OpenAI's tax-agent case study is not interesting because a model can help with taxes. It is interesting because the correction becomes infrastructure. In OpenAI's account, practitioner feedback can be captured as traces and evaluation material for Codex, turning professional review into product fuel. [1]

Thursday's paper argued that Codex was being sold as an enterprise correction loop. Friday's narrower receipt shows what that loop looks like when the profession is tax, a domain where small errors produce penalties, liability, and embarrassment.

The company describes a system in which tax practitioners review agent work, correct mistakes, and feed those corrections back into evaluation and improvement. [1] That is not merely quality assurance. It is a claim about how specialized labor becomes legible to a software platform.

OpenAI's enterprise governance material supplies the other half of the story. Codex is not being presented only as a clever assistant. It is being wrapped in controls around permissions, policy, and managed use. [2] The tax example gives those controls a mundane setting: the accountant's desk, not the demo stage.

The divergence is predictable. Mainstream attention will see AI moving into tax preparation or professional services. X will split between automation panic and developer enthusiasm. The paper's question is ownership. Who owns the correction data? Who can audit the trace? Does a firm know whether its expert's fix became a local quality record, a customer-controlled evaluation, or material that improves a vendor's product?

Tax is a useful stress test because the work is both rule-bound and judgment-heavy. A form can be checked. A classification can be disputed. A deduction can be legal in one fact pattern and reckless in another. That means the most valuable correction is often not a red mark on arithmetic. It is the expert's explanation of why the model's plausible answer failed the professional standard.

If Codex captures that explanation, the platform learns from more than mistakes. It learns from the profession's boundary lines. Those lines are hard to write down in advance and expensive to teach one file at a time. The case study makes them operational.

Those questions matter because expert correction is expensive. A partner, associate, enrolled agent, or reviewer is not merely annotating a model's homework. They are translating professional judgment into machine-readable form. Once that judgment is captured, the platform can price, reuse, measure, or restrict it.

There is a benign version of this story. Tax software gets better because experts catch errors. Firms gain an audit trail. Junior staff learn from visible corrections. A model that repeats a mistake can be tested against a new evaluation and fixed. [1]

There is also a harder version. The most valuable thing in the workflow may not be the model's first answer but the corpus of corrected professional mistakes. In that version, the platform company is not just selling assistance. It is instrumenting expertise.

That is why the tax example belongs in the AI state-power thread. Power is not only in model size, funding rounds, or leaderboard claims. It is in defaults, traces, evaluation files, permission settings, and the quiet capture of how professionals say no.

-- DAVID CHEN, Beijing