OpenAI and Thrive Holdings spent six months building a Codex — powered Tax AI intended to assist a network of more than 30 accounting firms across Crete with increasingly complex tax return preparation. The collaboration produced an agent designed not only to support practitioners in handling complicated filings but to learn from day-to-day use. That matters for the firms involved because the system’s Codex — driven feedback loop is intended to reduce dependency on manual engineer interventions by turning real-world use and practitioner inputs into signals the agent can act on autonomously.
The development effort centered on adapting Codex capabilities to the specific demands of tax accounting workflows. Over the six-month period the partners concentrated on shaping the agent’s capabilities to align with the kinds of problems accountants face when preparing intricate returns, ensuring the tool could be integrated into a network of 30 — plus firms whose caseloads include higher — complexity filings. The stated goal was assistance and augmentation of practitioner work rather than replacing professional judgment.
At the heart of the system is a Codex — driven feedback loop that takes two primary inputs: production use data and direct practitioner inputs. According to the project description, those inputs are converted into structured signals that the agent consumes. The conversion process is framed as a mechanism for translating messy, real-world interactions — such as corrections, clarifications and usage patterns — into organized information the model can use to refine how it prompts itself and how it behaves in follow — up interactions.
Sources
Replies (0)
No replies in this topic yet.
