OpenAI and Thrive Holdings have co-developed a self-improving AI system for tax preparation that was piloted across Crete’s network of more than 30 accounting firms during the recent US tax season.
The system, called Tax AI, processed around 7,000 tax returns and was designed to automate document extraction, field mapping, and portions of tax return preparation while keeping human accountants in the review loop.
According to OpenAI, the system reduced the time accountants spent on tax preparation by roughly one-third and increased overall throughput by about 50%.
Rather than operating as a fully autonomous filing system, Tax AI functions as a bounded automation layer. Human practitioners continue to review and approve final returns, while ambiguous cases are routed back to product and engineering teams.
The companies built the system on top of OpenAI’s Codex model using a three-part improvement loop in which practitioner corrections are captured as structured feedback, repeated failures are converted into evaluation datasets, and Codex is used to investigate issues, propose fixes, and validate improvements against regression tests before deployment.
OpenAI says the system “drafts returns with up to 97% accuracy,” though the blog’s primary benchmark focuses on “correct field completion” rates rather than a single aggregate accuracy metric.
When Tax AI was first deployed, only about a quarter of returns were completed with at least 75% of fields filled correctly before human review.
Within six weeks, that figure rose to 86%. The company also said nearly 60% of returns were at least 90% complete before correction, while around 9.5% required no field-level corrections.
Initially focused on simpler W-2 and 1099 workflows, the system later expanded into more complex filings involving K-1 forms, rental real estate schedules, and multi-document reconciliation tasks.
The broader significance of the project lies less in tax software itself and more in the engineering framework behind it. OpenAI and Thrive describe Tax AI as a model for building “self-improving agents” that learn from production feedback through structured evaluation pipelines and agent-assisted software iteration.
The companies say they are now applying the same architecture to bookkeeping, auditing, and operational workflows, such as IT help desk automation.
The deployment is part of a broader partnership announced earlier this year in which OpenAI took an equity stake in Thrive Holdings, aiming to modernise industries that still rely heavily on legacy operational systems.
ALSO READ: When Money Moves Itself: Why Agent‑Readable Banks Still Need Human Guardians
