OpenAI, Thrive, and Crete Created a Self-Learning Tax Agent with Codex

OpenAI, together with tax platform Thrive and fintech company Crete, developed a self-learning AI agent that takes on manual work in preparing tax returns, freeing specialists for more complex tasks.

Who Built This and Why

The project combined three different areas of expertise. OpenAI provided Codex—a specialized model for working with code and structured data. Thrive, a major tax platform, brought deep knowledge of the tax process: which forms are filled out, in what order, where each piece of data goes. Crete provided integration with real financial systems and banking APIs. The result is a system that, in fully automatic mode, parses a set of taxpayer documents and fills in the necessary fields in the tax return. This is not simply OCR text recognition, but a full-fledged data analysis with logic and consistency checks.

How the System Works in Practice

The process begins with uploading documents: bank statements, income certificates, payment receipts, deduction information. Codex parses the structure of each document, identifies relevant data, and maps it to tax form fields. But that's not all. The key feature is logical verification. The system knows the rules: amounts across different lines must match, certain deductions are only possible with supporting documents, some income excludes others. If the agent finds a contradiction, it doesn't simply overlook it, but tries to understand what happened and correct the error.

"This is not just automation—it's a learning process.

With each processed document, the system gets better."

Self-learning is another important part. When a tax consultant reviews a form filled out by the agent and makes corrections, this information is not simply archived. The system analyzes where it made mistakes and adjusts its rules.

Results in Numbers and Practice

• Speed: Processing a single return now takes minutes instead of 2-3 hours of manual work
• Accuracy: Errors have significantly decreased thanks to built-in logic verification
• Scalability: One AI agent can process thousands of returns around the clock
• Costs: Thrive clients can reduce spending on tax consultants by approximately 40-60%
• Time savings: Consultants now focus on complex cases and consulting instead of routine work

The pilot has already begun with a group of Thrive clients. Results show that even in the first version, the system handles most standard returns without errors.

What This Means for the Industry

This is a serious signal about how AI is moving from laboratories into real business processes. Tax returns are an ideal use case for such an application: the process is strictly regulated, the logic is clearly formalized, and the volume of documents is enormous. If the approach works in the tax sphere (and preliminary results suggest it will), similar systems can be adapted to accounting reports, insurance claims, credit applications, and other documents. This could lead to a wave of productivity in office functions where tens of millions of people work.