In healthcare, the guardrails are the product.
A patient-facing assistant for engagement and education, where the engineering effort went into what it refuses to do at least as much as into what it says.
Patients have questions between appointments. Clinicians have no time.
Most patient questions are not clinical judgement calls. They are about what a medication is for, how to prepare for a procedure, what a term in a letter means, and what happens next. All of it is answerable from approved material, and all of it currently consumes clinical time or goes unanswered.
Unanswered is the expensive outcome. Patients who do not understand their care plan follow it less consistently, and the cost of that lands later and larger.
But healthcare is the domain where a fluent, confident, wrong answer does real harm. A general-purpose assistant will answer a symptom question because it is capable of producing text, and that capability is exactly the hazard.
A narrow scope, enforced, and an escalation path that always works.
Scope was defined by exclusion before anything was built. The assistant handles education, preparation and navigation. It does not diagnose, does not triage, does not interpret results and does not advise on medication changes. Those boundaries are enforced in the system, not requested in a prompt.
Answers are generated only from clinician-approved source material, with the source shown. Where a question falls outside the approved corpus, the assistant says so and routes to a human. It does not attempt a general answer.
Anything resembling an urgent symptom short-circuits the conversation entirely and surfaces the appropriate escalation route immediately. That path is tested as a first-class feature, not as an edge case.
Personalisation operates on care pathway and stage rather than on inferred clinical state; a patient preparing for a procedure gets the material for that procedure, which is genuinely useful and carries none of the risk of inference.
What the assistant does.
Education delivery
Approved material presented conversationally and at the patient's reading level, not as a leaflet.
Pathway personalisation
Content matched to care pathway and stage, not to an inferred clinical state.
Preparation support
Procedure and appointment preparation, the highest-volume and lowest-risk category of question.
Source attribution
Every answer traceable to approved material, visible to the patient.
Scope enforcement
Diagnosis, triage and medication questions refused in the system, not merely discouraged in a prompt.
Escalation
Urgent-symptom detection that short-circuits the conversation and surfaces the right route immediately.
Clinical review
A queue where clinicians see what is being asked and correct the corpus, which is how the system improves.
Built with.
Model layer
Content
Safety
Application
What changed.
- Routine questions answered without clinical time. Education, preparation and navigation handled at scale, at the hour the patient has the question.
- Scope held under pressure. Refusal enforced in the system, not requested of the model, which is the only version of this that survives contact with real users.
- Escalation treated as a feature. Urgent-symptom routing tested as a primary path, not as an exception, because it is the path that matters most and fires least.
- The corpus improves from real questions. A clinical review queue turns what patients ask into approved material, rather than leaving gaps to be discovered by complaint.
