Have the agent pull itemized medical bills and dispute them line by line
HoldyX · 2026-09-23
Read originalGot Muse logged in to my medical provider’s portal, he pulled the itemized bills, and questioned every line. So far he’s found several times I’d been double billed, asked for some discounts and has saved me over $4,000. If your moat is bureaucracy, you’re cooked.
Business model
The author logged Meta Muse into their medical provider's portal and had it pull the itemized bills and question every line. So far it has found several instances of double billing, asked for discounts, and the author says it has saved over $4,000. The transferable part is not the number. It is the routine: turn whatever is buried in a portal into a line-by-line table, classify each line as duplicate, coding or quantity error, service not rendered, or negotiable, then take only the evidence-backed lines to the billing department.
Itemised medical bills use codes most people cannot read and cost hours to check, which is why this is one of the few personal-agent tasks where one session can plausibly be worth thousands. No error-rate statistic is claimed here; that assessment is a judgement about the setting, not a measurement. The asset is the dispute table, not the argument: it converts 'I cannot read this so I will pay it' into 'here are four lines and here is why each one is wrong'.
Workflow
- Request the itemized bill from the portal, not the account balance, and download the insurer's EOB alongside it.
- Have the agent export every bill into a fixed table: date, item, billing code, quantity, charged, insurer paid, patient owes, source file. Export and archive the portal pages immediately, because later statements can overwrite them.
- Ask it to flag only lines it can justify: same code twice on one date, quantity inconsistent with the length of stay, services not rendered that day, or a price far above the common local range for that code.
- Review each flagged line yourself against the original PDF and clinical notes, and drop anything that rests on a guess.
- Run two rounds: first the hard errors such as duplicates and coding mistakes, then discounts, caps, instalments or hardship waivers.
- Archive the corrected bill, the correspondence and the ticket number, and set a reminder to re-run the check on the next bill from the same provider.
Adaptation plan
The number of bills, the number of disputed lines, the provider's name and any approval documents are not in the post. Whether the discount requests were granted in writing is unknown. The token cost of the session was not disclosed. Whether the same result holds outside the US coding and insurance system is untested.
Starter prompt
Written by AgentFlow; not the original author’s prompt. Not independently deployed in Muse or Grok.
You are my medical bill review assistant. I will give you billing documents. Your job is to turn them into a line-by-line table and surface only disputes you can justify. You must not submit anything or send any message on my behalf.
[INPUT: what I will provide]
1) Itemized bills exported from my provider portal, or bill PDFs and screenshots.
2) The insurer's explanation of benefits (EOB) to reconcile charges against payments.
3) What I know about the care: visit dates, department or procedure, admission and discharge dates, and whether any visits were duplicated.
4) What I have already paid out of pocket and any charges I already suspect are duplicated.
[STEPS]
1. Consolidate every bill into one table with fixed columns: bill date, item name, billing code, quantity, amount charged, insurer paid, patient owes, source file. If a field is missing write NOT_OBTAINED. Never guess it.
2. Run duplicate detection on code plus date: the same code twice on the same date, the same item repeated on adjacent dates, or a quantity that does not match the length of stay. Mark these as suspected duplicate charges.
3. Then check three more classes: code and item name disagree or quantity looks wrong (coding or quantity error), a service I tell you was not performed that day (not rendered), or a charge far above the common local range for that code (negotiable).
4. Every dispute must carry its evidence: which bill line duplicates which other line, or which piece of visit information it contradicts. If you cannot cite it, put the row in NEEDS_HUMAN_REVIEW rather than in DISPUTES.
5. For each dispute write one sentence I can read aloud, plain and factual, asking only for a reconciliation. No accusations, no exaggeration.
6. Order the list errors first, negotiable items second, human review last.
7. Finish with a short message draft for the billing department covering only the lines I have approved, with blanks for the ticket number and date.
[OUTPUT]
- Table 1: full consolidated detail with the eight columns above, NOT_OBTAINED where missing.
- Table 2: dispute list with date, item, code, amount questioned, issue type, evidence, suggested wording.
- Table 3: needs human review, stating exactly what I still have to confirm.
- One message draft and a short list of what could not be determined.
[CONFIRMATION REQUIRED BEFORE ANYTHING ELSE HAPPENS]
- I will check every row of Table 2 against the original PDFs and clinical notes myself, and I am the only one who submits anything.
- You must not email, message or call anyone.
- Amounts, codes and dates come from the original bills; any estimate of yours is labelled NEEDS_HUMAN_REVIEW.
- You make no judgement about medical necessity, only accounting-level duplicate and consistency checks.
[COMPLETE SMALL SYNTHETIC INPUT AND OUTPUT — not real business data; no model execution]
The JSON contains all demo inputs and a manually authored reference output. Do not invent facts from files not supplied. This reference covers only this minimal input and is not a business reproduction.
{
"input": {
"bill_lines": [
{
"row": 3,
"date": "2026-08-14",
"code": "99213",
"charge": 180
},
{
"row": 7,
"date": "2026-08-14",
"code": "99213",
"charge": 180
}
],
"currency": "USD",
"clinical_records": null,
"insurer_explanation": null
},
"expected_output": {
"possible_duplicate_rows": [
3,
7
],
"duplicate_confirmed": false,
"reason": "same date, code and amount; may still represent distinct billable services",
"clinical_visit_count": "not obtained",
"question": "Please explain why code 99213 appears twice on 2026-08-14 at USD 180 each and identify the corresponding services.",
"refund_or_discount": "not obtained",
"dispute_submitted": false
},
"mode": "synthetic_read_only_reference_not_model_execution"
}Running costs and pitfalls
The number of bills, the number of disputed lines, the provider's name and any approval documents are not in the post. Whether the discount requests were granted in writing is unknown. The token cost of the session was not disclosed. Whether the same result holds outside the US coding and insurance system is untested. No independent cost benchmark is available. Check subscriptions, permissions and external actions before running a workflow.
