The tax code doesn't care if you're tired, and now neither does your filing agent.
The Summary
- OpenAI partnered with Thrive and Crete to build a tax agent using Codex that automates filings, self-corrects errors, and gets faster over time
- The "self-improving" part matters: the agent learns from corrections and edge cases, building a feedback loop that compounds accuracy
- Tax preparation is a $14B market in the US alone, built on human expertise that now has a shelf life
The Signal
Tax work is rule-dense, deadline-driven, and expensive to get wrong. It's also exactly the kind of structured-but-complex domain where AI agents stop being demos and start being infrastructure. OpenAI's collaboration with Thrive and Crete shows what happens when you point a code-native LLM at a problem that's technically code: tax law is logic trees, deduction cascades, and if-then statements wrapped in legalese.
The Codex-based agent doesn't just fill out forms. It reads tax regulations, interprets client data, flags ambiguities for human review, and files electronically. More importantly, it logs every correction a human accountant makes and updates its internal reasoning. Next time it sees a similar case, it doesn't repeat the mistake. That's the self-improving loop, and it's what separates a useful tool from a compounding asset.
"The agent learns from corrections and edge cases, building a feedback loop that compounds accuracy."
Here's why this matters beyond tax season:
- Tax prep is high-stakes but low-variance once you know the rules—ideal training ground for agents
- The market is gigantic and relies heavily on seasonal labor that's hard to scale
- Mistakes are measurable and fixable, which means the agent gets better with every filing cycle
Thrive and Crete aren't household names, but they're in the business of compliance automation for mid-market companies. The fact that they're productizing this with OpenAI signals two things. First, tax and compliance work is moving from "human-reviewed" to "human-supervised" faster than most CFOs are ready for. Second, the bottleneck isn't the AI, it's the workflow integration and trust-building with firms that have liability on the line.
The technical play here is Codex, OpenAI's code-generation model. It translates natural language into executable logic, which is basically what a tax return is: a structured output derived from messy inputs. Feed it a 1099, a list of deductions, and the current tax code, and it can generate the forms, check them against IRS rules, and flag edge cases. The self-improving layer comes from fine-tuning on real corrections, building a corpus of "this situation means this outcome" that gets richer every quarter.
The Implication
If you're in accounting, compliance, or any domain where your value is "knowing the rules and applying them correctly," your time horizon just shortened. The question isn't whether agents can do this work—they can—it's whether your firm will adopt them or get outcompeted by one that does. For businesses, the play is finding partners who are already operationalizing these systems, not waiting for the Big Four to catch up. For technologists, tax is a wedge. Master the self-improving loop here, and you've got a blueprint for legal research, regulatory compliance, insurance underwriting, and any other rules-based domain that pays people to be right.