Agents with Payment Authority: Would We Trust AI with Our Money?

Alexander opens with a household problem: keeping the children supplied with groceries while his wife is away. Repeating an old order is insufficient because the family varies meals, snacks, and other products. He wants to delegate responsibility for a stocked fridge rather than specify every item. This extends the previous episode’s discussion of autonomy into decisions with consequences in the physical world.

The hosts separate purchasing into five steps: discovery, comparison, negotiation, accepting an offer, and payment. Assistants can help with the earlier steps, while independent selection and spending raise harder questions. A shortlist, an approved basket, and an authorized transfer are distinct outcomes. Keeping the final decision with the buyer can still leave substantial useful work for the assistant.

Evgeny describes an agent that processes a family inbox of things to buy or sell, searches Vinted, and proposes bargaining strategies. The family still decides what to purchase. In another experiment, an Amazon assistant ordered unnecessary extension leads alongside a drill, explaining that the combination was a better deal. Evgeny says he could return the extras. Optimizing an apparent bargain did not mean understanding what he needed.

Aleksey points to missing context about everyday life. Much of the context for coding is already digital, whereas buying equipment requires knowledge of room dimensions, connections, and intended use. Even a request for running shoes hides personal preferences. People discover what they omitted while correcting the agent’s suggestions. Secure execution of a payment cannot make an unsuitable product useful to its buyer.

Access to payment is a separate obstacle. Aleksey encounters assistants that complete forms but hand payment details and confirmation back to the user. He would prefer mechanisms designed for agents over browser imitation of a cardholder’s actions. Throughout the discussion, payment security and understanding preferences remain independent requirements. Neither solves the other.

Alexander explains AP2 by distinguishing Checkout from Payment: what may be ordered, and what funds may be transferred. Each authorization can pin down a specific choice or allow discretion within defined constraints. One particular pair of shoes differs from a permitted set of alternatives; paying for a specified order differs from spending an allocated budget. The hosts distinguish protocols and demonstrations from evidence that every buyer can already use every scenario.

Aleksey describes companies wanting lower fees and less dependence on payment intermediaries. Building their own infrastructure nevertheless requires agreements, maintenance, and attention to requirements across countries. He discusses conversion between ordinary money and cryptocurrency as another route. Alexander brings the conversation back to the buyer: if goods never arrive or are defective, someone must provide a practical way to raise a complaint.

The participants differ over the maturity of cryptocurrency refunds. Aleksey mentions services that credit an internal wallet; Alexander asks who guarantees the outcome. The conversation does not establish a universal refund mechanism. For a product team, this leaves a question beyond successful settlement: what obligations remain with the seller, and who helps the customer resolve a dispute?

Evgeny suggests marketing as a promising setting: an agent works within an allocated budget, changes creative material, and evaluates results through available analytics. The objective and spending boundaries are easier to describe than open-ended shopping preferences. Cloud resources offer a related example of standardized products and predefined billing. The hosts also recall unexpected spending growth and consider how automated transactions may appear to fraud detection systems.

For an ordinary shopper, collecting offers and requesting confirmation may be sufficient. Evgeny sees no reason to remove that step automatically when it helps the person assess the choice. Paid access to website content appears as a possible transaction between agents. It remains a proposed use case: having two agents negotiate does not by itself establish who benefits or what the service should charge for.

Information search and ordering have different freshness requirements. Knowing an item was available last week is not enough to buy it now. A price or stock level can change between selecting alternatives and confirming the order. Airline tickets illustrate why a promising recommendation may fail before completion: the relevant information changes during the workflow. Websites’ protection against automated access creates another obstacle.

When discussing shopping inside a chat, the hosts consider assortment, inventory updates, fees, and the merchant’s loss of its own interface. Their explanation of difficulties in this model remains their assessment. Sellers need more than payment execution: they value customer relationships, analytics, and opportunities to offer other products. An integration must account for those incentives; technical ordering capability alone does not secure merchant participation.

If someone selects everything through one assistant, that assistant becomes an intermediary between the person and shops. Evgeny notes that a seller integrating through tools may see individual calls but little of the customer’s selection process. Alexander compares this with banks that risk becoming invisible settlement infrastructure and consequently seek to build shopping services of their own.

A personal assistant also depends on access to user context. The hosts discuss knowledge accumulated in chats and search, alongside devices that could recognize a need without an explicit request. Purchases could then become proactive. Richer personalization, however, requires greater trust in the device provider. Anecdotes about advertising after conversations do not establish covert recording; the participants offer different explanations for what they observe.

Aleksey compares responsibility for an assistant to parental responsibility: in his view, the person launching and paying for it should answer for its actions. This is a participant’s position rather than a legal rule covering every transaction. The hosts return to levels of delegation: gather information, advise, prepare a purchase for approval, act on a trigger, and manage a budget. Calling all of these autonomy obscures which decisions the person has actually handed over.

The closing discussion presents two views of fraud prevention. An assistant might ask questions and introduce a pause before a transfer, but misleading product descriptions or injected instructions could manipulate the assistant itself. Lower fraud therefore remains a hypothesis. The hosts expect new scenarios and interfaces without identifying an inevitable winner. They connect broad adoption to user convenience, workable economics, and consequences that people can understand and control.