Why AI should propose payments, not send them
The case for keeping the power to move money out of every AI model's reach.
AI assistants are now good enough to draft a payment from a sentence: who to pay, how much, from which account, with which reference. That is genuinely useful. The question for a finance team is what happens next.
In IXFXN the answer is fixed: the AI proposes and a person confirms. The Assistant can prepare the payment, check it against your rules and explain what will happen, but it cannot send it. A person confirms the draft, and the payment then needs the approvals your rules require.
Why not let the AI send small payments?
Because the risk is not the size of one payment; it is the size of a mistake repeated. A model that misreads an invoice, or is fed a cleverly worded instruction, can make the same wrong decision many times very quickly. A person confirming each payment is the cheapest control that stops that.
Where the line is enforced
The line is not a setting in the chat. The part of IXFXN that actually sends money, the payment engine, has its own credentials and cannot use any AI component; an automated check enforces that on every code change. It only acts on a permit that is issued after approval, works once, and expires within 60 seconds. Even a fully compromised assistant has nothing to call.
What the AI is still good for
Drafting, checking, explaining and chasing. In practice that removes most of the typing and most of the waiting, while the decision stays with the people accountable for it.