OpenAI, Thrive, and Crete build self-improving tax agent with Codex

OpenAI, in partnership with Thrive Capital (a key investor) and Crete (a tax automation platform), has developed an agent using Codex—a model that converts natural language into executable code—to automate tax filings. This agent is designed to self-improve over time, adjusting its code generation based on new tax laws or client feedback. The move signals OpenAI's ambition to embed AI directly into enterprise workflows, particularly in finance and compliance. The US tax preparation software market alone is estimated at over $10 billion annually, making it a lucrative target. With OpenAI's valuation exceeding $80 billion and over $13 billion in funding, this partnership aims to drive API usage and enterprise licensing.

Existing players like Intuit's TurboTax rely on deterministic question-and-answer interfaces with some machine learning, but they do not self-improve or generate code autonomously. H&R Block's AI tools assist humans but do not file independently. TaxJar, now part of Stripe, automates sales tax but is narrow in scope. The Codex-based agent is different: it uses a generative model to write the actual filing logic, aiming for end-to-end automation without human intervention. This represents a chasm between assistive and autonomous compliance tools. However, the underlying model's tendency to hallucinate or produce incorrect code introduces a fundamental risk absent in rule-based systems.

The self-improving nature is both a feature and a liability. Tax laws vary by jurisdiction and change frequently. An agent that learns from past filings could inadvertently encode errors as patterns, leading to systematic mistakes. Moreover, the IRS and other tax authorities require accurate, auditable filings; an opaque model that generates code based on natural language prompts may not satisfy compliance standards. Data privacy is another concern—tax returns contain sensitive personal and financial information. The legal liability for errors is immense. Even with human oversight, the efficiency gains diminish. This tension suggests that fully autonomous tax agents are several regulatory cycles away from mainstream adoption.

The bottom line: OpenAI's tax agent is a powerful proof-of-concept but reveals the deep gap between AI automation and regulatory compliance. For enterprises, the short-term value may lie in assisted rather than autonomous filing, using Codex to draft code that humans then verify.