Under the Council's PSR position, an AI agent's purchase that the customer asked for cannot skip SCA as a merchant-initiated transaction.
The test is who triggers the payment. When the customer says "buy me this", the payer did, so it is not an MIT at all.
I read the rest of Article 3(67) against an agent, and it fails there too: the mandate has to go to the payee, under an agreement between payer and payee, and an agent's mandate typically runs to the agent rather than to the merchant it pays. Article 91(2) then lifts SCA for payee-initiated payments only where they happen "without any interaction or involvement of the payer".
What survives, if you want the MIT route for agent purchases, is a standing mandate to the merchant itself, with SCA if it is set up remotely with the PSP involved (Article 91(3)).
If you were hoping 3DS would settle it, it won't. A 3RI request can carry threeRIInd=11 "Other payment", for an agent's purchase, but the indicator only records what the requestor says the payment is, which is not the legal test.
The Council adopted its first-reading position on 29 October, so it isn't law yet, but the definition is unchanged in substance from the compromise text, and PSD2 has no definition at all.
If you run authentication at an issuer or PSP: are agent purchases already reaching you flagged as merchant-initiated?
First posted on LinkedIn on 3 November 2026, where the discussion is. It answers the question I left open on 13 October, in The human in the loop is already in the spec.
This post is from Out of band, the shorter arguments I make on LinkedIn between essays, kept here with their sources. For the long form, read the essays.