OpenAI details self-improving tax agents built with Codex

OpenAI says it built a tax-preparation system that gets better by learning from real work, not just lab tests. The project, called Tax AI, was co-developed over the past six months by OpenAI forward-deployed engineers and researchers working with Thrive Holdings engineers and Crete’s network of more than 30 accounting firms.

According to OpenAI, the system was designed to help accountants prepare increasingly complex tax returns during the busiest part of tax season. Crete practitioners handle tens of thousands of returns each season and work through millions of underlying documents, the company said. For medium- to large-complexity filings, OpenAI said data entry alone can take up to eight hours per return because accountants must deal with messy source files, prior-year documents, and manual extraction and calculation.

OpenAI said Tax AI processed 7,000 tax returns across the Crete firms that took part in the pilot this tax season. The system automates much of the work involved in preparing 1040 and 1041 tax returns, and the company said it now saves practitioners about a third of their time on tax preparation. OpenAI also said the tool drafts returns with up to 97% accuracy and increases throughput by about 50%.

The more notable claim, though, is that the system improved during deployment. OpenAI said Tax AI is “measurably better” than the version first deployed three months ago. The company said it measures that progress by looking at how many returns reach 75%, 90%, or 100% correct field completion.

At launch, OpenAI said only a quarter of returns were at 75% correct field completion. Within six weeks, that figure rose to 86%, with faster gains at the 90% and 100% levels as well. The company said those thresholds are meant to show how much practitioner follow-up each return still needs.

OpenAI said the system started with simpler work such as W-2s and 1099s, then moved into more complex returns involving K-1s, schedules, and harder edge cases. The company said each new capability saved more time than the last because the tasks were harder and more time-consuming to do manually.

The company’s explanation centers on three parts: expert practitioner feedback, production traces, and a Codex-driven improvement loop. In OpenAI’s description, production traces capture the path from source documents to extracted fields, mapped tax-engine values, and practitioner corrections. Those traces are then used to create targeted evals, or evaluation sets, that Codex can work against.

OpenAI said this matters because not every correction means the model made a mistake. A changed value could reflect an extraction miss, a mapping issue, a value carried over from a prior-year return, or ordinary workflow noise. The company said practitioners helped sort those cases so the system could distinguish actionable failures from expected variation.

In one example, OpenAI described rental property income on Schedule E. The system reads source material, extracts fields with citations, maps those values into the tax engine, and then preserves enough evidence for a practitioner to approve or correct the result. Repeated corrections are grouped into findings, then turned into eval targets. Codex can then inspect the trace, repo, and evals, propose changes, validate them, and surface a candidate pull request for engineering review.

OpenAI said the same pattern can apply beyond rental properties. It said the work on rental-property filings helped create reusable abstractions and review artifacts that can support other complex schedules such as Schedule C and Schedule A. The company also said it is using the same approach as a blueprint for other workflows inside Thrive Holdings, including bookkeeping, audit, and IT help desk automation.

One senior accountant cited by OpenAI said she spent 180 hours on tax prep last year and only 15 hours this year. According to OpenAI, she used the extra time to call clients and expand her service offerings.

OpenAI said the broader goal is to build agents that improve over time by combining practitioner expertise, structured production data, and eval-backed engineering. The company said ambiguous cases still route back to product teams rather than being forced through automation.