🛠️OpenAI's Self-Improving Tax Agent Hits 97% Draft Accuracy
TL;DR
OpenAI published its tax-agent deployment with Thrive Holdings on May 27, taking draft accuracy from ~25% at launch to 97% within weeks by routing CPA corrections back into Codex as a training signal. Prep time fell about a third across 7,000 returns.
OpenAI published its tax-agent deployment with Thrive Holdings on May 27, taking draft accuracy from ~25% at launch to 97% within weeks by routing CPA corrections back into Codex as a training signal. Prep time fell about a third across 7,000 returns.
Key Points
Draft tax-return accuracy: ~25% at launch → 97% within weeks
7,000 returns processed; preparation time down roughly 33%
Built with Thrive Holdings and Crete Professionals Alliance over a six-month engagement
Practitioner corrections drive evals; Codex makes the targeted code changes
OpenAI took an equity stake in Thrive in December 2025; Thrive owns the resulting IP
Why It Matters
The self-improving feedback loop is the part nobody had shown working in production. Publishing one with audited numbers will pull every vertical SaaS roadmap toward this architecture.
Quick Facts
Frequently Asked Questions
Why does this matter?
The self-improving feedback loop is the part nobody had shown working in production. Publishing one with audited numbers will pull every vertical SaaS roadmap toward this architecture.
What happened?
OpenAI published its tax-agent deployment with Thrive Holdings on May 27, taking draft accuracy from ~25% at launch to 97% within weeks by routing CPA corrections back into Codex as a training signal. Prep time fell about a third across 7,000 returns.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.