AI Judge changed title after evaluation, original title: "OpenAI's Boris Power builds a self-improving tax agent that reached 90% completion on nearly 60% of tasks"
The system increased tax firm throughput by 50%.
Brandon McKinzie#1004
At @ThriveHoldings, we built a product with @OpenAI to automate tax prep for the 30+ accounting firms we own across the country.
This season, it processed 7k+ returns. But what I think is more interesting is that the product meaningfully self-improved as accountants used it.
Many users praised OpenAI and Thrive's self-improving tax AI for delivering major time savings in tax prep and fitting verifiable domains well, while others flagged risks of error compounding and added regulatory overhead.
No Digg Deeper questions have been answered for this story yet.
OpenAI for self-improving tax agents:
At @ThriveHoldings, we built a product with @OpenAI to automate tax prep for the 30+ accounting firms we own across the country.
This season, it processed 7k+ returns. But what I think is more interesting is that the product meaningfully self-improved as accountants used it.
Read more about our work in detail with @OpenAI below.
https://openai.com/index/building-self-improving-tax-agents-with-codex/
The unlock was a self-improvement loop.
We record production misses: unsupported fields, wrong predictions, and corrections. Codex then uses that context to autonomously create evals from production data, hillclimb against them, and open candidate PRs for engineers to review. ⚡️
OpenAI and Thrive just built a self-improving tax agent with up to 97% accuracy.
Tax AI processed 7,000 returns across 30+ accounting firms, saved about one-third of preparation time, reached up to 97% accuracy, and raised throughput by about 50%.
The hard part was not reading W-2s or 1099s, but handling messy K-1s, rental schedules, notes, spreadsheets, prior-year files, and values that must match across documents.
The system records the full trace: source file, extracted field, citation, tax-engine mapping, accountant correction, and final filed value.
Repeated corrections become eval targets, so Codex gets a narrow task with evidence, code, tests, and a pass condition.
A wrong tax field can come from many places: bad extraction, weak mapping, unsupported workflow, prior-year carryover, or human judgment.
The clever part was not simply using Codex to write fixes, but building a product environment where repeated practitioner corrections became bounded, testable engineering tasks.
In the rental-property example, the agent could inspect source documents, extraction traces, mapper behavior, expected outputs, and regression tests before proposing a change.
I really appreciate the lessons and technical ideas @samaysham & team were able to share about their tax agent system, which learns from production traces to self-improve via detailed tracing tightly integrated into deployment + an autonomous AI engineer.
At @ThriveHoldings, we built a product with @OpenAI to automate tax prep for the 30+ accounting firms we own across the country.
This season, it processed 7k+ returns. But what I think is more interesting is that the product meaningfully self-improved as accountants used it.
Building evaluations and fast feedback and iteration loops is the most reliable way of solving hard problems and improving the product rapidly.
Here, the amazing Crete team with collaboration from OpenAI automated majority of the field extraction work, done during tax preparation within weeks!
A glimpse of an exciting future where tax professionals will be able to spend more time advising customers and explaining the tax returns!
At @ThriveHoldings, we built a product with @OpenAI to automate tax prep for the 30+ accounting firms we own across the country.
This season, it processed 7k+ returns. But what I think is more interesting is that the product meaningfully self-improved as accountants used it.
Read the full blog on how this was achieved: https://openai.com/index/building-self-improving-tax-agents-with-codex/
Building evaluations and fast feedback and iteration loops is the most reliable way of solving hard problems and improving the product rapidly.
Here, the amazing Crete team with collaboration from OpenAI automated majority of the field extraction work, done during tax preparation within weeks!
One accountant told us she spent 180 hours on tax prep last year. This year she spent 15.
She used that time to call every client and walk them through their return.
That’s what gets me excited. Less time buried in prep and more time as a strategic advisor to clients.
https://openai.com/index/building-self-improving-tax-agents-with-codex/
OpenAI and Thrive just built a self-improving tax agent with up to 97% accuracy.
Tax AI processed 7,000 returns across 30+ accounting firms, saved about one-third of preparation time, reached up to 97% accuracy, and raised throughput by about 50%.
The hard part was not reading W-2s or 1099s, but handling messy K-1s, rental schedules, notes, spreadsheets, prior-year files, and values that must match across documents.
The system records the full trace: source file, extracted field, citation, tax-engine mapping, accountant correction, and final filed value.
Repeated corrections become eval targets, so Codex gets a narrow task with evidence, code, tests, and a pass condition.
A wrong tax field can come from many places: bad extraction, weak mapping, unsupported workflow, prior-year carryover, or human judgment.
The clever part was not simply using Codex to write fixes, but building a product environment where repeated practitioner corrections became bounded, testable engineering tasks.
In the rental-property example, the agent could inspect source documents, extraction traces, mapper behavior, expected outputs, and regression tests before proposing a change.
@thesephist 🚀
@gdb POV: You lost the game of Rawls’ Veil of Ignorance, and were born as a sentient tax AGI agent.
@OpenAI And thank you to my incredible colleagues who made everything we've accomplished possible: @aravind_svn, Arthur (OAI), John (OAI), @haversang, @tuckerhaas, @snbafana, @stephensu_, Jon, @sachityagi, Vamsi, James, Jared, @AnujMehndiratta, & @kareemszaki. 🚀
@gdb GPT-4o for self improving, return it!!!
@gdb Tax agents sound like more regulatory overhead.
@gdb cool, now do self-improving tax code that deletes loopholes and audits itself.
@gdb A tax agent earns trust in the redlines: missed deduction, weird receipt, partner note, reviewer override. Self-improvement only matters if each correction makes the next return easier to audit, not harder to explain.
@gdb 7k returns logged and the prompts got tuned on the fixes. same loop we built for contract review tools years ago
@gdb Greg, I dont understaand CODEX at all (for finance) for making my own dashboards etc.. is this happening in the future?
@gdb The revenue service as an O1-style reasoning loop—each audit trains the next. Probably the most defensible govtech moat there is.
