A Deferred Purchase Is Not a Deferred Authorization

On September 9, 2026, Poland's BLIK payment system reported the first pilot in which an AI agent completed a purchase and paid for it with no user present at the moment of the transaction. The item had been out of stock. The user authorized the purchase in advance — a six-digit code, a banking-app approval, conditions and a spending limit — and the agent waited. When the product came back, the agent bought it for 19.99 złoty. Everything in the account is about what the user set before the wait. The question this series asks is about what was true when the wait ended.

What happens

On September 9, 2026, Polski Standard Płatności — the operator of BLIK, the mobile payment standard available in virtually every Polish banking app — published the result of a pilot with three partners: Juo, which built the agent chat inside the Your KAYA online store; PayU, which processed the payment; and the store itself.

A logged-in user opened a chat with the store’s agent and gave it a task: a specific variant of a hand cream was sold out; buy it when it comes back. The user then authorized the purchase in advance — entering a six-digit BLIK code and approving the consent in their banking app — and, in BLIK’s words, “defining the conditions under which the agent could continue acting on their behalf.” The agent monitored the store’s inventory. When the variant became available again, the agent detected it, added it to the basket, and completed the purchase and the BLIK payment for PLN 19.99. “At the moment the transaction was finalised,” the release says, “no additional action was required from the user — the agent acted autonomously within the boundaries of the payment consent previously granted in the banking app.”

BLIK’s vice president frames the control model plainly: the agent “operates exclusively within boundaries previously defined by the user and cannot complete a transaction without the user’s prior authorisation.” PayU’s head of product adds the three elements that bound it — “previously granted payment consent, defined spending limits and clearly specified conditions.” And BLIK names the purchasing models the pilot is meant to support: a one-off purchase completed immediately with the customer present; a one-off purchase executed later, on the customer’s instruction — when a size is unavailable, an item is out of stock, or delivery is postponed; and subscriptions.

The pilot is a milestone in the terms the release claims: a system spanning the agent, merchant, processor, and payment method completed a purchase with the buyer absent at the moment of the transaction.

The usual reading

Read as commerce, this closes a gap. Agents could already find, compare, and recommend; now one can pay. Read as control, the reading is reassuring: nothing happened outside what the user authorized in advance, through a payment consent approved in the user’s banking app.

Both readings are about the consent. Neither is about the interval.

The interval

The distinctive thing about this pilot is not that an agent paid. It is that the authorization and the execution are separated by a wait whose length nobody chose. The user consented at one moment; the purchase happened at a later moment determined by a warehouse. The consent did exactly what a consent does: it fixed, at the moment it was given, what the agent could do — this product, this variant, this limit, these conditions.

What a consent does not do is hold the world still while it waits. Between the two moments, any of the following can change without touching the consent at all: the price of the variant; the user’s account status or available funds; the merchant’s standing with the processor; the user’s own intent — they bought it elsewhere, or no longer want it; a fraud or risk signal on the account; the consent’s own validity, if the bank or the user withdrew it. The condition the agent was watching for — “available again” — is one fact about the world. It is the fact the user cared about. It is not the only fact on which the eligibility of a payment depends.

BLIK’s deferred model makes this explicit as a product feature: a purchase “executed at a later time on the customer’s instruction.” The instruction is given once. The execution happens whenever the condition is met. The longer that interval, the more the consent is a statement about the past by the time it is used.

Two different objects

The prior payment consent. A scope fixed at the moment of authorization: product, variant, spending limit, conditions, confirmed through the user’s bank. It is valuable because it is stable — the agent can be trusted to wait precisely because the boundaries were set and cannot be widened by the agent.

Execution eligibility at the moment of purchase. Whether, when the condition is met and the agent moves, the payment may open now — given the account, the merchant, the price, the consent’s validity, and whatever else has changed since the authorization. This is a property of the later moment, not of the consent.

The pilot demonstrates the first object well. The public account says the transaction was executed within consent, limits, and conditions — that is, the action matched the scope. What the account does not describe is what, if anything, was re-established at the moment of execution beyond that match: whether the consent carried an expiry; whether the user could withdraw it and how quickly a withdrawal would take effect against an agent already watching; whether a price change on the variant would have been treated as a new condition or absorbed under the limit; whether the account and the merchant were re-checked when the agent moved rather than when the user consented. None of this is a finding that those checks are absent — a press release about a first pilot does not enumerate them, and the banking-app consent and the processor may well carry more than the release says. It is the question the interval makes unavoidable, and it is the question this series keeps finding at the moment an action opens: a match to the granted scope establishes that the action was within what the user permitted; it does not by itself establish that the conditions the payment depends on still held when the agent acted.

Where this is heading

The three models BLIK names — immediate, deferred, subscription — are three lengths of interval. In the first, the interval may be only seconds and the user remains present; consent and execution are closely coupled, but they are still distinct events. In the second, the interval is however long the stock takes. In the third, it is indefinite by design. As agentic payment moves along that line, the prior consent becomes an increasingly distant description of what the user once agreed to, and the work of keeping the payment eligible shifts from the moment of consent to the moment of execution — to whatever re-establishes, when the agent finally moves, that the account, the counterparty, the price, and the consent itself still satisfy the conditions required for execution.

This series has said the same thing about mandates in a payments protocol and about approvals in a payment batch. What the BLIK pilot adds is the simplest possible instance, with a warehouse as the clock: the user did everything a user can do in advance, and the interval still happened.

The proposition

A deferred purchase separates two moments. The first is when the user authorizes: scope, limit, conditions, confirmed through the bank. The second is when the condition is met and the agent acts. The consent from the first moment supplies the prior permission on which the second moment may rely. It does not, on its own, make the second moment eligible — because eligibility depends on the state of the world at the second moment, and the world did not agree to wait.

The order can wait. The authorization has to be current when the order stops waiting.


Source: BLIK — First pilot BLIK transaction completed independently by an AI agent (Polski Standard Płatności, Sept. 9, 2026). The account of the pilot and its control model is taken from this release; it is not the result of independent verification of the implementation.