BusinessWednesday, May 27, 2026· 2 min read

Self-Improving Tax Agents: How Codex Automates Filings and Boosts Accuracy

Source: OpenAI Blog

TL;DR

OpenAI, Thrive, and Crete built a self-improving tax agent using Codex that automates tax filings, reduces errors, and speeds up workflows. The project showcases how agentic AI can continuously learn from feedback to deliver more accurate, scalable tax processing for businesses and accounting teams.

Key Takeaways

  • 1A Codex-powered agent was created in partnership with Thrive and Crete to automate tax filings and related workflows.
  • 2The agent uses a self-improvement loop—learn from feedback and refine behavior—to increase accuracy over time.
  • 3Teams see faster processing and fewer manual errors, letting staff focus on higher-value advisory work.
  • 4This demonstrates the real-world potential of AI agents to streamline regulated, detail-oriented business processes.

AI + accounting: a practical, self-improving solution

OpenAI partnered with Thrive and Crete to build a Codex-powered tax agent that automates filing tasks, extracts and formats required data, and iteratively improves its performance from corrective feedback. Rather than a one-off automation script, this agent is designed to learn from real-world interactions and edge cases, improving accuracy and reducing the need for manual fixes.

The project highlights three immediate benefits: faster workflows, fewer errors, and scalable operations. By automating repetitive steps in tax preparation and submission, the agent shortens turnaround times for routine filings. At the same time, the self-improvement loop helps the system handle tricky exceptions more reliably as it encounters more examples.

Teams that adopt this kind of agent can reallocate human expertise from repetitive processing to higher-value advisory and review work. Accounting teams and small businesses benefit from lower operational friction and more consistent compliance, while service providers can scale support without proportional headcount increases.

What this means for business automation

  • Automated filing and data handling reduce manual workloads and error rates.
  • Continuous feedback-based improvement makes the agent more robust over time.
  • Practical deployment with partners demonstrates agent-driven workflows are ready for real business use.

Overall, the Codex tax agent is an encouraging example of how AI agents can deliver concrete productivity gains in regulated, high-value domains like tax and accounting. It points the way toward broader adoption of self-improving automation across many business functions.

Get AI Wins in Your Inbox

The best positive AI stories delivered to your inbox. No spam, unsubscribe anytime.