AI can now file your taxes, but it can’t defend the return

Sarah believed she had completed her taxes. Her software meticulously scanned all her documents, filled out the forms automatically, and explained each deduction as if it were a supportive coworker. It was like having TurboTax’s Intuit Assist or H&R Block’s AI Tax Assist holding her hand through the more complicated parts.

However, hours later, she received a notification from the IRS: Subject: Discrepancy detected – Line 28. There was no waiting on hold, just a system correcting her own. For some context, the IRS has been employing data-driven methods, including AI, to manage complex tax returns. While this is primarily focused on larger, more complicated filings, the same anomaly detection can quickly flag issues even in standard returns.

Her AI and the IRS’s AI were already in an argument she hadn’t even joined.

Where Models Commonly Collide

Across town, Michael, who operates a small candle shop online, noticed a yellow alert flashing on his sales tax dashboard. His accounting AI had filed returns for him in forty-two states, but one of them, North Carolina, had "adjusted" his figures.

The issue was with a specific item: the Vanilla Bean Candle. His system categorized it as “home fragrance,” but the state’s classifier interpreted "vanilla" as "food". This kind of misclassification often occurs in states with specific rules for food and candy; just a small change in labeling can affect the tax rate. While sellers typically rely on AI classifiers like Avalara to minimize these errors, conflicts still arise. Two models, differing by just one token, and now the bill was his to pay.

What Tax Season Feels Like Now

Machines handle the filing, machines conduct the checks, and they often disagree, in real time. What once took months ,the mismatch, the notification, the resolution, now unfolds before your coffee has a chance to cool.Even the government side is becoming more "assistant-like". By 2025, IRS Direct File will be able to import information directly from your IRS account, allowing parts of your return to be pre-filled before you even review it. It's quicker, yes, but it still demands careful judgment.

The Human Step-In

Sarah didn’t panic. Instead, she opened her filing assistant and requested it to draft a clarification. The initial version sounded like a robotic lawyer, too lengthy and overly confident. So, she decided to rewrite it herself, cutting down on the jargon and adding a personal touch with one simple line:

"The course was part of a degree program, and the provider qualifies as an eligible institution according to the definition outlined in IRS Publication 970."(Pub 970 = education credits; you don’t need the whole thing—just the relevant section.)

She uploaded the letter along with her proof of enrollment. Two days later, she received confirmation: adjustment accepted. Her refund was reduced by $400, but the matter was resolved smoothly.

Michael's situation, however, took longer to sort out. The export from his platform, a massive wall of JSON detailing why "vanilla" triggered the incorrect code, was completely unreadable. Frustrated, he copied the key lines and added a brief note:

"This is the classification log indicating that the state misclassified SKU A13 as food. Attached are the order IDs and UPCs for your reference."

A week later, the penalty vanished. There was no apology, just silence, the bureaucratic equivalent of saying, “you were right.”

AI has made taxes faster, but not simpler. Your software can fill out the forms and dispute the numbers, but it can’t convey your story. That responsibility still falls on you: to bridge the gap between what the algorithm perceives and what the rules actually entail. The new essential skill isn’t just data entry; it’s context control.

Because the next time you receive a notice titled “Discrepancy Detected,” you’ll understand what it truly signifies: two machines have started discussing you. And the one that prevails will be the one you trained to communicate like a human.

Disclaimer: Informational only, not tax or legal advice. US-centric examples; state rules vary.

Sources (single line): Intuit Assist (TurboTax); H&R Block “AI Tax Assist”; IRS statements on AI/data-driven selection and enforcement focus; IRS Direct File (2025) account-data import/prefill; North Carolina DOR pages on food vs. candy/soft drinks vs. prepared food and state/local rates; Avalara AI-assisted product/tariff classification; IRS Publication 970 (education credits and eligible institutions).