On Mastercard's results call on 29 October, Michael Miebach said the company had signed a broad partnership with OpenAI that includes Agent Pay. He also said that before agentic payments really take off, "consumers need to feel comfortable". On 2 November Mastercard and Tencent's TenPay Global announced a memorandum to explore agentic payments together.
My first question was a simple one. How is such a payment confirmed? An agent may pay, but I would expect the confirmation to still happen in my bank's app.
How it is confirmed today
Mastercard's developer playbook for agentic commerce describes it in a few steps. The customer authenticates with the agent platform and authorises it to buy. The platform requests a token from the network, and "certified agents manage consumer authentication before initiating any purchase". For each purchase the agent has to capture the customer's consent to the merchant, the amount and the details. Mastercard says leading implementations use a payment passkey, unlocked on the customer's own device. The agent may not use a card already saved at the merchant, because then the bank could not tell that an agent was paying. Visa describes the same shape: the agent asks the customer to confirm a payment instruction with a passkey, and the network checks the credential requests that follow against it. Visa's own example of signing up includes verifying identity with the card issuer and creating that passkey.
The same Mastercard page explains why my question is a fair one. A standard 3-D Secure challenge on an agent's payment "typically fails because the agent cannot complete the interactive challenge on behalf of the cardholder". The answer it points to is EMV 3-D Secure 2.3, with enhanced decoupled authentication that lets the cardholder confirm out-of-band, "via a push notification in their banking app". The bank and the checkout both have to support it, and Mastercard describes it as still rolling out.
So today the customer confirms through the agent, on their own device, and the bank sees a token marked as an agent's and can put limits on it. What I want is the confirmation in my bank's app first, and then payments within limits the bank holds. What happens when such a purchase turns out to be wrong is a separate question, and I wrote about it in September.
Start with confirming everything
From my point of view, it's better to start with confirmation of everything. After that we can give the agent a limit in terms of amount and in terms of certain merchants where it is allowed to pay. But still, we have to limit what the customer delegates to the agent.
We already see how this works in BLIK. For a certain merchant you establish a trusted connection, and after that you don't need to generate a code to pay, but you still confirm in the application of your bank. Usually it's not difficult if the bank did a good job and created a convenient experience in the mobile app. It's only a confirmation with one button.
But it can be annoying when you pay to the same merchants and the same accounts every time, and the agent is not allowed to do it. For example, I pay for my child's school dinners every month. The amount is different every month, but it's within a predictable limit, and the school doesn't have a convenient payment mechanism. So every month I spend my time finding out how many days to pay for and what the amount is, and trying not to forget to pay on a certain day. Here an agent could be very, very convenient, and I would certainly give it the possibility to make those payments on my behalf.
The limit belongs in the bank
On the bank side I'd rather have one more limit. When I give the agent an allowance, I would like to tell my bank: from this token you can accept only one payment per month, in an amount not more than this. Then I will be calm for sure that my agent is controlled by the bank.
Part of this already exists. Mastercard's guidance describes agent tokens that can be limited by merchant, amount or time window, and since 5 August 2026 issuers in Agent Pay can define spending rules for each token, which Mastercard then enforces. Putting those rules in front of the customer in the app is the bank's own work, and whether a rule as narrow as mine fits is for each bank to check.
What I want to see in my banking app
In my banking app I would like to see a list of agents, their limits and the conditions to pay. Of course, a pause button should be there. That's already implemented in banks for regular payments. It works better or worse in different banks, but it works.
Good practice is also when a bank adds a notification about upcoming payments. I get a notification about a payment that Apple will take automatically tomorrow from my Revolut account. Apple is doing something automatic, Revolut warns me that the payment is tomorrow, and I have time to stop it or reconsider it. With an agent it's the same.
My experience with 3-D Secure
From my experience in banks, 3-D Secure was accepted very well. In the early days it was difficult, because you had to use a form that was made for desktop and didn't work well in mobile applications. Nowadays push or biometric confirmation works well. SMS is a bit old school, but if the phone can catch the SMS and put the code into the field, then why not?
Whom I would let pay without asking
I trust Claude from Anthropic. I have harnesses and an operating system around it, and I believe I can let Claude pay certain amounts without additional confirmation. But only if on the bank side I am sure that for this particular token of my Claude there are limits which I approved and which cannot be overcome.
Somewhere in the middle
In the buzz around agent payments I see that people are either afraid of it at all or want to trust everything to agents. I think it's somewhere in the middle, and there is no need to allow agents to do everything. Agents will certainly help with difficult payments and with regular payments to the same shop, where the transactions are already trusted. For unique, one-off transactions I don't believe it's necessary right now.
But frankly speaking, sometimes I'd like to ask Claude to fill in the payment details, because transfer details are difficult every time. You have to fill in the title properly, you have to fill in the details properly and check that there is no mistake. That work I would delegate to the bank's own AI agent.
What a bank should do now
From my point of view, a bank should partner with Mastercard and Visa and include agent payments with some limitations as soon as possible. First of all to get statistics, to get feedback from customers and to understand the limitations, so it is prepared for further growth of this service.
Agent tokens in your banking app
We build mobile banking apps and the integrations behind them. If you are planning how customers will see, limit and pause their agents, bring us the flow.
How does a customer confirm a payment made by an AI agent?
The customer authorises the purchase or the payment instruction, and the journey depends on the implementation. Mastercard's guidance has the customer authenticate with the agent platform, and the agent capture consent to the merchant, the amount and the purchase details; it says leading implementations use a payment passkey. Visa describes the customer confirming a payment instruction with a passkey, which the network then checks the agent's credential requests against.
Does 3-D Secure work for payments made by an AI agent?
It depends on the flow. Mastercard's developer guidance warns that a standard interactive 3-D Secure challenge typically fails when the agent cannot complete it for the cardholder. It points to the enhanced decoupled authentication in EMV 3-D Secure 2.3, which lets the cardholder confirm out-of-band, for example with a push notification in the banking app, where the bank and the checkout support it.
Can a bank limit what an AI agent spends?
Mastercard's guidance describes agent tokens that can be limited by merchant, amount or time window, and since 5 August 2026 issuers in Agent Pay can define spending rules for each token, which Mastercard enforces. Whether the customer can see and change those limits in the banking app depends on what the bank builds.
Can an AI agent pay with a card already saved at the merchant?
Not in the flow Mastercard describes. Its guidance says agents cannot use the customer's existing cards on file with merchants, because stored credentials would stop the issuer from recognising the transaction as agentic, so the agent needs a token issued for agent payments.
Should every payment an agent makes be confirmed in the bank's app?
The author's preference is to confirm every payment in the banking app at first, then allow selected regular payments within limits the customer has approved and the bank holds on the agent's token. The payment flow still has to meet its own consent and authentication requirements. One-off payments he would keep under the customer's own control for now.