Tax season is nobody’s idea of a good time. Not for accountants, not for clients, and definitely not for the poor souls reconciling messy K-1 data at 2 a.m. OpenAI and Thrive Holdings think they’ve built something that can help: a self-improving AI system that drafts tax returns with up to 97% accuracy and gets better the more it’s used.
The pilot program ran through Crete Professional Alliance, a network of over 30 accounting firms, and processed 7,000 tax returns focused primarily on 1040 and 1041 filings. The results were, by enterprise AI standards, genuinely impressive.
How the system actually works
The Tax AI is built on OpenAI’s Codex technology, and its defining feature isn’t the accuracy number itself. It’s the self-improvement loop.
Here’s how it works. The system ingests direct feedback from practitioners, full production traces of prior corrections and filings, and then uses Codex to run targeted evaluations and code modifications in response to frequent issues. In English: every time a human accountant fixes something the AI got wrong, the system learns from that correction and updates itself.
That feedback loop produced measurable results over a remarkably short timeline. At launch, only 25% of returns hit 75% correct field completion. Six weeks later, that number had climbed to 86%. The system also posted gains at the 90% and 100% accuracy thresholds.
The initial deployment wasn’t aimed at simple W-2 returns that practically file themselves. It targeted medium- to high-complexity returns involving messy documents and K-1 data, the kind of work that eats up the most billable hours and produces the most errors.
Practitioners using the system reported saving roughly one-third of their preparation time, while throughput increased by approximately 50%.
The business relationship behind the tech
OpenAI and Thrive Holdings didn’t just shake hands on a licensing deal. OpenAI acquired an equity stake in Thrive Holdings back in December 2025, and their engineers collaborated for six months to build the system. The technical framework was formally announced on May 27, 2026.
One detail worth noting: Thrive Holdings owns the resulting intellectual property and products from the collaboration. That’s a somewhat unusual arrangement in the world of Big Tech AI partnerships, where the model provider typically retains significant IP rights.