Better Bill GPT: How LLMs Performed Against Human Legal Invoice Reviewers 

92% invoice-level accuracy, compared to 72% for experienced lawyers. 22x faster than manual review. 
Source: Better Bill GPT, arXiv:2504.02881 

Legal invoice review is difficult to scale. A single invoice can contain hundreds of charges, each of which needs to be checked against billing guidelines, engagement terms, approved rates, and matter-specific instructions. As volume grows, maintaining that level of scrutiny across every invoice gets harder. 

Onit’s AI Center of Excellence set out to test whether large language models could help. The study pitted six general-purpose LLMs against experienced lawyers, early-career lawyers, and experienced legal operations professionals, all reviewing the same 50 invoices and 492 line items against a standardized set of billing guidelines. The results: the top-performing models outperformed experienced lawyers on accuracy, speed, and cost. 

In this white paper, you’ll find: 

• The full accuracy comparison: 92% invoice-level accuracy compared to 72% for experienced lawyers, plus the wider methodology behind that number 
• Why the gap widened even further at the line-item level, where the best-performing model scored nearly double the top human reviewers 
• The complete speed and cost breakdown behind the “22x faster” figure, including per-invoice processing cost 
• What happened when human reviewers used discretion instead of strictly enforcing billing guidelines 
• Five governance questions legal teams should answer before increasing automation in invoice review 

This white paper is designed for General Counsel, legal operations leaders, and finance professionals who want to understand where AI can reliably support invoice review, and where human judgment still needs to lead.