My Tax Software Had Bugs. My AI Agent Found Them.
Six errors. One would’ve stopped me from e-filing. My AI agent caught all six.
I prepare my own tax returns every year, including an S-corp and a partnership LLC for my media business.
This year, instead of working through the TaxAct interview myself like I always do, I handed the whole thing to Claude Cowork to see what it could do.
I shared the results on Episode 504 of The Accounting Podcast.
It started the way any good preparer would, pulling the GL from both entities, one in Xero and one in QuickBooks, and reconciling the trial balance and P&L against last year's return.
Along the way, it caught a QuickBooks bug I wasn’t aware of: the software was running an accrual report when it thought it was running cash, thanks to a broken dropdown.
Claude figured that out, reran the correct numbers, and kept going.
Then it built the book-to-tax workpaper, asked me the same adjustment questions any preparer would, and opened TaxAct to start the interview.
This is where it got interesting
It ran TaxAct's built-in error checks, as anyone would. Then it kept going and started catching mistakes in the software's calculations.
It found:
Understated non-deductible expenses. TaxAct's autofill ignored the fact that entertainment is 100% disallowed under Section 274(a). There wasn't even a line on Schedule M-1 for that addback, so the agent built a custom one.
Distributions defaulted to zero dollars when the real number was tens of thousands. Left alone, they wouldn’t have shown up on the K-1.
Ending retained earnings auto-computed to the wrong figure. That one's serious because Schedule L wouldn't have balanced.
A name-change box was checked incorrectly on the federal form. The agent caught it because the same field was unchecked on the state return, so it fixed the federal one.
An Arizona e-filing checkbox resets to unchecked every time you run the software. Left unchecked, the state return wouldn't have transmitted.
Claude fixed it all, tied it out, and verified it against a fresh independent read of the GL before I transmitted the returns myself.
Each return took the agent roughly an hour, spread across a few hours of me checking in.
The concerning part for tax professionals
We've built an entire industry on the assumption that tax software gets the mechanics right and preparers supply the judgment calls.
That was never really true.
Preparers have been catching software bugs for years. It's just been invisible, folded into "reviewing the return" like it's part of the job description instead of a sign the tools aren’t truly trustworthy.
Something finally checked the software’s work with the rigor a senior preparer would, at every step, without getting tired and without phoning in the review because it's 9 p.m. during crunch week.
The value of AI is that it never skips a checklist item, no matter how many times it's run that checklist before.
That's what tax software has been missing all along: something that never gets complacent.
Firms treating AI as a way to speed up data entry miss the bigger opportunity
Sure, we can prepare returns faster, but the real value is double-checking returns without cutting corners when the software itself has a baked-in bug.
If your only quality control is a human skimming a printed return before it goes out the door, you're relying on the same weakness that made six separate mistakes in my tax software.
If you're in tax prep and you haven't handed a full return to an AI agent yet, not just for drafting but for working it end to end and checking its own output against the source data, you're missing something bigger than a productivity gain.
