A health plan must answer a prior-authorizationPrior authorization: a plan's approval that a treatment is covered before it happens. CMS-0057-FCMS-0057-F: a US rule requiring plans to answer prior-auth requests within 7 business days (urgent: 72 hours from 2027), exposed over a standard health-data format. sets the deadlines. request in days — and show its reasoning if a regulator asks. A generic chatbot just guesses from text. In healthcare, a confident wrong answer is a real risk.
What MediGovern is given to answer one question
Policy PDF
Holds
Coverage & prior-auth rules
Example
MRI criteria
Benefit table
Holds
What this plan covers
Format
CSV per procedure
FHIR record
Holds
Maria's clinical context
Source
Member system
The question
Asks
Is a lumbar MRI covered?
Code
CPT 72148
3 governed documents + 1 question → one cited answer, or an honest "not sure".
So what
The goal isn't a chatty answer — it's a defensible one: grounded in the real policy, with its source, and logged so the decision can be replayed.
2
Load the governed documents — labelled, traceable
Policy PDFs, benefit tables and the member's FHIRFHIR: the standard format health systems use to exchange data (records, claims, coverage). Loading FHIR-aware means the system understands those fields, not just raw text. record are read in and tagged — so every fact can be traced back to the exact document it came from.
What happens behind the screen
3 source documentspolicy · benefits · FHIR
→
Loaded & taggedCPT · ICD · payer · plan
→
Traceable to sourcenothing anonymous
Why tag everything? When the answer later cites "Radiology Policy §4.2", the system can point to the exact source document and page — that's what turns an answer into evidence.
So what
Like stamping every page with where it came from before filing it — so any claim can be traced back, not taken on faith.
3
A six-rule quality gate — before anything is used
Every document passes a data-quality gateData-quality (DQ) gate: a set of checks each document must pass before it's trusted. Each rule returns pass, warn, or quarantine. first. A document that fails — say, a stale policy version — is quarantinedQuarantined: held aside so a faulty or outdated document can't influence an answer. It isn't deleted — it waits for review., not silently used.
Six rule families (sample results)
96%
91%
98%
99%
88%
Caught here: an older copy of the radiology policy is flagged as superseded and quarantined — so the answer is built only from current, valid rules.
So what
Garbage in, garbage out — so the gate stops the garbage at the door. Bad or outdated rules never reach a coverage decision.
4
Turn documents into searchable evidence
Clean documents are split into small chunks and turned into embeddingsEmbedding: a numeric "fingerprint" of a piece of text that lets a computer find passages by meaning, not just exact keywords. stored in a vector storeVector store (pgvector): a database that searches by meaning. Ask "MRI for back pain" and it finds the relevant policy paragraph even if those exact words don't appear. — each chunk indexed by its metadata.
From document to findable evidence
Clean documents
→
Split into chunks
→
Embedded into pgvectorindexed by metadata
This is the "R" in RAGRAG (Retrieval-Augmented Generation): the AI first retrieves the relevant trusted text, then writes its answer from it — grounded in your documents, not invented.: retrieval. The AI will pull from these governed chunks — never from the open internet.
So what
It's a librarian indexing the policy so the right paragraph can be found in milliseconds — by meaning, not just matching words.
5
Filter first, then search by meaning
Maria's request doesn't search everything. It first filters to the right plan and procedure, then runs a meaning-based search — so it lands on the exact governing paragraph.
Retrieving the right rule
"Lumbar MRI · CPT 72148 · this plan?"
→
Metadata pre-filterthis plan · this procedure
→
Semantic search
The question
Is a lumbar MRI (CPT 72148) covered for Maria's plan?
0.91top match
Radiology Policy §4.2
MRI lumbar spine: covered with prior auth if ≥6 weeks conservative therapy is documented.
So what
The pre-filter is what keeps it honest — it can only answer from this member's plan and procedure, not a lookalike policy from somewhere else.
6
An answer with its receipt attached
The system writes a groundedGrounded / cited: the answer is built only from the retrieved policy text, and it shows which passage it used — so it can be checked, not just trusted. answer from the retrieved policy and shows the citation, plus a confidence bandConfidence band: HIGH / MEDIUM / LOW, computed from how strong and consistent the evidence is — not a medical opinion. for how strong the evidence is.
Maria's coverage answer
✓
Covered — with prior authorizationHIGH confidence · 3 cited sources
Every line is traceable. If a reviewer or regulator asks "why?", the citation points straight to the governing paragraph.
So what
The answer isn't "trust me" — it's "here's the rule, here's where it's written, and here's how sure I am." That's what makes it usable in healthcare.
7
A confident "I don't know" beats a confident guess
This is the heart of the governance. If the retrieved sources conflict — or none clearly applies — MediGovern ABSTAINSABSTAIN: the system deliberately refuses to answer when the evidence is conflicting or insufficient, and routes the case to a human instead of guessing. rather than inventing coverage.
What happens when two policies disagree
Policy v2 (older)
MRI lumbar spine: not covered without specialist referral.
?they conflict
Policy v3 (current)
MRI lumbar spine: covered with prior auth and conservative therapy.
✓ Sources agree → cited answer
✕ Sources conflict → ABSTAIN, route to a human
So what
A system that knows the limits of its evidence is safer than one that always sounds certain. ABSTAIN is a feature, not a failure.
8
Two different things people often confuse
The confidence band measures how strong the evidence is — not whether the treatment is medically right. They answer different questions.
Confidence: HIGH
"How strong is the evidence behind this answer?"
HIGH 3 cited sources agree
vs
Clinical decision
"Is this the right care for Maria?"
Human a clinician's call
Read it as: "We're highly confident the policy says this is covered with prior auth." It does not replace a clinician deciding what's appropriate.
So what
Confidence measures trust in the evidence. The clinical call stays with people. Never present the two as the same thing.
9
Every decision is logged — and replayable
The whole decision is written to an immutable audit packetAudit packet: a tamper-evident record of one query — the question, the exact chunks retrieved, the answer, the citations, the confidence, and the time. Exportable as JSON + HTML.. When a regulator asks "how did you decide on this case?", the answer already exists.
What's recorded for Maria's query
Field
Logged value
Question
Lumbar MRI (CPT 72148) coverage — Maria's plan
Evidence used
3 chunks · Radiology Policy §4.2 + benefit table
Answer
Covered with prior auth · ≥6 wk conservative therapy
Confidence
HIGH (sources agree)
Timestamp
Recorded, immutable
So what
Compliance stops being a scramble. The proof of how every answer was reached is captured automatically, the moment it's made.
10
Governed AI, not "chat with a PDF"
That's the journey. A coverage question became a fast, cited, auditable answer — with the discipline that healthcare and regulators require built in underneath.
1 questionasked in plain English
→
1 cited answergrounded · confident · logged
…or an honest ABSTAIN when the evidence conflicts. It never invents coverage — and never touches real PHIPHI (Protected Health Information): identifiable patient data. MediGovern runs entirely on synthetic data, so there's none to leak..
Faster decisionsPrior-auth answers in seconds, each cited.
Regulator-readyEvery query logged and replayable.
Safe by designABSTAIN on conflict · 0 PHI · 100% synthetic.
So what
The governance lives underneath the AI — a quality gate before answers, citations on every claim, an honest ABSTAIN, and a full audit trail.