Dayjob: Tax – evaluating AI agents on 50 professional tax assignments

DAYJOB: Tax is a benchmark for whether AI agents can complete professional tax assignments from an initial client request through a finished return, review, workpaper, or recommendation. Its 50 realistic work environments span eight fields, including individual and business tax, partnerships, estates, international tax, and disputes. Agents must work through 1,789 varied records - such as returns, ledgers, contracts, emails, and handwritten receipts - while resolving conflicts, recognizing missing evidence, and deciding which issues need attention even when clients do not ask about them. Tax experts created and reviewed the assignments.

The results show how rarely current models meet a high professional standard: Claude Sonnet 5.5 led with a 10.8% pass rate, followed by Claude Fable 5.1 at 9.6% and Claude Opus 5.5 at 9.2%. A pass requires meeting at least 95% of a task’s criteria, covering calculations, evidence, consequential issues, and the deliverable. Illustrative failures include overlooking the deadline after which a homeowner could lose a large-gain exclusion, and recommending how to divide retirement accounts without resolving whether a couple was legally married. The benchmark’s central test is therefore broader than answering tax questions or calculating returns: agents must identify the real problem, investigate consequential uncertainties, and carry sound judgment through to usable professional work.