What the analyzer does
Medical Bill Reader accepts an uploaded medical bill or Explanation of Benefits document, sends it to an AI model for analysis, and returns a plain-English breakdown organized into five sections: what the document appears to be, visible charges and insurance fields, Amounts Shown, items to verify, and questions and next steps. The amounts section reports only clearly labeled figures visible in the document. It does not determine a final amount due or establish what the user legally owes. The output is a written explanation, not a diagnosis or a payment decision.
What model powers the analysis
The analyzer calls the Anthropic Claude API. The current model is claude-sonnet-4-6, Anthropic's general-purpose multimodal model. The API request is made server-side from a Next.js API route. The Anthropic API key is read from a Vercel environment variable and is never embedded in client code.
The model receives the uploaded file (image or PDF) plus a system instruction that treats document text as untrusted data and limits the response to five fixed sections. It is told to act as a cautious document explainer, not as a clinician, insurer, attorney, coder, or billing specialist. It does not look up codes against a live authoritative database, read your plan documents, or access records beyond the file you submit.
What the analyzer identifies
The model is asked to surface the items most useful to a patient:
- Procedure and service codes such as CPT (Current Procedural Terminology), HCPCS Level II, ICD-10-CM diagnosis codes, and NDC drug codes when present.
- Service dates and provider information as printed on the bill.
- Charge structure: total charges, the insurer-allowed or negotiated amount, plan payment, and patient responsibility (deductible, coinsurance, copay).
- Items to verify: exact-looking duplicates, visible mismatches, missing or unclear fields, unfamiliar services, or figures that do not reconcile from the document alone. These are questions, not findings that a charge, code, coverage decision, or party is wrong.
For a glossary of the specific code systems and abbreviations the analyzer references, see the codes explained page.
How bill data is handled
Medical bills contain sensitive personal and health information. Handling reflects that:
- Server-side processing: the file is sent over HTTPS to a Next.js API route on Vercel, which forwards the content to the Anthropic API and returns the response.
- No persistence: the bill content is held in memory only for the duration of a single request. The application does not write the bill to a database, an object store, or a log file. The response text is returned to your browser and is not retained server-side either.
- No application profile: the free analyzer does not require an account. A bill can still contain names, identifiers, and health information, so users are asked to redact unnecessary identifiers before submitting it.
- No advertising or analytics data flow: the analyzer does not load advertising or analytics code and does not emit upload, analysis, report, or payment events to those systems. The site also uses a no-referrer policy.
- Anthropic processing: the file is processed by Anthropic's API to produce the response. Per Anthropic's published standard commercial API policy, inputs and outputs are not used to train models by default and are automatically deleted from Anthropic's systems within 30 days. Anthropic says inputs and outputs flagged by automated trust and safety systems may be retained for up to two years and associated classification scores for up to seven years; legal, policy-enforcement, customer-controlled service, and agreed-term exceptions may also apply. This site does not claim a zero-data-retention agreement or Business Associate Agreement. Review Anthropic's retention policy and the Consumer Health Data Privacy Notice before uploading.
Honest limitations
The tool is useful as a first-pass plain-English explanation. It is not a substitute for professional review. Specifically:
- The analysis is informational, not medical or financial advice.
- The model can misread poor-quality scans, low-resolution photos, handwriting, or heavily redacted documents.
- The model cannot determine fraud, upcoding, unbundling, medical necessity, coverage, or legal compliance from a bill alone. Those questions require the underlying records, payer rules, and a qualified professional.
- The tool does not have access to the contracted reimbursement rate that the provider negotiated with your specific insurance plan. It cannot tell you what the "right" price should have been.
- Insurance regulations and patient protections vary by state and by plan type (employer, ACA, Medicare, Medicaid, TRICARE). The model cannot account for those differences.
- The model can hallucinate. Verify any specific code or amount it references against your itemized bill before acting on it.
Why this matters
Bills and EOBs can be difficult to compare. Official CMS guidance recommends checking that listed services and supplies match what was received and comparing the provider bill with the EOB. Separate appeal and No Surprises Act protections may apply depending on the plan and bill. The tool's limited goal is to organize visible information into questions a user can verify with the provider, insurer, or qualified adviser.
References
- CMS (Centers for Medicare and Medicaid Services): cms.gov.
- American Medical Association, CPT code overview: ama-assn.org/practice-management/cpt.
- HHS, federal No Surprises Act resources: cms.gov/nosurprises.
- Patient Advocate Foundation, Medical Billing Tip Sheet: patientadvocate.org.
- Anthropic, Claude model and API documentation: docs.anthropic.com.
- Related guides on this site: blog, codes explained.