Introduction
Professional tax preparation is a labor-intensive process fraught with complex edge cases, but a collaborative project between OpenAI and Thrive Holdings is redefining efficiency. Their Codex-powered Tax AI isn't just an automation tool—it's a self-improving agent that learns from real-world practitioner feedback to get better over time.
News Analysis
News Title: Building self-improving tax agents with Codex (May 27, 2026)
Importance Score: 8.2/10
News Summary: OpenAI and Thrive Holdings co-developed Tax AI, a self-improving agent integrated with Crete's accounting firms, automating 1040 and 1041 tax return preparation while using Codex to turn production feedback into continuous accuracy gains.
Key insights from the project:
- Self-Improvement Through a Three-Part LoopTax AI's core strength lies in its iterative improvement cycle. It leverages practitioner corrections to identify failures, uses production traces to turn those corrections into structured evaluation targets, and relies on Codex to investigate root causes, implement fixes, and validate improvements. In just six weeks, the share of returns reaching 75% correct field completion jumped from 25% to 86%, with even faster growth at higher accuracy thresholds.
- Transforming Accountant WorkflowsThe AI cuts tax preparation time by 33%, boosts throughput by 50%, and achieves up to 97% accuracy. This frees accountants from manual data entry and calculation tasks, allowing them to focus on high-value client interactions. For example, one senior accountant reduced tax prep time from 180 hours to 15 hours, reallocating time to client consultations and business expansion.
- Scalable Model for Cross-Domain UseThe three-part loop developed for tax preparation is reusable across other complex tasks. The team has already applied it to additional tax schedules (Schedule C, Schedule A) and plans to expand into bookkeeping, audit, and IT help desk automation. Thrive Holdings' unique structure as owner-operator enables direct collaboration with practitioners, accelerating cross-domain adaptation.
Conclusion & Commentary
Tax AI represents a significant leap forward in bridging the gap between lab-based AI and real-world production systems. By centering practitioner expertise in the improvement loop, it addresses a critical pain point of slow, manual feedback cycles in AI development. The model proves that self-improving agents can deliver tangible, measurable value in professional services, enhancing both efficiency and human productivity.
As this approach expands to new domains, it sets a precedent for how AI can be integrated into industries where precision and human judgment are paramount. The collaboration between OpenAI and Thrive Holdings demonstrates that combining cutting-edge AI with on-the-ground practitioner insights is key to building robust, trusted, and continuously evolving systems.