October 15 is behind us. The extensions are in, the stragglers are filed, and for the first time since January the office is quiet enough to think. I like to use this week for one uncomfortable question: of everything that went wrong this season, how much of it started with three words?

"It looked right."

I've heard those words in every firm I've worked in. I've said them myself. They come out after a return goes in with a wage figure off by two swapped digits, or a 1099 from the wrong year, or interest pulled from the statement before the corrected one arrived. Nobody skipped the work. The number was sitting in the right box, formatted properly, and it looked exactly like a correct number would look. So it went through.

My argument is simple. "It looked right" is the most expensive phrase in a tax practice, and it's about to get more expensive, because the tools we're all adopting are very good at producing numbers that look right.

The short version

  • AI tools that read tax documents return wrong values with the same confidence as right ones, so errors no longer look like errors.
  • Tax review was built to catch things that look off. A confident wrong number gives it nothing to catch.
  • The fix is in how the work is laid out: the source document next to the number, an honest blank instead of a guess, a screen that doesn't claim checks it never ran, and a person who approves last.

Why did errors used to be easier to catch?

For most of this profession's history, mistakes gave themselves away. Handwriting was hard to read, so you slowed down. A total didn't foot, so you went back. A keyed number was off by a factor of ten and it jumped out on review because it was absurd. Manual work was miserable, but it had one hidden virtue. When something went wrong, something usually looked wrong.

Good reviewers built their instincts around that. Nobody in March reads every line with equal suspicion. You scan for the odd one: the figure that's too round, the box that's empty when it shouldn't be, the number that doesn't match last year for no reason. In practice, review is pattern-matching for wrongness.

That works well, as long as wrong things look wrong.

What changes when AI reads the documents?

More and more of the numbers reaching a reviewer's desk didn't come from a tired person keying off a page. They came from a tool that takes a document in and hands a value back. And those tools have a property the old failure mode never had. They are never unsure.

Ask one for a number and you get a number. Clean, formatted, in the right field, delivered with the same confidence whether the box was crisp, smudged, cut off at the edge of the scan, or empty. A correct read and a bad guess look identical on the screen.

So the risk has moved. It used to be a typo you could spot. Now it's a clean, assured answer nobody thought to check, because nothing about it asked to be checked. A confident wrong number doesn't look wrong. It looks finished.

And review, tuned for a world where wrong things look wrong, walks right past it.

Why does this hit tax harder than other industries?

Lots of industries are dealing with tools that are confidently wrong. In most of them, a plausible-looking error costs an awkward correction. In ours, the error goes onto a return, and the return goes out with a signature under it.

That signature carries a specific duty. Circular 230 lets a practitioner rely in good faith on information a client provides. It also says we can't ignore the implications of what's in front of us, and we have to make reasonable inquiries when something looks incorrect, inconsistent, or incomplete. The whole standard assumes a diligent professional will notice when something is off.

Now picture the case where nothing looks off, because the error was smoothed over before it ever reached you. The duty hasn't moved. Whoever signs still owns the number. But the cues we've always used to meet that duty are gone.

The bill doesn't come in April, either. It comes months later as a notice the client has forgotten about, an amended return, an awkward phone call, and a small dent in the one thing a firm really sells: that its numbers can be trusted. Each of those starts with a reviewer who looked carefully and saw nothing wrong.

Should firms stop using AI for data entry?

No. I use these tools, and I'd be foolish not to. The volume this profession carries with the same or fewer people isn't survivable without help. Anyone who says otherwise hasn't worked a February with two people out sick.

My point is narrower. Speed is easy to buy now. Most vendors sell it and most deliver it. What's harder to find, and what's actually worth paying for, is the moment where a person looks at a confident answer and can tell whether it's also the correct one.

You can't bolt that on at the end. It depends on how the work is laid out in front of the person reviewing it. A reviewer can only verify what they can see. If what they see is a bare number in a field, cut off from the page it came from, then "verification" means trusting the field. That isn't checking. It's hoping, with a tick mark.

How do you review AI-extracted tax data? Four habits

Take the confident-wrong problem seriously and four habits follow. None of them depend on which software you use.

1. Put the source next to the number. The most useful thing you can do for a reviewer is put the original document beside the value, so checking is a glance across the screen, not a hunt through a folder. When the evidence is right there, people look at it. When it's three clicks away, at 9 p.m. in March, they don't.

2. Prefer an honest blank to a confident guess. When a box can't be read cleanly, the right output isn't a plausible number. It's an empty field with a short note saying why. A blank draws the eye and says a person needs to decide this one. A guess hides in plain sight. Any process that would rather fill a field than admit it doesn't know is producing exactly the errors your review can't see.

3. Don't let the screen claim checks that never ran. Green ticks, "verified" badges, and confidence scores that don't measure anything real are worse than nothing. They tell a tired reviewer to stop looking. An honest screen is quieter than a confident one, and it's the only kind worth trusting at the end of a long day.

4. Keep a person as the last check, every time. Approval should be someone's deliberate act. Not a default, not a timeout, not a status the software gives itself. Nothing moves forward until the person who will answer for it has looked and said yes.

A 30-minute exercise for the week after the deadline

Here's what I'd do this week while it's quiet. Pull three returns from this season where something went wrong after filing: a notice, an amendment, a client correction. Trace each one back to the moment the bad number entered the file. Then ask one question. At that moment, did the number look wrong?

My guess is that most of the time it didn't. It looked right. It looked finished. And your review did what it was built to do. It looked for something that seemed off and found nothing.

That's a design problem more than a people problem, and every firm can start fixing it before the next season lands on the desk.

Speed you can buy. The check is the job.

I'd like to hear from other practitioners: what's the error that looked right and got past you this year?

Frequently asked questions

Can a tax practitioner rely on numbers from AI or OCR software?
Circular 230 allows good-faith reliance on information a client furnishes, but a practitioner can't ignore the implications of what they know or see, and must make reasonable inquiries when information appears incorrect, inconsistent, or incomplete. The person who signs remains responsible for the return.

Why are AI extraction errors hard to catch in review?
Because a tool returns a wrong value in the same clean format and with the same confidence as a right one. Traditional review looks for things that seem off; a confident wrong number gives it nothing to react to.

What is the best way to review AI-extracted tax data?
Show the source document beside each value being checked, leave unreadable fields blank with a reason rather than guessing, remove on-screen markers that don't reflect a real check, and require a person's explicit approval before anything moves forward.

Yatin Miglani is an Enrolled Agent in Phoenix, Arizona, and the founder of Sophicor. He writes The Verified Practice on the AI-tax era from the practitioner's chair — the problems he lived before he built for them.

Sources referenced: Treasury Department Circular No. 230 (31 CFR Part 10), §10.34(d), relying on information furnished by clients: a practitioner may rely in good faith without verification upon information furnished by the client, but may not ignore the implications of information furnished to, or actually known by, the practitioner, and must make reasonable inquiries if the information as furnished appears to be incorrect, inconsistent with an important fact or another factual assumption, or incomplete; §10.22(a), a practitioner must exercise due diligence in preparing, approving, and filing tax returns. 31 CFR Part 10, eCFR / law.cornell.edu.

— Yatin Miglani

Enrolled Agent · Phoenix, Arizona
Founder, Sophicor · sophicor.com