Universal Document Processing at Gusto: From One-Off Parsers to a Self-Service Platform

Thomas Taylor, Praveen Awasthy23 Feb 2026

Problem

Gusto processes millions of documents a year: payroll tax forms, benefits enrollment, identity verification, compliance notices. Each type has regional and yearly variants, and every new product area or jurisdiction brought documents that needed their own parsing logic. Parsers broke silently when a field moved or an agency issued a new form version. Teams built similar pipelines independently, and every case automation could not handle went to a person, so manual review became the norm.

Example from the article: a state agency releases a new version of a form, the existing parser breaks silently or produces bad data.

Solution in brief

Gusto replaced per-document parsers with one platform of five stages, each behind its own API. What to extract is declared as configuration, and the model, prompt and fallback behind each stage can be swapped. Every extraction carries a confidence score that decides whether to accept, retry or send the document to a person.

Questions this case answers
  • How to extract structured fields from many evolving document types with one shared pipeline instead of one parser per type
  • How to route model outputs between automatic acceptance, retry and human review when the model can be wrong without warning
  • How to stop writing a new parser for every document type and every form revision?
  • How to decide which extracted fields can go straight into business systems and which need human review?

Tools and methods

  • Staged pipeline behind APIs — ingestion, classification, extraction, validation, mapping, each replaceable on its own
  • Declarative document schemas — a new document type is a config change, not a code release
  • LLM extraction and classification — different models per stage and per document type
  • Deterministic validation — hard rules (for example, an SSN must be 9 digits) next to model confidence
  • Confidence-based routing and human review — retry, escalate or reprocess when confidence is low
Continue reading — it's free

Google or a one-time email code. No password.