client filed with AI-generated numbers — now facing an accuracy penalty. real case.
need to vent and also warn everyone.
client (self-employed consultant, ~$200k gross) used ChatGPT to "do his taxes" in March. didn't tell me. filed the return himself through FreeTaxUSA using the numbers ChatGPT gave him.
what ChatGPT got wrong:
- treated his home office as a depreciation deduction instead of the simplified method he'd been using (you can't switch mid-stream without recapture consequences)
- applied a 20% QBI deduction without checking that his income was over the threshold AND he was in an SSTB
- missed his estimated tax payments entirely — just didn't account for them
- overstated his health insurance deduction by including his wife's premiums (she has employer coverage)
net result: understated tax liability by about $14,000. IRS sent a CP2000. now we're dealing with:
- amended return
- accuracy-related penalty (§6662) — 20% of the underpayment
- interest from original due date
- potential negligence finding
total exposure is roughly $18,000 including penalties and interest. the client is furious — at ChatGPT, at himself, and unfortunately at me for "not warning him."
i'm now adding a paragraph to every engagement letter about AI-prepared returns.
4 replies
this is the case study everyone in this community needs to read.
$14k underpayment → $18k total exposure after penalties and interest. that's a 28% premium for using AI without professional oversight. the $2,000 he "saved" by not hiring an accountant cost him $18,000.
the engagement letter addition is smart. ours now says: "Client acknowledges that tax positions generated by artificial intelligence tools may contain errors. If Client files a return based on AI-generated calculations without prior review by Firm, Client assumes full responsibility for any resulting penalties, interest, or professional consequences."
harsh? yes. necessary? absolutely.
the QBI/SSTB error is particularly nasty because it SOUNDS right. "you're self-employed, you get a 20% deduction." that's true for many people. but if you're a specified service trade or business above the income threshold, it phases out to ZERO.
this is why AI is dangerous for tax: the rules have exceptions, and the exceptions have exceptions. AI gives you the general rule and misses the specific limitation.
maria — for the §6662 penalty, consider making a reasonable cause argument. the client relied on a "sophisticated technology tool" in good faith. it's not a slam dunk but i've seen it work for first-time offenders. cite Treas. Reg. §1.6664-4(b).
can we talk about the depreciation/simplified method switch? this is a trap that even HUMAN preparers fall into.
§280A(c)(5) simplified method: $5/sqft, max 300 sqft. once you've been using it, switching to actual method requires you to calculate the depreciation you WOULD have taken. and if you're going back... there's potential recapture.
ChatGPT doesn't know which method the client used last year. it doesn't have prior return data. so it just picks one. that's not "wrong" in isolation — it's wrong in CONTEXT.
this is the fundamental limitation: AI doesn't have continuity across tax years. every return is year-one to it.
the "client is furious at me for not warning him" part hits hard.
we're going to see more of this. clients will use AI, mess up their taxes, and then blame their accountant for not preventing it. even when the accountant wasn't involved.
this is actually a huge opportunity for this community. if we can build enough awareness that "AI tax prep needs professional verification," we create demand for exactly what we do. maria's client is never going to file an AI return without getting it checked again.
Sign in as a verified accountant to reply.
Sign in