The feed
The things that would otherwise be a LinkedIn post.
One position per post, each one arguable. Open the ones you disagree with, because those are the ones with the reasoning underneath. No account, no algorithm, and every post has a permalink you can send to somebody.
Your AI prioritisation matrix scored that project 8.4. Ask the person who wrote the 8.4 where it came from.
The operators had already built the ranking. It lived in their heads and a spreadsheet three of them maintained. Nobody had ever asked them to write it down.
Per-seat pricing on a feature with a per-use cost breaks quietly, at exactly the moment customers start liking it.
The evaluation set is the cheapest line on any AI build estimate and the first one teams cut.
"We could switch model providers easily." When did you last try?
Two quotes for the same AI build differ by five times. It is almost never the model work.
Search for an AI due diligence checklist and you get a generic venture template with an AI heading pasted on.
You shipped it without taking a baseline. You can still build one afterwards, from data you already have.
A real share of what gets scoped as an AI feature is a rules problem, a search problem, or a form with too many fields.
Next step
If one of these described a problem you recognise more precisely than you would like, that is the conversation.