OpenAI and Thrive Holdings have built a tax preparation AI that essentially teaches itself to get better. The system, deployed across Crete’s network of more than 30 accounting firms, started with a 25% accuracy rate at launch and climbed to 97% through autonomous feedback loops. The Tax AI uses OpenAI’s Codex models to transform real-world usage data into improvement signals, letting the system refine itself continuously without a team of engineers scrambling to patch every identified error.
From rough draft to near-perfect in six weeks
At launch, only 25% of returns hit a 75% accuracy threshold. Six weeks later, that figure jumped to 86%. The system kept climbing from there, eventually reaching up to 97% accuracy on drafted tax returns.
Preparation time dropped by roughly one-third, and overall throughput increased by approximately 50%. The AI produces tax engine submissions that are ready for human review, allowing tax professionals to focus on direct client interaction and complex judgment calls.
OpenAI’s quiet bet on service industries
