The demo always goes well. The person at the keyboard knows which question to ask, skips the one that breaks, and narrates over the three-second pause so it feels like thinking. The room applauds. Someone says ship it.
Monday is different. The person using it has a queue, a customer on hold, and no interest in your prompt. They ask the question they actually have. The answer comes back confident and slightly wrong. They will not file a bug. They will go back to the spreadsheet.
test it at 4:45
Sit with someone who was not in the demo. Give them one real item from this week. Do not help. Watch where they hesitate, what they copy into another tool, and whether they can tell a good answer from a fluent one.
If they cannot tell, the interface is lying. Correct answers and nonsense, styled the same, train people to trust nothing. Show the source. Say when the system is guessing. Offer the way out.
Users do not experience your model. They experience what the screen implies about the model.
ship the monday version
The Monday version is narrower. Three jobs the tool is allowed to help with, an example from last week, and a person to hand it to when those three are not enough. Breadth can come later, after someone has relied on it without a chaperone.
A tour is easy to applaud. A Monday is the only review that counts.