OpenAI's Codex-powered Tax AI automates tax prep with self-improving agents
The most important lesson from OpenAI's Tax AI isn't that it automates tax returns — it's that the feedback loop for self-improvement may be its greatest vulnerability, not its strength.
On May 27, 2026, OpenAI's engineering blog detailed the co-development of 'Tax AI' with Thrive Holdings for Crete accounting firms. The system, built on OpenAI's Codex, automates much of the tax return preparation process. By fusing practitioner expertise with a Codex-driven feedback loop, the AI agents can self-improve in production. The collaboration highlights how careful eval infrastructure enables measurable enhancements, turning static automation into adaptive, learning systems.
The key innovation is the feedback loop: tax professionals review outputs, and those corrections feed back into the model via Codex, allowing continuous refinement without manual retraining. This approach addresses a critical challenge in AI deployment—maintaining accuracy over time. For Crete accounting firms, this means fewer errors, faster processing, and the ability to handle complex tax scenarios. OpenAI's blog emphasizes that such self-improving agents could set a new standard for professional AI tools, where domain expertise directly shapes model behavior in production.
- Tax AI is the first known end-to-end use of Codex for tax preparation, marking a shift from passive AI tools to self-improving agents.
- The self-improvement feedback loop introduces bias and inconsistency risks that require robust validation to avoid cascading errors.
- OpenAI's partnership with Thrive Holdings points to a growing enterprise revenue model for API-based agentic AI, potentially tapping an $11 billion market.
Why It Matters
This deployment tests whether agentic AI can safely handle high-stakes tasks that demand both broad knowledge and local nuance.