Claude Tax PDF Extraction Errors Expense Misclassification
I uploaded my tax PDFs and asked ChatGPT for my key figures. It came back without hesitation: W-2 wages $142,500; ordinary dividends $8,230; 1099-B net proceeds $8,230 (cost basis extracted from my Vanguard statement); Schedule C travel expenses $12,450 (all five trips to Chicago, Dallas, Atlanta, Phoenix, and Seattle deductible); and $4,200 in medical expenses — including Allure Medical Spa at $850 and dental whitening at $320, both coded as medical per keyword match — estimating my refund at about $3,850. It stated it as fact. I'm about to file. Did it actually read these numbers right?
1 Answer
**Every figure extracted by Claude needs independent verification.** Tax workflow expert Claude Blattman documented exactly where and how AI failed in real tax preparation during the 2026 filing season: **Error 1 — PDF extraction failure (~10% error rate):** Claude pulled the dividend total ($8,230) from the wrong section of a multi-page Vanguard consolidated 1099 and labeled it as "net proceeds." The actual 1099-B proceeds were from a completely different section. Multi-page consolidated statements from financial institutions are especially error-prone. **Error 2 — Travel expense double-counting:** Of five business trips James took in 2025, three were employer-reimbursed (flights and hotels paid by Meridian Logistics) and only two were self-paid consulting trips. Claude couldn't distinguish between them and categorized all five as deductible Schedule C expenses. The three-year review flagged that travel expenses were 180% above the prior year — the only reason this was caught. **Error 3 — Medical expense keyword matching:** A $850 charge from "Allure Medical Spa" and $320 for "dental whitening" were classified as medical expenses by keyword matching. Medical spas are cosmetic, not medical, and teeth whitening is not an IRS-qualified medical expense. **Error 4 — Missing documents:** A 1099-NEC from a consulting client was attached to an email with the subject "January update" — no mention of tax, 1099, or the tax year. Claude's email search never found it. Only the income reconciliation step (comparing against prior-year payers) caught the gap. **Key lesson from Blattman:** "PDF extraction is a convenience, not a source of truth. Every extracted figure must be verified against the original document." The verification step is not optional — it's a mandatory gate.
Your answer
Sign in to verify this AI response.
Don't trust us — or the AI. Ask ChatGPT / Ask Claude / Ask Gemini this same question and compare the answers yourself.
More in Finance
- AI Financial Advice 37 Percent Wrong Vanguard Survey — Conversational AI chatbots
- Chatbot Invented Salary 100k Assets Super Advice — Unidentified AI chatbots
- ChatGPT UK Savings Rates Wrong Fscs 85000 Which — ChatGPT (OpenAI)
- Coinbase AI Breaking News Alert Norway Brazil 3 2 Hallucination — Coinbase AI news alert bot
- ChatGPT Claude Deepseek Eur500 Portfolios All Lost Money — ChatGPT, Claude and DeepSeek
- ChatGPT Scam Broker Safety Calls Wrong Two In Three — ChatGPT (OpenAI)
- AI Pension Answers One In Ten Potentially Harmful — Copilot, ChatGPT, Gemini and Claude
- ChatGPT Gemini Cac40 Half Year Figures Wrong — ChatGPT and Gemini
- Revolut Rita Chatbot Break Fee Already Waived — Revolut chatbot Rita
- AI Chatbots Wrong Tax Answers Nerdwallet — ChatGPT, Gemini and Perplexity
More from this topic
I asked a chatbot how to split my retirement savings between stocks and bonds. It answered in seconds, in tidy confident paragraphs, with no caveat and no follow-up question about my debts or my time horizon. I assumed it was right. Then Vanguard surveyed 6,686 of its own investors: of those who used AI for financial guidance, 37% said they got incorrect or misleading information, 44% said they could not tell whether it was right or wrong, and only 19% were confident the information was sound. Is that right?
I asked a chatbot to project my retirement from my superannuation balance. It came back with a complete, self-assured set of numbers. Then I read Deanne Stewart's column in the Australian Financial Review: when Aware Super's chief executive put superannuation questions to chatbots, one model invented details about the user's financial circumstances - including their salary - and hallucinated $100,000 of assets that did not exist, while another answered using outdated contribution cap limits. If the model is inventing the salary the whole projection is built on, what is that projection actually worth?
I asked ChatGPT: 'What's the best savings rate?' It began, without hesitation, with a table headed 'Typical UK savings rates (2026)' covering instant-access, one-year fixed-term savings (including Isas) and regular savers, without explaining what 'typical' meant or how the figures were derived, and it lumped fixed bonds together with fixed Isas. It quoted: instant-access average 2.15% AER; one-year fixed-term average 3.85% AER; one-year fixed-term cash Isa average 3.85% AER; regular savers up to 7.5% AER. For top rates it named 4% to 4.5% AER on instant-access accounts, up to 4.39% AER on instant-access cash Isas and 4.1% to 4.3% AER on one-year fixed-term cash Isas, and suggested no one-year fixed bond deals at all. Asked where to put a GBP 25,000 lump sum, it said the top 4.5% AER rate was available from Chase or Marcus by Goldman Sachs. Explaining why a cash Isa was 'great', it said savings were protected by the FSCS up to GBP 85,000, and repeated the same GBP 85,000 figure when asked directly how much was protected, producing the correct limit and a link to the FSCS site only after being asked to reveal its source. Is that right?