Building self-improving tax agents with Codex

Thrive Holdings and OpenAI collaborated to develop Tax AI, an AI system that assists accountants with preparing complex tax returns. Traditional real-world systems often fail in unexpected ways, leading to slow and manual feedback loops for improvements. Tax AI addresses this by using a Codex-driven self-improvement loop, allowing it to learn autonomously from production use. This not only enhances accuracy and efficiency but also reduces the need for constant engineer intervention.

Over the past six months, this collaboration, including OpenAI forward-deployed engineers and Thrive Holdings’ engineers, developed Tax AI specifically for Crete’s network of over 30 accounting firms. The system processes tens of thousands of tax returns each season, navigating millions of underlying documents. For medium to high complexity filings, data entry alone can take up to eight hours per return, highlighting a significant bottleneck during peak tax season.

Tax AI significantly streamlines this process, automating much of the time-intensive work for 1040 and 1041 tax returns. Participating firms in the pilot processed 7,000 tax returns, demonstrating considerable efficiency gains. Notably, the system has shown measurable self-improvement, performing significantly better than its initial deployment three months prior.

Practitioners upload source files and client notes, and Tax AI generates a tax engine submission ready for review. This typically saves practitioners about a third of their time on tax preparation, drafts returns with up to 97% accuracy, and increases throughput by approximately 50%. This allows accountants to dedicate more time to client interaction.

Quantifiable improvements have been observed. At launch, only a quarter of returns achieved 75% correct field completion. Within six weeks, this figure rose to 86%, with even faster growth at the 90% and 100% completion levels. Initially, Tax AI handled simpler tasks such as W-2s and 1099s, but as the season progressed, it advanced to more complex returns involving K-1s and various schedules, continuously improving its capabilities.

This self-improving capability is built upon three pillars: expert practitioner feedback, production traces (structured histories from input to output), and a Codex-driven iteration loop based on tailored evaluations. This approach enables continuous and rapid product development, showing how practitioner expertise is critical for shaping system quality and data processing in complex domains.