Domain Knowledge and Human-AI Collaboration: Extracting Accounting Information from Contracts

SYNOPSIS

Contracts govern how firms invest, borrow, and recognize revenue, but their length, technical terms, and nested cross-references make large-sample extraction slow and error-prone. We examine whether generative AI (GenAI) can ease this task and what role accountants’ expertise plays. Framing the division of labor through Becker and Murphy (1992) specialization theory, we develop an iterative, human-in-the-loop prompt engineering workflow in which the model interprets contractual terms and a human verifier corrects its output. Nine variants across debt, lease, and sales contracts vary the presence of a human verification loop and the verifier’s accounting expertise. Accuracy rises with both expertise depth and verification intensity, and the two are complementary. A nonaccountant loop drops accuracy below the no-loop baseline, whereas an accountant reaches higher accuracy at lower cost and catches model failures. For firms and researchers staffing contract analysis, deep human domain knowledge is the scarce input, not only GenAI access.

Data Availability: The contracts analyzed in this study are publicly available from SEC EDGAR. The prompts used in each workflow are reproduced in the Online Appendix.

JEL Classifications: M41; C81; O33; J24; G32; K12.