😺 Watch: AI can do your taxes now
The Neuron · Matthew Robinson · 2026-07-01
OpenAI and Thrive Holdings built Tax AI, a Codex-powered agent for complex tax return preparation that improves by converting accountant corrections into structured evaluation signals, demonstrating a feedback-loop architecture for expert AI workflows.
Extraction
Topics: ai-agentsexpert-feedback-loopstax-automationagentic-eval-designenterprise-ai-deployment
Claims
- Tax AI, built on OpenAI's Codex, can parse complex heterogeneous document types — PDFs, spreadsheets, images, handwritten notes, and tax forms — to prepare returns for accountant review.
- The system's core improvement mechanism converts expert corrections into structured signals: repeated accountant overrides on the same field surface measurable edge cases that engineers can fix and test before shipping.
- The harness — the instruction layer, workflow, review interface, evals, and deployment discipline — is the product, not the underlying model.
- In some cases during development, the model extracted the correct value while the existing ground-truth training data was wrong.
- Reviewing a mostly correct AI-prepared return requires substantially less effort than manual data entry, shifting expert work from production to review.
Key quotes
agents get safer when their mistakes are visible, measurable, and useful. The future of expert AI may look less like a chatbot taking over and more like a review system that learns every time the expert says, 'Nope, this field goes over here.'
The harness is the product: Arthur explains that the self-improving part is the instruction and workflow layer around Codex, not a vague promise that the model magically fixes itself.
before you build the agent, build the scoreboard.