Naming recommendation: keep this client anonymous. It combines a named prescription medication with patient conversations. Even with permission, naming the client invites questions about data handling that a marketing page is the wrong place to answer. “A US telehealth provider” carries the same credibility with none of the exposure.
Context
Our client runs an obesity care programme built around semaglutide treatment. Patients on that pathway have questions constantly — about dosing, side effects, what to expect, what is normal — and most of those questions arrive outside clinic hours.
Problem
Two things make a healthcare assistant harder than a support chatbot.
The first is that the cost of a wrong answer is not a bad review. A general-purpose model asked about medication dosing will answer confidently whether or not it should, and in this domain confident and wrong is the failure mode that matters.
The second is that the useful answers are specific to this programme — its protocol, its escalation rules, its patient journey — not to obesity treatment in general. A model relying on its training data would give plausible, generic, unusable answers.
So the design problem was a narrow one: be genuinely helpful within the programme’s boundaries, and reliably decline or escalate outside them.
Approach
NEEDS INPUT. The three things a reader wants to know, one sentence each:
- How the assistant was grounded in the client’s own protocol content — retrieval, a curated knowledge base, structured flows, or a combination.
- How out-of-scope and clinical questions are detected and escalated to a human.
- How responses were evaluated before launch, and whether clinicians reviewed them.
This section is what separates an engineering firm from an agency that wired up an API key.
What we built
A conversational AI assistant supporting patient engagement and guided assistance across the obesity care journey.
NEEDS INPUT: channel (web, SMS, in-app), whether it hands off to a human, and what content it draws on.
Results
NEEDS INPUT. Conversations handled, proportion resolved without escalation, or out-of-hours coverage gained.
Related
Retrieval-Augmented Generation for an Education Platform · Healthcare Management CRM · AI Engineering
