AI is trained to agree with you.
Harvard Business Review just proved it across 15,000 conversations and every major model. Change the order in which you list your options (not the facts or the context, just the order), and the advice shifts by 19%.
The models aren’t analyzing your situation. They’re confirming whatever you led with.
In some industries, that’s an inconvenience. In the building enclosure, it shows up as a leak five years after the project closed.
An architect prompts: “We’re planning to use silicone sealant at the storefront-to-EIFS interface. Any concerns?” The AI finds the proposed solution within the question and confirms it. It won’t tell you the sealant needs to bond to the base coat, not the finish coat. It won’t tell you the sealant should be ultra-low modulus. And it won’t warn you that the same approach that works at a storefront-to-EIFS joint can fail at an EIFS-to-EIFS joint five feet away. The architect walks away confident. The joint fails in year three.
A real review doesn’t just check what’s there. It also tests what isn’t. A real review also points out the items the designer didn’t know to ask about. That knowledge comes from watching buildings perform over time, not from reading about them.
AI doesn’t have that experience. It has text about that experience.
That’s the difference between a tool and a reviewer. A tool speeds up the work, but it doesn’t decide what needs to be checked or why.
Specialized knowledge and professional judgment aren’t something you can prompt your way to.
It’s the domain expertise filter that makes the answer mean something.