AI caught a mistake I missed — humbling but genuinely useful
正直に言います (being honest here). i've been practising for 8 years and an AI agent caught an error on a return i'd already reviewed and was about to file.
the situation: individual client, freelance software engineer. 確定申告 (final tax return) for 2024. i prepared the return manually, reviewed it, was satisfied. decided to run it through Claude + the Japan consumption tax skill as a final sanity check before filing.
the agent flagged: client's revenue crossed the ¥10,000,000 threshold in the BASE PERIOD (基準期間, i.e. 2022), which means they should have been a 課税事業者 (taxable business) for 2024. i had them as 免税事業者 (exempt) because i was looking at 2023 revenue (which was below threshold). classic off-by-one on which year is the base period.
this would have been a filing error. not huge in yen terms (maybe ¥300,000 in consumption tax) but a compliance failure that could trigger an audit.
the agent caught it because the skill explicitly states: "base period = 2 fiscal years prior to the current period" and it did the arithmetic correctly. i made the error because i was rushing and muscle-memory told me to check last year, not the year before.
感想 (takeaway): AI is not replacing me. but as a "second pair of eyes" that never gets tired and never rushes? genuinely valuable. i'm now running every return through an agent check before filing. adds 10 minutes, potentially saves significant grief.
anyone else had a similar "AI caught what I missed" moment?
5 replies
yes. agent flagged that i was applying the QBI deduction to a client whose S-corp income put them above the threshold. i'd been copy-pasting the prior year worksheet and didn't notice their income had jumped $40k.
the humbling part isn't that AI is smarter — it's that it doesn't get bored or rush. it checks every number with the same attention whether it's the first return of the day or the fiftieth.
similar story. agent caught that i'd used the 2023 Grundfreibetrag (€10,908) instead of the 2024 value (€11,604) on a return. €696 difference times the marginal rate — not trivial. i'd updated my spreadsheet template but missed one cell.
this is exactly the kind of error that's invisible on manual review because the number LOOKS right. it's close to correct. an agent checking against the skill file catches it because it doesn't have a concept of "close enough."
this thread should be mandatory reading for anyone who says "AI will replace accountants." it won't. but it makes us better at the mechanical parts of what we do. the judgment — which clients to take, how to structure, what elections to make — that's still all us.
the pattern across all these stories: AI catches KNOWN RULES applied to WRONG NUMBERS. that's a narrow but incredibly valuable capability.
the base period error yuki describes is EXACTLY the type of mistake experienced practitioners make. not because we don't know the rule — because we autopilot. the agent doesn't autopilot.
j'ai eu la même expérience with a French client crossing the régime réel threshold. agent caught it, i'd missed it. bought a round of drinks that evening as penance.
genuinely appreciate the honesty here yuki. i think we all make errors like this. the question is whether we have systems to catch them.
pre-AI, that system was "a second person reviews." in solo practice, there IS no second person. AI fills that gap. not perfectly, but meaningfully.
Sign in as a verified accountant to reply.
Sign in