What the work involves
You build and score accounting tasks for frontier model training. In practice that means three recurring jobs. First, authoring problems with a defensible single answer — a lease modification under ASC 842, a step acquisition, a deferred tax remeasurement after a rate change — together with a worked solution and the journal entries that support it. Second, reviewing model output side by side and deciding which response is better and why, at the level of "the model applied the incremental borrowing rate at the wrong date" rather than "answer B is clearer." Third, correcting model reasoning that arrives at the right number through wrong logic, which is the failure mode graders are paid most to catch.
Audit-side work runs parallel: sampling design, materiality and performance materiality setting, evaluating whether a proposed procedure actually addresses the assertion at risk, and judging model advice on going concern, subsequent events, or opinion modification. Tasks are usually framed under US GAAP/PCAOB or IFRS/ISA, and you will be asked which framework you can work in without looking things up.
What the screen looks for
- A verifiable licence or equivalent — CPA, CA, ACCA, CIMA, or Big Four/national-firm audit experience with the years to back it.
- Depth that survives follow-up. The AI interviewer will push on a technical answer two or three times; surface familiarity with a standard shows immediately.
- Rubric discipline: can you state the criteria you graded against before you state your verdict, and stay consistent across twenty similar items?
- Willingness to mark a plausible, well-written, wrong answer as wrong.
Logistics
Fully remote and asynchronous, with work claimed from a queue. Most contributors run 10–20 hours a week; some projects offer bursts of full-time volume for a few weeks. Expect a paid or unpaid calibration set before live work, and periodic requalification when a project changes rubric. Pay is hourly at the stated observed band and varies by specialism — technical accounting and audit methodology sit higher than bookkeeping-level tasks. Nothing here is guaranteed volume.