Executive summary

Reference documentation: not a medical device

This is a public reference implementation evaluated on 100% synthetic data. It is a capability and readiness reference, not a compliance certification, legal advice, or a medical device. It is not clinically validated and handles no production PHI.

In one paragraph

This is a measurement-first approach to using generative AI in a regulated industry, shown end to end on a concrete case: a conversational health agent for medication adherence whose every clinical statement must cite a verified knowledge-base entry, or the agent declines, paired with an automated evaluation harness that proves it before anything ships. It exists to answer the question a decision-maker actually asks: can this be trusted, is it worth it, and what does “ready” look like.

The problem it addresses

Most generative-AI pilots never reach measurable business impact; industry experience puts the stall rate at roughly 70-95%. Four causes do most of the damage, and each is treated here as a deliverable rather than a hope:

• Data readiness - whether the underlying information is good enough to build on.
• Risk and compliance - whether the system can stay inside a safe, defensible envelope.
• Cost - what it actually costs to run at volume.
• Value - the measured business case, including where the answer is not AI.

What it delivers

The work is framed as two sprints a buyer can recognize:

• Discovery - is this the right opportunity, is the data ready, and what are the value and the risk, decided before a line of production code.
• PoC-to-Production - a measured, safe, production-shaped system with the evidence needed to ship it.

How trust and safety are established

• Cite or refuse. Every clinical assertion must cite a verified knowledge-base entry; with no support, the agent declines rather than guesses.
• Synthetic data by design. The system is built and evaluated on entirely synthetic, representative data, so real patient information never has to touch an unproven system.
• A safe envelope. The agent is held to a defined scope; outside it, the deterministic behavior is to refuse or escalate to a person, not to improvise.
• A calibration gate. Automated safety judgments are checked against human judgment, and launch is blocked when they do not agree closely enough.
• A wellness reference, by design. It is deliberately scoped as a wellness reference, not a medical device.

Compliance and risk posture

The governance package maps the system against the frameworks decision-makers ask about, and states the boundaries plainly: HIPAA readiness, the EU AI Act, the NIST AI Risk Management Framework, ISO/IEC 42001 and SOC 2, and regional data-protection regimes such as Chile’s Ley 19.628 and Brazil’s LGPD. The point is not a certification claim; it is a clear, mapped readiness position.

Honest scope: where AI is not the answer

Credibility comes from declining the wrong use as clearly as embracing the right one. The opportunity map names what was deliberately deprioritized and where a simpler, non-AI path is the better call.

What you receive

The reference delivers a prioritized opportunity map, a value model on the case’s own numbers, a data-readiness scorecard you can run on your own data, the evaluation harness and its human-vs-judge calibration gate, an operations runbook for the handoff, and a costed blueprint from demo to production. See What you receive and the full reference library.

Bottom line

This shows what buying down the four failure modes looks like in practice, with the evidence to prove it, on a wellness reference built entirely on synthetic data. It is designed to make the trust, value, and readiness questions answerable before committing to build.