Before signing with a supplier I have never met, I will spend an hour on what counts as delivered. What the output looks like. When it is late. Who decides if the quality is arguable.
The colleague I have worked with for three years gets a sentence. Someone on my own team gets half a sentence and a nod. The person I live with gets a look and an assumption.
We give our most careful definitions to strangers, whose failure would cost money, and our sloppiest to the people whose failure will cost a month of atmosphere. It holds up, mostly, because shared history quietly fills in the other two hundred words we never said.
Then something turns up with no shared history at all. An AI agent is the colleague who started this morning: trained, articulate, and completely ignorant of what your data means or what you would sign off as done. Give it the same eight words and it hands you the wrong thing in ten seconds.
That speed is the useful part. The failure arrives while you can still remember what you asked for.
I went into this with Dmitrijus for an hour: what you can safely hand an agent, what happens when the model underneath your process gets retired, and why organisations grow scar tissue where they should have had a drawing.
The conversation is in Lithuanian: AI agentai versle: ką jiems galima deleguoti ir kur prasideda pavojus on Verslo klubas Borromean (brmn.lt).