OpenAI and Thrive Holdings Deploy Self-Improving Tax AI That Processes 7,000 Returns With 97% Accuracy

OpenAI and Thrive Holdings have deployed a self-improving Tax AI across 30+ accounting firms that processed 7,000 returns this season with 97% drafting accuracy, saving practitioners a third of their time while autonomously correcting its own mistakes through a continuous feedback loop.

• Thrive Holdings and OpenAI have co-developed Tax AI for Crete's network of 30+ accounting firms, processing 7,000 tax returns this season and saving practitioners roughly a third of their time while achieving up to 97% drafting accuracy.
• The system self-improves through a three-part loop: practitioners provide expert corrections, production traces convert those corrections into structured evaluation targets, and Codex autonomously investigates failures, implements fixes, and validates changes against targeted and regression evals before surfacing pull requests for engineering review.
• Starting from only 25% of returns hitting 75% correct field completion at launch, the system reached 86% within six weeks, with the same self-improving blueprint now being expanded to other domains including bookkeeping, audit, and IT help desk automation.