to read the message templates before the first send. I write the system; I do not sign off on the regulatory position, and any automation engineer who tells you they will is selling you something.

What it costs and what to measure

Running cost lands around 40 to 150 USD per month — the orchestration layer, model usage and messaging — because a single clinic's send volume is genuinely low. The real expense is the EHR integration and the human triage path, and both are one-time project costs. If a vendor is quoting you a large monthly figure for this, you are paying for their margin, not for compute.

Measure four things, and resist the temptation to lead with the flattering one:

• Month-three survival rate, cohort by cohort, against your pre-build baseline. This is the only number that means the automation worked.
• Median days from lapse to human contact. If this drops from thirty to three, you have bought yourself a real chance at the save regardless of what the sequence says.
• Human-handoff rate per hundred sends. Counterintuitively you want this reasonably high. A sequence nobody replies to is not being polite, it is being ignored.
• Opt-out and complaint rate as a circuit breaker. Set a threshold, and have the workflow pause itself automatically when it is crossed rather than waiting for someone to notice.

The honest caveat

An automation cannot make a patient tolerate a medication, and it cannot make a program affordable. If your twenty phone calls come back saying cost, or saying the side effects were unbearable and nobody warned them, then the fix is a pricing decision and a better onboarding conversation, and the automation will only surface that faster. What this build genuinely does is remove the silence. It makes sure that a patient who is struggling in week three gets asked, by a person, whether they want help — instead of drifting out of a program they paid for and quietly concluding it did not work.

The same pattern ports cleanly to any local business selling a multi-month protocol where the customer's early experience is worse than their eventual result: orthodontics and aligner compliance, hearing aid acclimatisation, hair restoration, allergy immunotherapy, and long-course dermatology. The clock changes, the compliance surface changes, but the structure — program-week trigger, brutal suppression list, model confined to scheduling copy, human owns every clinical word — does not. If you want the broader argument for keeping these systems small and supervised, I made it in the AI agents that actually make money are narrow and boring, and you can see the range of what I build on my AI automation services page.