However the chart arrives — paste, on-screen read, PDF, or photo OCR — Cheryl turns it into one block of processable text. One entry point, so every step below is written once, not per-source.

An AI brain runs the case like a senior pharmacist — it decides the next step, asks back when a fact is missing, and acts only through audited tool calls. No keyword routing.

A deterministic rules engine and an LLM read the chart in parallel — diagnoses, BMI, A1C, prior drugs. Iron rule: every fact must appear verbatim in the source, or it's dropped. Nothing is invented.

Raw facts are cleaned in a deliberate order — de-duplicate, fix time-labels, correct coding, resolve conflicts — so the same condition is never counted twice or mis-coded.

Every fact is checked against public registries — ICD-10, RxNorm, payer medical policy, CMS, NPI, real trial evidence. Cheryl cites; it never recalls from memory.

The draft is checked line-by-line against this payer's hard criteria, returning three verdicts: can we draft, can it be submitted, will it be approved — with the odds and likely push-back.

On a denial appeal, Cheryl cites precedent — real Independent-Review decisions where a similar denial was overturned. Commercial and Medicare lanes are kept structurally separate.

A clean, submit-ready payer letter — and a pharmacist-only dashboard of verdict, approval-odds, blockers and risks. Internal notes never bleed into the letter body.

Missing a key fact — no current BMI, no prior drugs tried? Cheryl asks the pharmacist one item at a time from that payer's checklist, then finishes the letter.

Every payer outcome and pharmacist edit flows back as a PHI-free training signal — aggregated across ≥3 clinics and human-reviewed before any change — so the next letter is sharper.