The fastest way to lose trust in an AI tool is to have it confidently say something wrong. Once happens, you can shrug. Twice happens, you start double-checking everything. Three times, you stop using it.
Founding users were starting to double-check Assist by the second week of launch. The answers were usually right. The problem wasn’t accuracy. The problem was that even when Assist was correct, there was no way to tell if it was correct because of the right reasons. Did it read your saved meeting transcript? Or did it generate a plausible-sounding answer that happened to be true? You couldn’t tell.
We shipped reasoning cards.
Every Assist answer now ships with a visible reasoning card under it. The card names the context Assist routed to: your saved meetings, a specific deal record, a relationship’s signal history, a manifesto section, a content brief, the network graph. It explains in one line why it picked that source. And it links to the underlying record so you can read the source for yourself.
You can collapse the card for a clean reading view. You can expand it when the answer matters. The default is collapsed, because most of the time you don’t need to audit. The point isn’t to make you audit every answer. The point is that you can.
What changes for you. Trust in Assist becomes structural, not anecdotal. The first time Assist says something that surprises you, you can check the source in two clicks. The second time, you stop being surprised, because you’ve learned where Assist is reading from.
