What “auditing a bill” actually means
Auditing a bill is pattern recognition. There are around two dozen categories of common billing error — duplicate charges, upcoding, unbundling, balance billing, expired promotional rates, estimated meter reads, equipment fees for returned hardware, ETFs miscalculated, and so on. Each category has a recognisable signature in the bill text.
A trained billing specialist scans for those signatures. AI does the same thing, faster, on more bills, without getting tired.
The pipeline
- Text extraction. A PDF bill is parsed for its text content. A photo bill is processed through OCR (Fix My Bill uses GPT-4o's vision model because it handles real-world creased, skewed, hand-annotated bills better than classic OCR).
- Sanitisation. Bill text can contain instructions that look like commands to the AI (“ignore previous instructions”). The pipeline redacts those before they reach the model — a defence against prompt injection.
- Contextual prompting. The model receives a system prompt with the audit rules, plus a country context (e.g. US medical: check CPT codes, look for balance-billing violations under the No Surprises Act) and a bill-type context (e.g. medical: focus on duplicate charges, upcoding, unbundling).
- Structured output. The model returns a JSON document with each finding — category (error/warning/info), title, description, line item, estimated overcharge, confidence score, regulation reference, and a dispute snippet.
- Validation. The output is validated against a strict schema. If it doesn't match, the scan is rejected and the credit is refunded.
What AI is good at
- Catching duplicates and unit errors across many line items — humans miss these on long bills.
- Decoding CPT/HCPCS, ICD-10, revenue codes, and tariff codes quickly across multiple jurisdictions.
- Identifying patterns like upcoding (level-5 ED for a sprain) and unbundling (services that should have been packaged).
- Spotting hidden fees — surcharges with official-sounding names that aren't government taxes.
- Cross-referencing the bill against the relevant law: No Surprises Act for US medical, Ofgem rules for UK energy, ACL/NECF for AU energy.
- Drafting dispute letters in seconds using country- and tone- appropriate language.
What AI is not good at
- Verifying that a service was actually rendered. Only your medical records or your memory can confirm that. AI can only flag suspicion.
- Reading badly photographed images. Get a flat, well-lit shot or use a PDF.
- Knowing your specific insurance plan. The audit pass identifies common violations; whether a charge violates your specific coverage requires the EOB and the plan document.
- Replacing legal advice. Findings have confidence scores for a reason. For large or complex disputes, a billing advocate or attorney is still worth it.
- Catching truly novel errors. AI is good at common patterns. A unique miscoding scheme may slip through.
Why a confidence score, not a yes/no
Every finding has a 0–1 confidence score. A 0.95 finding (“duplicate CT scan, same date”) is the kind you can dispute on the strength of the bill alone. A 0.6 finding (“possible upcoding”) is worth raising but may need clinical-record verification first.
Why findings always cite a regulation
A dispute letter is much more effective when it cites the specific legal basis for the dispute. Every finding includes the relevant regulation (No Surprises Act, Ofgem back-billing principles, ACL, etc.) so you can quote it directly.
AI bill auditing is decision support, not legal advice. The bill is yours, the dispute is yours, and the final call is yours. Verify findings with your provider and insurer before acting.