I Trusted an AI With My Taxes. It Was Mostly Right, Which Is the Scary Part
Not financial advice, and I still sent it to a human. But watching the model do my return was a clean look at where AI helps and where it quietly guesses.
I am not recommending this, and I paid a person to check it. But I ran my own return through an AI first, because I wanted to see exactly where it helped and where it lied with a straight face.
Where it helped was the grunt work. Pulling the documents into one place, matching the forms to the line items, catching the deduction I'd forgotten because it was buried in a pdf. That part was faster and more thorough than I am, and I'd use it again just for that.
Where it guessed was the dangerous part. On a weird one-off, a grant that had two possible treatments, it picked the cleaner one and presented it like fact. It wasn't fact. It was a coin flip dressed as a recommendation, and the only reason I caught it was that I happened to know the rule.
The lesson isn't 'AI is bad at taxes.' It's that the model is confident precisely when it shouldn't be, on exactly the inputs where you have the least margin for error. The grunt work it does honestly. The judgment call it does loudly.
So my actual workflow now: let it assemble and flag, me decide the weird bits, human sign off. The AI saved me an afternoon. The human saved me an audit. Both earned their place, and only one of them should ever be the last word.