Most AI Review Processes Don't Catch Anything Real

If your review step for AI-assisted work is "read it, and approve it if it seems fine," you don't have a review process. You have a formality with a sign-off attached. Three stories this year, none of them framed as being about evaluation, land on the same three decisions a working review step actually requires — made in advance, not improvised while you're reading the output.

Three stories, one gap

Stanford's 2026 AI Index found that roughly 40% of real workplace tasks still come back needing significant correction, despite 88% of organizations now using AI. That's not a story about the tools getting worse. It's a story about almost nobody having written down, in advance, what "needs correction" actually means for the specific work they do — so every review defaults to a feeling instead of a standard.

A guide published by Hamel Husain and Shreya Shankar this fall named the same gap from the other direction: most teams pick a metric or build a dashboard before anyone has read a single real output. A score chosen before a failure has been read isn't measuring quality. It's measuring a guess about what quality would look like.

And a NEJM AI study of 376 stuck diagnostic cases showed what happens when the roles in a review are actually assigned. o3 Deep Research generated ranked, evidence-linked hypotheses; physicians decided which were worth testing, ran the tests, and made the call — 18 patients got answers because the model's job stopped at "generate" and the human's job started at "filter," with nobody collapsing the two.

Three different domains, one shared mechanic: a review step only catches something real when someone decided, before looking at the output, what counts as wrong, who is allowed to decide what, and how precisely a problem has to be named to count as fixed.

What to decide before you review anything

What counts as a real failure. Pull 20 to 30 outputs your workflow actually produced — not hypothetical ones — and read them the way a skeptical client would, before you trust or build any score. The failure modes you find by reading are almost never the ones a generic rating would have caught.

Who decides what, at which step. If AI is generating candidates, name the separate step where a human filters them, and the separate step where a human confirms the result. The moment "generate" runs straight into "confirm" with nobody filtering in between, the review isn't happening — it's being skipped at the one point it mattered.

How precise a correction has to be. Accountants reviewing OpenAI's Tax AI only improved the system's next run when their corrections named the specific source of the error — the wrong year's schedule, the wrong line. "This is wrong" fixes nothing twice.

Owen: the step that was missing, not the people

Owen runs a six-person underwriting team at a regional insurance broker. An AI tool drafts a first-pass risk summary for each new commercial policy application, and for a year the process was: an underwriter reads the summary, and if nothing jumps out, signs off. Nobody had ever asked, in those terms, whether "generate" and "confirm" had quietly become the same step with no filter in between. He added one explicit stage: before any underwriter can sign off, they have to name which two data points in the summary they actually traced back to the application file, not just read as plausible. The sign-off rate dropped for the first two weeks. The number of corrected policies after issue dropped faster.

Renata: the correction that stopped repeating

Renata leads a four-person content team at a 140-person HR software company. Her review notes on AI-drafted blog posts used to be quick: "tighten this," "doesn't sound like us." The same looseness kept showing up in the next draft regardless. She switched to naming the exact miss — "paragraph 3 claims a stat with no source; either cite the report or cut the line" — and that specific failure stopped recurring within two drafts. The correction gave the next pass something to fix instead of a mood to guess at.

The one thing

None of these three stories is about AI getting better or worse at the underlying task. They're about whether anyone decided, before the output showed up, what a real failure looks like, who gets to filter it, and how specific a fix has to be. Skip any one of those three decisions and the review step still happens — it just stops catching anything.