Skip to content

Engraved illustration of a rough sea. The picture is decorative; the site works without it.

Essay or blogExcerpt

Building self-improving tax agents with Codex

Aravind Srinivasan; Samay Shamdasani; Arthur Fernandes Araujo; John de Wasseige · OpenAI / Thrive Holdings ·

Research excerpt

Describes a practitioner-feedback and evaluation loop for Tax AI. Reports a 7,000-return pilot, roughly one-third less preparation time, and about 50% greater throughput.

The complete text is not republished here. This short description is editorial metadata, not the article itself.

Open original source ↗

Research notes and sources

Limitations: Joint participant report. Up to 97% accuracy is not a universal error rate or permission to remove professional review.

StatementEvidenceAs ofSources
OpenAI article reports drafting accuracy up to 97%. Current OpenAI (reporter)
Caveat

Not reconciled with later 98% claim.

Company reported1
OpenAI article reports throughput increase about 50%. Current OpenAI (reporter)Company reported1
Share of returns meeting 75% correct-field threshold rose from 25% to 86% in six weeks. Current
Caveat

A threshold distribution, not 86% field accuracy; raw eval data unavailable.

Company reported1
Codex proposes tested changes for engineering review; ambiguous cases return to product team. Thrive HoldingsCompany reported1
Thrive/OpenAI describe a pilot involving 7,000 tax returns. Thrive Holdings Current OpenAI (reporter)Company reported1
Thrive/OpenAI report approximately one-third less preparation time. Thrive Holdings Current OpenAI (reporter)Company reported1
Thrive/OpenAI report approximately 50% greater throughput. Thrive Holdings Current OpenAI (reporter)Company reported1
Availability
Excerpt
Rights basis
No public full-text republication basis recorded
Research access
reviewed

Continue reading

Building self-improving tax agents with Codex · AI Holding Company Research