OpenAI and Thrive Build Self-Improving Tax Agent With 97 Percent Accuracy

What this is about

In late May 2026, OpenAI and the company Thrive Holdings introduced a self-improving tax agent built with OpenAI's Codex technology. In a pilot through the Crete Professional Alliance, a network of more than 30 tax firms, the system processed 7,000 tax returns, mostly US forms of types 1040 and 1041. OpenAI had taken a stake in Thrive Holdings in December 2025.

What the system actually does

The technically interesting part is less the accuracy than the feedback mechanism. The system records full traces: source file, extracted field, the AI's mapping, the accountant's correction, and the result ultimately filed. When the same error occurs repeatedly, it is bundled into a testable engineering task that Codex then fixes automatically. According to the figures, the system reached 97 percent accuracy, cut processing time by about a third, and increased throughput by roughly 50 percent.

Why it matters

This is what self-improving AI looks like in practice: not a system that teaches itself to think, but a bounded system that detects recurring errors, turns them into solvable tasks, and ships the correction without a human writing code each time. For knowledge-based services such as tax advice, accounting, or law, this signals an architecture in which an AI agent works with a correction loop. Tax processing is just the first field of application.

In plain language

Imagine an apprentice who writes down every mistake in a notebook. If the same mistake happens often, it becomes a fixed rule the apprentice then follows automatically. That way the master does not have to step in every time, and the work becomes more reliable over time.

A practical example

A tax firm in Austria with 12 employees processes hundreds of returns in high season. With an agent built on this pattern, routine cases could be pre-filled and only reviewed by staff. Recurring mapping errors, for example with a certain type of receipt, would be collected and fixed once instead of being corrected manually each time. Before deployment, a review under the EU AI Act and data protection, a test on anonymized past cases, and clear human approval rules would be necessary.

Scope and limits

First, the cited figures on accuracy, time, and throughput come from the involved companies' account and are not independently verified. Second, tax law is country-specific; a system trained on US forms is not readily transferable to the DACH region. Third, 97 percent accuracy means errors remain, so a final human check and clear liability rules are indispensable for legally binding filings.

SEO & GEO keywords

OpenAI, Thrive Holdings, Codex, AI tax agent, self-improving AI, Crete Professional Alliance, tax return, agentic AI, 97 percent accuracy, accounting, professional services, 2026