Opportunity map

Before building anything, Discovery scores the candidate opportunities and picks deliberately. This page shows the map for the demo’s own case: why medication adherence was chosen, what was set aside, and - the part most decks skip - where a conversational AI is the wrong tool.

How the map is scored

Each candidate is scored on a few plain dimensions, not a black-box number:

Impact - how much does moving it matter, clinically and financially?
Feasibility - can a conversational agent actually do it well today?
Risk boundary - can the unsafe cases be drawn and enforced deterministically?
Data readiness - is there usable, licensable, privacy-safe data? (see the data-readiness scorecard)
Value clarity - can the value be evidenced, not just asserted? (see the value model)

The weighting behind a real engagement’s scores is calibrated per client; the dimensions above are the public shape.

Why medication adherence

Medication adherence scored highest on this map:

Impact: non-adherence is a large, well-documented, avoidable cost driver.
Feasibility: adherence support is conversational, educational, and motivational - squarely what an LLM does well - and it needs no diagnosis and no dosing decision.
Risk boundary: the dangerous cases (acute symptoms, dosing changes, diagnosis) are crisp and can be refused or escalated deterministically, so the safe envelope is enforceable, not aspirational.
Data readiness: the domain can be covered by synthetic, public-source- paraphrased content with no PHI (see the dataset card).
Value clarity: the value turns on a measurable adherence lift.

Where AI is not the answer

The honest headline most opportunity maps omit. A conversational agent is the wrong tool for:

Acute emergencies. Chest pain, suicidal ideation, severe bleeding need an immediate deterministic escalation to a human or emergency pathway, not a generated answer. The system handles them with a rule-based floor that runs before any model call - precisely because generation is not the answer here.
Dosing and diagnosis. “What dose should I take?” or “What do I have?” is a medical decision a wellness agent must refuse, not attempt; these are out of scope by policy and enforced by the scope classifier.
Anything that is really a medical device. If the intended use is to diagnose, treat, or drive a clinical decision, the answer is a regulated device pathway, not a demo agent (see the regulatory posture).
Problems a deterministic rule or a human solves better. If a lookup table, a form, or a nurse call is cheaper and safer, that is the right answer; an LLM there adds cost and risk for no gain.

Naming these is not a disclaimer - it is the selection discipline that makes the chosen opportunity credible.

What was deprioritized

Adjacent opportunities considered and set aside, each with its reason:

Candidate	Why deprioritized
Symptom triage / diagnosis support	Risk boundary too blurred for a wellness agent; drifts toward a medical-device intended use.
Dosing calculators	A clinical decision; out of scope by policy and refused by design.
Full EHR-integrated care management	Data-readiness and compliance load (real PHI, integrations) dwarf the demo’s purpose.
Open-domain medical Q&A	Unbounded scope defeats the enforceable safe envelope that makes the adherence case safe.

Each was set aside for a stated reason, not silently dropped - which is what a real Discovery deliverable looks like.