The end-to-end workflow — every node, in plain language. Data flows top → bottom; the learning loop feeds back up.
However the rep gets the chart, cheryl turns it into one block of processable text. One entry point, four sources — so every step below is written once, not per-source.
Drop in the forwarded chart text.
One click reads any on-screen page — EHR or payer portal.
Read to text on-device; nothing leaves the machine.
Image-only PDFs go through AI vision OCR.
all four converge on a single, encrypted, token-authenticated intake
An AI brain runs the case like a senior pharmacist: it decides what to do next — gather a missing fact, or write the letter. No keyword routing; it genuinely understands the case, then calls a standardized, auditable tool.
Which drug, which payer, which scenario — new, renewal, or appeal.
Missing facts → it questions the rep, never fabricates.
Every action is a controlled, logged tool call.
Two engines run in parallel to pull the facts out of the chart — diagnoses, BMI, A1C, prior drugs tried, side effects.
Hundreds of hand-written deterministic rules. Owns the numbers & codes — stable, predictable, explainable.
Reads the hard sentences: "no sleep apnea", "family history of diabetes", "tried but stopped".
The raw facts are cleaned up before use — and the order matters.
A chronic condition mislabeled "past" → set to "current".
Gestational / drug-induced / pre-diabetes → their own codes, not type-2.
"present" vs "absent" → AI's full-sentence read wins.
Rules + AI hit the same fact → merge to one. Runs last: dedup would otherwise collapse a present/absent pair before the conflict is judged.
Each fact is checked against public, authoritative data sources — the backbone of trust. cheryl cites; it doesn't recall from memory.
Diagnosis codes — must be specific, no vague placeholders.
NLM's standard drug vocabulary.
Clinical terms & lab codes (kept in reserve).
Real trial evidence, each with a verifiable PubMed ID.
Per-drug, per-plan criteria — the published policy of record.
Federal eligibility & appeal criteria from CMS.
National provider registry — verifies the prescriber is real.
Independent-review rulings — powers node 7.
The draft is checked line-by-line against this payer's hard criteria, returning three independent verdicts — not one vague yes/no.
If facts can't support a real letter, refuse — don't fabricate.
What, if anything, the rep still needs to fill in.
A probability + "how the payer may push back" + rebuttal prep.
the goal isn't a letter that looks good — it's a letter that gets approved
On a denial appeal, cheryl does what others don't: it cites precedent. It pulls real Independent Review decisions where a similar denial was overturned, and argues from them — like a lawyer citing case law.
Denied, then independent medical review. Corpus: California DMHC — 1,625 published decisions.
Independent review (IRE) → administrative-law judge (ALJ) → Medicare Appeals Council. Its own precedent.
The result splits into two deliverables that are physically separated in the product.
Clean, compliant, submit-ready — only what a letter should contain.
Verdict, approval-%, blockers, strategic risks & checklist — pharmacist-only, never in the letter body.
internal risk notes inside a letter read as defects to a reviewer — so they live apart
Missing a key fact (no current BMI, no prior drugs tried)? Before writing, cheryl asks the pharmacist — one item at a time, from that payer's standard checklist — then completes the letter.
↻ feeds back to The Orchestrator (node 2)Every rep action flows back out as a training signal, so the system gets sharper with use.
Approved / denied / partial, logged with drug, payer, scenario — the ground truth.
The code a pharmacist fixes becomes the strongest correction signal.
Each rep decision emits a PHI-free training signal into the letter-training funnel — generated → submitted → labeled.
Nightly roll-up across clinics; ≥3 clinics required (privacy Safe Harbor), then human-reviewed before any rule change.