How Thrive and OpenAI turned tax-review clicks into a self-improving Codex loop

What shipped for Crete's 30+ firms

Over six months, OpenAI forward-deployed engineers and Thrive Holdings built Tax AI for Crete's network of 30-plus accounting firms, and ran it across 7,000 1040 and 1041 returns during this pilot season. The starting problem was concrete: data entry alone can take eight hours on a medium-to-large-complexity return, working through messy prior-year documents, spreadsheets, and handwritten notes.

The headline results are efficiency numbers — about a third of prep time saved, up to 97% accuracy, roughly 50% more throughput. But the claim the authors care about is that the system got better while running. At launch, only a quarter of returns reached 75% correct field completion; within six weeks, 86% did, with faster gains at the 90% and 100% thresholds as the agent moved from W-2s and 1099s into K-1s and rental schedules.

The correction screen is the data pipeline

The frontend design decision that carries the whole system is treating practitioner review not as a terminal step but as a structured event. When an accountant approves or edits an extracted value, the product records exactly what Tax AI proposed, what the practitioner modified, and what ultimately went into the filed return. That triple is the raw material for everything downstream.

This is easy to underrate. Most review UIs capture the final answer and throw away the disagreement. Here the disagreement is the product. The team explicitly designed the workflow so that 'the people doing the work steer what the product learns' — practitioners generate the improvement signal simply by doing their jobs, without a separate labeling task or feedback form bolted onto the side.

Why provenance had to be a first-class surface

A raw correction is ambiguous. A changed value before filing might be a true extraction miss, a mapping problem, a value carried forward from a prior-year return, a practitioner preference, or just workflow noise. If the interface only stored input and output, engineers would still have to reconstruct which of those it was — the manual, slow loop the post opens by describing.

Tax AI avoids that by preserving the full path: documents organized and classified, rental-property fields extracted with citations back to the source material, values mapped into the tax engine, then the practitioner edit. That provenance trail is what lets the system produce field-level review rows, group recurring failures — repeatedly missing 'fair rental days,' mishandling 'other expenses,' confusing multiple properties in one package — and separate genuine product bugs from expected noise before anything reaches Codex.

Codex then works inside a bounded environment that mirrors this separation: a writable worktree holding the rental-income product surface and its targeted plus regression evals, and a read-only context providing the production trace, source documents, prediction, and filed return. The agent can inspect the evidence without mutating it, propose a pull request, and route ambiguous cases back to engineers rather than forcing them through the loop.

The implication: build the review surface before the agent

The reusable lesson here is ordering. The rental-property capability took about six weeks and heavy engineering oversight to reach 90% precision and recall, but the payoff was reusable abstractions, eval conventions, and review artifacts that made Schedule C and Schedule A cheaper. None of that was possible until the correction surface emitted structured, cited, groupable evidence.

For teams building domain agents where expert judgment defines quality, the takeaway is that the frontend where humans correct the machine is not cosmetic — it is the ingestion layer for self-improvement. Get it wrong and every fix stays manual; get it right and each shipped change generates the evidence for the next cycle. Thrive's own summary of the principle is worth sitting with:

The best agents are steered by people to learn to become more capable, more trusted, and more valuable over time.Montana Labs

The human proof point is a senior accountant who spent 180 hours on tax prep last year and 15 this year — time she redirected toward calling every client and taking on new work. That outcome only exists because the interface she used to correct the agent was, quietly, the thing teaching it.