In a May 27, 2026 post, OpenAI engineers and collaborators at Thrive Holdings outlined how they used Codex to create a self-improving Tax AI deployed with Crete’s network of accounting firms. The pilot processed about 7,000 tax returns and, according to the teams, reduced practitioner time on preparation by roughly one third while increasing throughput by about 50% and producing draft returns with up to 97% accuracy.
Pilot results and measurable improvement
OpenAI and Thrive report that accuracy was tracked by share of returns reaching 75%, 90%, or 100% correct field completion. At launch, roughly 25% of returns hit the 75% threshold; within six weeks that figure rose to 86%. Early coverage handled simpler forms such as W-2s and 1099s and later expanded to more complex filings including K-1s, Schedule E, and other reconciliations for 1040 and 1041 returns.
The three-part improvement loop
The teams describe three pillars that enabled continuous improvement: 1) practitioner feedback to surface which errors matter, 2) production traces that record the full path from source documents to final filing, and 3) a Codex-driven iteration loop that turns repeated, reviewed failures into bounded eval targets and candidate engineering changes. The product captures corrections as structured data so each practitioner intervention becomes a traceable signal.
How a field-level failure becomes a Codex task
An illustrated example focused on rental property income (reported on Schedule E). When practitioners repeatedly corrected a field such as the “fair rental days” value, the system grouped related review rows, separated recurring failures from workflow noise, and created a targeted eval set. Codex then inspects traces, repo artifacts, and evals to investigate root causes, propose targeted fixes (schema, source selection, mapper, or grader changes), and validate them against regression suites. Ambiguous cases are routed back to engineers rather than forced through automation.
According to the teams, rental-property work required about six weeks and substantial engineering oversight to reach roughly 90% precision and recall, but the patterns, artifacts, and eval conventions made it easier to tackle other complex schedules such as Schedule C and Schedule A. Engineers retain responsibility for architecture and shipping, while practitioners steer the improvement loop through their corrections and reviews.
Original source: OpenAI News