Amazon has been turning Meta's Muse shopping agent away since 20 September, saying it never authorised it. Shopify is letting Muse check out through Shop Pay. Under both sits one question: when an agent buys something, how does the cardholder's bank get the cardholder to say yes? The published EMV 3DS specification already has an answer, and it isn't a new protocol.
It's two existing pieces used together.
The first is 3RI, the device channel for an authentication the 3DS Requestor starts without the cardholder in session. Version 2.3.1.1 lists nineteen reasons a requestor can give, from recurring and instalment payments through split shipment and billing agreements to FIDO credential registration. Merchants already use it for the payments after the first in a series and for checking a card on file.
The second is decoupled authentication. The specification does not let an issuer challenge a 3RI request in session: Transaction Status C is not allowed on that channel. If the ACS wants the cardholder, it answers D. The requestor has already said how many minutes it will wait, the issuer reaches the cardholder by whatever route it chooses, its own app say, and reports the result afterwards.
Put the two together and you have a merchant-initiated transaction that ends with the cardholder approving it out of band. That is the "human in the loop" everyone is now describing for agents, and the fields for it have been in the specification since version 2.2.
It also answers a regulatory question. Under PSD2 strong customer authentication is owed where the payer initiates the transaction, and the RTS wants the payer made aware of the amount and the payee. 3RI exists for payments the payer does not initiate, which is why the later payments in a recurring series may skip SCA. An agent's purchase sits between the two: the customer set it going but is not there. The decoupled challenge is the one moment in the existing protocol where the issuer can put the amount and the payee in front of the customer and record the answer, which is the evidence I argued last week issuers will be asked for.
Two things are missing, and neither is a specification. Support: decoupled authentication has been used sparingly, and a 3RI request with a decoupled challenge needs the directory server and the issuer both to accept it. And the law: whether an agent's purchase counts as initiated by the payer, which decides whether SCA is owed at all, is not settled anywhere I can find.
So, for anyone running 3DS traffic: has a 3RI request with a decoupled challenge gone through your stack yet, and if an agent sent one tomorrow, what would your cardholder see?
First posted on LinkedIn on 13 October 2026, where the discussion is.
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.