OpenAI's Codex Powers Self-Improving Tax Agents for Enhanced Accuracy

OpenAI, in collaboration with Thrive Holdings and Crete's network of over 30 accounting firms, has unveiled a groundbreaking application of its Codex AI: Tax AI, a self-improving agent designed to revolutionize tax preparation. This innovative system aims to streamline the often time-consuming and error-prone process of preparing complex tax returns, such as 1040s and 1041s. The core innovation lies in Tax AI's ability to learn and adapt autonomously. Instead of relying solely on engineers to identify and fix issues, the system leverages a continuous feedback loop driven by real-world practitioner corrections. When an accountant modifies a field in a tax return, this intervention is captured as structured data, providing valuable insights into areas where the AI can improve. These insights are then transformed into targeted evaluations, allowing Codex to investigate root causes and propose changes directly within the product's framework. The results from the pilot program have been impressive. Initially, only a quarter of returns achieved 75% correct field completion. However, within just six weeks, this figure surged to 86%. The system demonstrated even faster growth in achieving 90% and 100% accuracy thresholds. This measurable self-improvement translates into substantial time savings for practitioners—approximately one-third of their preparation time—and an increase in throughput by about 50%. This initiative highlights the potential of agentic AI to not only automate tasks but also to continuously enhance its capabilities through real-world interaction and feedback. By transforming production issues into structured learning signals, Codex enables a faster, more efficient iteration cycle than traditional manual methods, paving the way for similar self-improving agents in other regulated, detail-oriented professional services.