An analyst signing off a valuation
She wrote the summary on page one. The assumption that carried it sat on page sixty-one, and it was cut two revisions ago. Nothing in the document says the summary now stands on nothing.
01
Everything in it can be true. The answer can still be wrong.
AI has made professional analysis faster and cheaper to produce. It has not made it safe to rely on.
A report can be fluent, well sourced and entirely plausible, and still not support the decision made with it.
The caveat sits on page nine. The recommendation sits on page one. Both are true. Read together they say something the summary does not.
In RAND's preliminary benchmark, three GPT-4.1 systems correctly judged how well technical-report claims were supported 48%, 54% and 53% of the time.
A prompt can ask for scrutiny. It cannot supply the controls that make a check dependable.
RAND, Evaluating Large Language Models' Abilities to Process and Understand Technical Policy Reports (April 2026).
02
A campaign budget decision
One part of the pilot shaped the next campaign.
13.4×Presented ROAS
7.2×Actual return when the whole pilot was included
$15.5MThe expectation gap
03
These are usually cheaper to resolve before the decision than after it.
The pilot report above was a numbers problem. Most are not.
A recruiter sending a shortlist
The ranking reads the same all the way down. Two of the twelve are ranked on things the interview notes do not say. Nothing marks where the notes stopped and the inference began.
An underwriter pricing a submission
The broker's report is complete on its own terms. It does not mention what is stored on the third site, because nothing in it had to. Nothing in a document marks what is missing from it.