How reliable is AI at doing math or fully preparing something like a K-1 entry without human review?

The consensus was that AI tools tend to be inconsistent with math because they aren't actually performing calculations the way traditional software does, which may explain some of the inconsistency. The recommended approach is to treat it as a combination of 'its brain and my brain' — for example, uploading a K-1 and having the AI explain what it sees and how it should be entered in your software, then verifying that it ties out and makes sense given your understanding of the entity. This differs from a taxpayer doing the same thing, because a professional knows when to push back if something doesn't look right — whereas the AI will often agree with you even when you're wrong ('Oh yeah, yeah, you're right') if you push back, regardless of whether your pushback is correct. This means firms need to train their team on what percentage of a task should be the AI's output versus a human's judgment, and reviewers need to understand that balance too.