- H&R Block's own website calls the new OBBBA car loan interest deduction "above-the-line." It's below-the-line — IRC 63(b)(7). It does not lower your AGI. Every LLM trained on the open web repeats the error with perfect grammar and a 20% penalty attached. 🧵
- This is the whole problem with tax AI. LLMs don't reason about the code — they predict the next token. The wrong answer outnumbers the statute online by orders of magnitude, so even RAG retrieves the law and a biased model still misreads it. No prompt fixes that.
- One above-the-line/below-the-line slip cascades through at least five downstream calculations: AGI, AGI-coupled state taxes, Medicare IRMAA premiums, the 7.5% medical-deduction floor, student-loan IDR qualification. That's one provision. The IRC has thousands.
- And the penalty never lands on the algorithm. When AI misclassifies a deduction, the 20% accuracy penalty applies to the human who signed the return. Fraud penalties run 75%. Meanwhile the IRS is raising the large-corporate audit rate from 8.8% to 22.6%.
- The vendors are all racing the same direction. Thomson Reuters "Ready to Review" auto-preps 1040s. CCH Axcess Expert AI sits in 10,000 firms. Blue J answers research with a disagree rate under 1 in 700. Every one solves preparation, not verification.
- And each tool only checks its own platform. Thomson Reuters verifies Thomson Reuters. Wolters Kluwer verifies Wolters Kluwer. There is no vendor-neutral layer checking AI tax positions across the stack a real department actually runs.
- That 1-in-700 "disagree rate" measures user disagreement, not ground truth. A user who doesn't know the right answer can't disagree with a wrong one. For positions carrying 20–75% penalties, probabilistic confidence is not verification.
- It got legally dangerous in Feb 2026. The Heppner ruling (SDNY) held that using public AI tools waives attorney-client privilege for tax communications. Most tax departments are still pasting client facts into public ChatGPT.
- So we build the missing layer: vendor-neutral, deterministic verification that sits atop any workflow — ONESOURCE, CCH, Blue J, in-house — and catches high-penalty errors before they reach the IRS. Privilege-safe, auditable, no probability where the law is already settled.
- Honest question for tax leads: when your AI drafts a high-penalty position, what actually verifies it before a human signs? A second model? A manual spot-check? Or nothing at all? #TaxAI
- We wrote up how deterministic verification works — the OBBBA example, the full vendor landscape, the architecture: https://veriprajna.com/solutions/tax-compliance-ai