OpenAI's Codex-powered Tax AI automates tax prep with self-improving agents

OpenAI engineers, collaborating with consulting firm Thrive Holdings, have deployed a system called 'Tax AI' that leverages Codex to automate tax return preparation for accounting firms. The system generates code directly from natural language instructions, producing complete tax forms. Its defining feature is a self-improvement mechanism: practitioners review outputs, flag errors, and the agent updates its behavior. Initial reports claim production enhancements, but this is far from a simple automation play — it is a live experiment in agentic AI applied to a domain where errors carry real financial and legal weight.

The competitive landscape highlights why this matters. Intuit’s TurboTax uses machine learning to optimize deductions for individuals, but its AI is a passive assistant, not an autonomous agent. Botkeeper automates bookkeeping with a human-in-the-loop model, but its tax features are secondary to broader accounting workflows. Blue J Legal predicts tax law outcomes, aiding research rather than preparation. Tax AI is distinct: it is an active agent that writes its own code to execute end-to-end preparation, then evolves based on feedback. This represents a shift from static AI tools to dynamic, learning systems that could eventually operate with minimal human oversight. The US tax preparation market alone is $11 billion annually, and OpenAI likely generates API revenue from Codex usage — potentially a lucrative new revenue stream for the company if this model scales.

The hidden risks, however, are severe. Tax law is a minefield of jurisdictional edge cases and ambiguous rules. A self-improving agent that learns from practitioner feedback inherits the biases, inconsistencies, and gaps in that feedback. Without a rigorous validation framework, mistakes could propagate, leading to incorrect filings, audits, or legal liability. Data privacy is another concern: tax returns contain highly sensitive personal and financial information, and any breach could erode trust. Moreover, the system must comply with IRS standards and regulatory requirements — a challenge that the announcement does not address. AI researcher Gary Marcus called it an 'interesting proof of concept,' but warned that regulators may push back without rigorous validation. The reliance on human feedback also raises scalability questions: as the system grows, ensuring quality oversight from thousands of practitioners becomes a logistical nightmare. Finally, the potential for job displacement among tax preparers, while not a technical flaw, introduces ethical and social friction that could slow adoption.

This deployment is a bellwether for agentic AI in professional services. If Tax AI succeeds, it will validate the model of self-improving agents in regulated, high-stakes domains, accelerating similar efforts in law, medicine, and accounting. If it fails — due to undetected errors, regulatory pushback, or privacy lapses — it will reinforce the need for gated deployment and rigorous safety architectures. The bottom line: Tax AI is not just a product; it is a live case study in the tension between the promise of autonomous agents and the messy reality of real-world expertise. The feedback loop that powers its improvement is also its Achilles' heel.