Not Whether You Checked It. When.

I spend a fair amount of time these days reading about AI in law and medicine, not accounting. Partly out of professional curiosity. Mostly because those two professions are a year or two ahead of us on the same road, and I would rather learn from what’s already happened to them than find it out the hard way in my own practice.

What I found wasn’t reassuring in the way I expected. Every AI vendor in every one of these fields says the same sentence: a licensed professional reviews the output before it goes anywhere near a client. That claim is now completely universal — and completely useless as a way to tell a safe product from a dangerous one, because everybody says it. The real question, the one almost nobody is asking yet, is simpler and sharper: reviewed when? Before the professional formed their own opinion, or after?

What law is already living through

Legal AI is having a moment that looks, from the outside, like pure success. Harvey was valued at $11 billion in March, up from $5 billion nine months earlier. Legora was in talks for $10 billion-plus as of this writing, having roughly doubled its own valuation twice in four months. Thomson Reuters’ CoCounsel passed one million users. Every one of these products puts “an attorney reviews this” in its marketing.

And courts are now tracking, by the thousand, what happens when that review doesn’t actually catch the mistake. Roughly 1,490 court decisions worldwide now involve reliance on AI-hallucinated material. An Oregon federal judge handed down a $110,000 sanction for 23 fabricated citations. Ninety-six lawyers have been sanctioned for AI misuse since the start of 2025, against two in the two years before that. One Nebraska attorney was suspended after 57 of 63 citations in his brief turned out to be defective — and the part that stood out to me wasn’t the AI use, it was that he’d denied using it. Courts keep punishing the cover-up harder than the original mistake.

None of those 96 lawyers skipped review. That’s the point. They had a policy that said “an attorney checks this,” the same policy every legal AI vendor tells its clients to adopt, and it didn’t stop the hallucinated citation from reaching a judge. A policy requiring review is not the same thing as review that actually works.

What medicine found when it actually checked

Medicine has the most useful data point in this whole comparison, and it didn’t come from a vendor. Ontario’s Auditor General tested all 20 AI medical scribes pre-approved for provincial use, running each one through two simulated doctor-patient conversations. Every single vendor produced errors. Sixty percent recorded a different drug than was actually prescribed. Seventeen of twenty missed key mental-health details. Nine fabricated treatment recommendations nobody in the room ever said. And in the province’s own procurement scoring, accuracy counted for four percent of the vendor-selection weight — domestic presence counted for thirty.

That’s the closest thing I’ve seen anywhere to an independent, government-run accuracy audit of a professional AI category. Every accuracy figure any vendor publishes about itself, in law, medicine, or accounting, is self-reported. This is the one exception, and it found that the review step — not the model — is doing all the actual work of catching errors, when it’s done at all.

But the single finding that changed how I think about this came from a 2026 mock-juror study on radiology. Researchers built two versions of the same scenario: a radiologist misses a bleed that the AI correctly flagged. In one version, the doctor sees the AI’s flag first, then forms a single read. In the other, the doctor reads the scan blind, forms an independent opinion,then sees what the AI thought. Same doctor. Same missed bleed. Same underlying mistake.

Seeing the AI’s answer before you form your own produced a plaintiff-favorable verdict 74.7% of the time. Forming your own opinion first, then checking it against the AI, dropped that to 52.9%. Same error. A 22-point swing in how a jury assigns fault, based purely on the order two people looked at the same piece of paper.

That is not a finding about whether review happened. It is a finding about sequence. And it is the piece I have not seen a single accounting-AI product, or accounting firm, treat as a designed decision rather than an afterthought.

Wealth management is converging on the same rule — with real teeth behind it

FINRA’s 2026 oversight report moved “from observation to enforcement-ready expectations” on AI supervision. The CFP Board stood up an AI working group. And the SEC has already fined two firms — Delphia and Global Predictions — for “AI washing,” claiming AI capabilities they didn’t actually have. Six such cases are now tracked, with more than $44 million in alleged fraud. If you build or sell anything you call AI-powered, that enforcement category applies to you whether you’re registered with the SEC or not. Say what the tool actually does. Nothing more.

I’ll also say plainly: the tools purpose-built for exactly the kind of Roth-conversion and retirement-distribution planning I do for clients already exist and are already good. Income Lab’s Tax Lab runs twenty parallel tax-aware withdrawal strategies with real IRMAA and bracket modeling. Holistiplan reads a 1040 in under a minute and surfaces the Roth-conversion window itself. Neither one waits for accounting to catch up — they’re already doing it, for the advisors who got there first.

Where accounting actually stands

Compared to the other three, accounting looks like the least regulated at the licensing-board level, for better or worse. The AICPA has published a small-firm generative-AI policy template and is clear that AI changes none of a CPA’s professional responsibilities — due diligence and professional judgment are still the standard, full stop. But I could not find a single state board of accountancy or NASBA rule addressing AI specifically, the way California’s bar is moving toward a binding duty to verify every AI output, or the way Alabama passed a law requiring a licensed human, not an AI recommendation alone, to make certain healthcare decisions.

That gap cuts two ways. Here’s the honest comparison across all three fields I looked at:

I don’t read that table as “accounting is safe.” I read it as “accounting hasn’t had its public reckoning yet,” which is a very different thing from not having one coming. The absence of a tracked incident record isn’t evidence nothing has gone wrong — tax returns and workpapers aren’t public record the way a court filing is. It’s entirely possible we simply can’t see it yet.

The rule I’m actually adopting

“A human reviews the AI’s work” is table stakes now — every product in every one of these professions claims it, which means it tells a client precisely nothing about whether the work is actually safe. The distinguishing question, the one the radiology study answers with real numbers, is whether the professional forms an independent judgment first, and only then checks it against what the machine produced — or whether the machine’s answer is the first thing anyone sees, quietly anchoring every “review” that follows.

Those are not the same discipline wearing two names. One of them catches the error a jury believes you missed on your own. The other one just has your signature on it.