Complex itemized bills from services such as cloud computing, healthcare, or telecommunications can be difficult to understand because the pricing terms that determine what a customer pays are straightforward, but the eligibility rules that determine whether a given line item, a single charge for a product and quantity, qualifies for a given price are scattered across user guides, service terms, and fine print. No single document describes them all, and without them, it is difficult to explain why a specific charge appears on a bill. We present a neuro-symbolic tool that learns these eligibility rules automatically from billing data and documentation. Pricing terms are evaluated against line items to produce labelled data indicating whether each pricing term was eligible for a given line item. An LLM agent then uses this labelled data, a knowledge base of public documentation, and a small set of rule-evaluation tools to generate rules that are accurate: no rule incorrectly labels a line item as ineligible for a pricing term, and explainable: every rule is backed by a passage from the knowledge base. We showcase the tool on the AWS billing system, evaluating millions of line items across hundreds of internal accounts. The tool learns a small set of rules that, together with the pricing terms, explain the majority of observed discrepancies, and we measure documentation coverage, rule stability across sample sizes, and the impact of ML pre-seeding on discovery cost.

@inproceedings{paperbot12424,
  title = {Why Is This on My Bill? Documentation-Grounded Rule Learning for Itemized Bill Explanation},
  author = {John Konstantinides and Jason Titus Lefever and Muhammed Numair Mansur and Martin Schaef and Willem Visser},
  booktitle = {ASE},
  year = {2026},
  doi = {10.1145/3832783.3834490}
}