An agent can answer a tax question correctly and still miss the issue that matters. Surge AI’s October 27 benchmark gives agents 50 professional assignments with 1,789 supporting files: incomplete requests, conflicting records and deliverables to finish.

The best reported pass rate is 10.8%, with a pass requiring at least 95% of a task’s rubric. The useful test is judgment as well as calculation: investigate assumptions before turning them into confident advice.

Two sample tasks and the evaluation harness are public; the full set requires access. These are vendor-reported results using an AI grader, not an independent audit or a measure of every specialist tax product.