How a bill is checked

1. Reading the documents

First, each document is checked for what it is: a bill, an Explanation of Benefits, a Good Faith Estimate, or something else, and a document that is plainly something else is turned away rather than read as one of them.

Then an AI model from Anthropic transcribes what is printed: provider, dates of service, billing codes, units, charges, allowed amounts, what the plan paid and what it says you owe. It copies figures; it does not interpret them. It is instructed to mark anything it cannot read as unreadable rather than guess, and a check with unreadable fields is reported as incomplete and is not offered for sale.

2. Matching the lines

Each line of the bill is paired with the line on the Explanation of Benefits for the same service, by billing code and date of service. The most certain pairs are made first — same code, same date — and a line paired with confidence is never reconsidered by a looser pass. Lines that pair with nothing are kept, because a charge with no match is itself something to look at.

3. The rules

There are fifteen rules, and every finding comes from exactly one of them. Which ones can run depends on what you send: without an Explanation of Benefits, or with a summary bill, the rules that need them do not run, and the check says so.

1. The bill asks for more than the Explanation of Benefits says you owe.
2. The plan denied a line because of a provider-side administrative error — late filing, a coding error, a missing authorization — and you are being billed for it anyway.
3. The same service is billed more times than the plan processed it.
4. The bill claims more units of a service than the plan processed.
5. A charge on the bill does not appear on the Explanation of Benefits. The most common cause is a line that was not read, so this is never presented as a confirmed error.
6. The bill was issued before the plan finished processing the claim — or there is no Explanation of Benefits yet.
7. An in-network provider is billing above what its contract lets it collect.
8. An out-of-network provider is billing for emergency care above in-network cost sharing, which the No Surprises Act limits.
9. An out-of-network provider at an in-network facility is billing beyond in-network cost sharing, which the No Surprises Act also limits.
10. You are paying without insurance and the bill is at least $400 above your Good Faith Estimate, which makes it eligible for the federal dispute process.
11. The hospital is a nonprofit, and must run a financial assistance program.
12. The bill is a summary with no line detail, so the line-level rules cannot run — and an itemized statement is the thing to ask for.
13. The same service appears several times on one date.
14. The plan's own figures — deductible, copay, coinsurance and anything not covered — do not add up to what it says you owe.
15. A document shows a credit rather than a balance owed.

How findings are reported

• Clear discrepancies are places where the two documents disagree about a number.
• Worth asking about are things to question where the amount cannot be known from the documents alone.
• The amount in question counts each dollar once. Two findings often describe the same dollars from different angles — a duplicate charge also makes the balance too high — so the findings are not added together.

What a check cannot know

Only what is on the documents. Not your income, which decides financial assistance. Not the terms of your plan beyond what the Explanation of Benefits states. Not whether the provider will agree. A finding is a precise question to ask, not proof that a bill is wrong.

How it is tested

The rules and the matching are covered by more than 2,000 automated tests. The reading is scored against a set of made-up bills and Explanations of Benefits whose correct answers are known, and each score is recorded, so a change that makes the reading worse shows up. Those test documents are clean PDFs; photographs and scans are harder to read than anything in that set, which is one reason unreadable fields are marked rather than guessed.

Last reviewed 27 September 2026.