Skip to content
Buy This GLP
← Research
Evidence

An AI coach raised retention, until the model adjusted for everything else

Adherence rose after a digital clinic added an AI support agent. The regression in the same paper found the opposite, and cost predicted dropping out either way.

Neil Sanders6 min read
Same study, two answersraw comparison47.3% → 53.2%adherence improvedadjusted modelOR 1.178attrition increasedWhat actually predicted dropping out:heavy month-1 tracking (OR 15.8) · pauses (2.5) · cost (2.5)Adherence here means orders placed, not doses taken.

Digital weight-loss services lose most of their customers, so anything that keeps people on treatment is worth real money to them. One such service in Australia added an asynchronous AI support agent and afterwards analyzed what happened across 16,556 people prescribed semaglutide [1]. The retention problem it is trying to solve is quantified in two thirds stop within a year.

The simple comparison looks like a success. Six-month adherence was 53.2% after the agent launched against 47.3% before, a difference comfortably clear of chance.

What the study identifies more confidently is what actually predicts people leaving, and the largest of those is the strangest. Intensive self-tracking during the first month, more than 25 logging events, was associated with a fifteenfold increase in the odds of dropping out. That is almost certainly not tracking causing attrition. Much more plausibly, someone logging obsessively in week two is someone already anxious about whether it is working — which makes it a warning sign worth a phone call rather than a behavior to discourage.

The number most relevant to this desk is plainer. Higher program cost was associated with roughly two and a half times the odds of attrition, alongside program pauses at a similar magnitude. Price is not merely what someone pays; it is a predictor of whether they finish, which is the argument against buying a cheap first month at a service whose maintenance pricing climbs — the pattern documented in the headline against the bill and what a twelve-month commitment locks you out of.

One definition limits everything above. Adherence here means at least six orders fulfilled within 183 days — a measure of purchasing, not of injecting. Someone who orders reliably and skips doses counts as adherent, and someone who stockpiles counts as lapsed. For a commercial service those are the data that exist, and they are a reasonable proxy, but the distinction matters when the same word is used to describe taking medicine, as in what patients said when asked.

Frequently asked

Did the AI support agent help?
The study does not settle it. Raw adherence was higher after launch, but the adjusted model associated that period with higher odds of dropping out, and an era comparison cannot separate the chatbot from everything else that changed at the same time.
Does tracking your progress make you more likely to quit?
Almost certainly not directly. Heavy logging in the first month predicted attrition strongly, but intensive early tracking most plausibly marks someone already anxious about whether treatment is working, rather than causing them to stop.
What predicted people staying on treatment?
Lower program cost and avoiding pauses. Higher cost was associated with roughly two and a half times the odds of dropping out.

Sources

  1. [1] Talay L, Xu C, Alderete J, Hom J, Tan M, Ahuja N (2026). AI Patient Support and 6-Month Medication Adherence in a Digital Obesity Program: A Retrospective Analysis Diabetes, Obesity & Metabolism. PMID 42575845

Where to get it

Price the injectable sellers

The desk lists every seller that publishes an injectable figure, with the advertised price struck against the one a buyer is billed.

Open the price desk

More in Evidence