The most dangerous AI answer can be the one that sounds convincing. Clear language and a confident explanation make an answer easy to accept, even when the evidence behind it is missing.
We ran an experiment to see how an AI tool would perform out of the box on cost and project management tasks. We supplied a large set of documents and expected something usable in return. Some answers had no supporting citations. Others referred to sources that did not exist.
Check what supports the answer
Had those answers passed through a review, they could have affected advice to clients or decisions by project teams. Incorrect advice can lead to rework and delays, especially once other decisions depend on it.
That experience changed how we approach review. Does the source exist? Is it the right document and revision? Does the cited passage actually support the answer? A reference is only useful if someone can follow it and check what it says.
A figure taken from an earlier revision may look reasonable while describing work that has since changed. The document date and the scope it covers belong alongside the number.
Make review part of the process
The work starts before asking the question. Define the task, select the relevant information, and make clear which documents apply. Then check the evidence before using the result. Missing information should remain visible so the team knows what still needs confirming.
For cost estimates and project advice, a transparent, auditable process gives the next person a practical way to question the numbers. They should be able to identify the source, follow the reasoning, and distinguish documented facts from assumptions.
AI helps us move through information faster. Human review remains part of the work, and the responsibility for the advice we give stays ours.