A $0, four-week proof of concept that reads federal healthcare regulations and drafts compliance question sets in minutes, built to prove the AI would actually work before asking anyone to fund it.
The compliance question-set product sells curated checklists that let a hospital self-assess against federal healthcare regulations. A two-person editorial team reads 30–40 page regulatory documents line by line, drafts a question for every requirement, and hand-populates 14 metadata fields per question: category, citation, risk score, entity type, and more. A single document takes days. Every existing question set then needs quarterly re-review against the current regulation, with no diff tool, so re-checking means re-reading the whole thing.
Mapping the workflow surfaced a clean split: the mechanical majority of the work (extraction, formatting, first-pass classification) is automatable. The expert minority, final judgment on correctness and risk, is not, and shouldn't be. Five-whys on "why is this still manual in 2026?" landed on the real root cause: a workflow assumption, humans must transcribe this, that was true for years and stopped being true once vision-capable LLMs got good enough to read dense regulatory prose. Nobody had re-checked the assumption.
Two pipelines: one reads a regulatory document and drafts a full question set matching the editorial team's exact template; the other snapshots a regulatory source, diffs it against the last version, and classifies what actually changed. Both are strictly advisory: the model never auto-publishes, and if it can't ground a citation in the source text, it flags [VERIFY] instead of guessing.
Assist vs. replace is the real product decision. In a trust-sensitive domain, that framing choice matters more than any single feature. It's the difference between a tool a compliance team adopts and one they fear.
A cheap POC is how you earn the expensive conversation. Proving AI quality at zero infrastructure cost is what unlocks the budget conversation for the real build, not the other way around.
Never let the model choose between citing and fabricating. Give it a third option, flag for review, and the failure mode becomes "too cautious," never "confidently wrong."