The FSB delivered its finalised AI practices to the G20 this week. We set the week's deployments against what it recommended: money-moving agents with no approval step named in coverage, a read-only Claude plugin, and M1's AI adviser with no human behind it. Each item graded for provenance.

A line a client can toggle is a feature. Ours is absolute: deciding stays with the holder. On Schwab, Robinhood, and the question above the question.

This week the question is whose hand is on the gate. Schwab fixes the line itself: its new assistant does not trade. Robinhood leaves the switch with the customer. At Sibos, banks kept people on payment release while a panel forecast the human coming off. Each item graded for provenance.

A bank asks where wealth came from for its own sake. We begin with who holds it. On Deutsche Bank’s source-of-wealth agent, and the person behind the file.

This week the question moved from where the AI agent stops to who answers when it gets it wrong. Deutsche Bank names an accountable executive; JPMorgan keeps compliance gates; Ant International offers a refund; Apollo's Torsten Slok warns of an "agentic bank run," where no one does. Each item graded for provenance.

An AI licensed by every institution still faces only the one that deployed it. On Anthropic's spread across the adviser stack, and whose table it sits at.

The human boundary on finance AI came loose from the agent this week: Accenture sold its guardrail list as the product, Ritholtz bought governance from Hadrius before rolling out Claude, and Oasis Group argued per-action costs will cap autonomy. Beneath it, one engine under six products. Each item graded for provenance.

A platform built from thousands of estates cannot be formed from yours. On Luminary, the estate-data layer, and the meaning no aggregate can hold.