Across three restaurants in Spain, a voice AI agent called Polina checks availability, creates or changes bookings and answers customer questions. At telehealth company AskDocDoc, AI agents carry out software development within a budget set through subscriptions. These examples show businesses managing software that can take actions, even where it has no authority to buy anything.
The distinction matters as new payment infrastructure makes it possible for agents to purchase digital resources. Cloudflare’s Monetization Gateway, launched in closed beta on 30 September, allows businesses to charge agents for access through payments settled in USDC on Base. The ability to pay adds another action for businesses to authorise and supervise.
Denys Romanov, founder of ReNewator, an AI and software company, says Polina is already interacting directly with restaurant business systems. The agent’s work involves checking information and making changes to bookings, rather than simply producing an answer for someone else to act on.
Romanov describes the operating principle as defined boundaries and logged actions. An agent should have permission to perform specific tasks, with limits on the systems it can access. When money becomes part of the workflow, he would apply the same approach to budgets, suppliers and approval thresholds.
What to Automate?
“Routine, low-risk actions can be automated, while anything above a certain amount or outside predefined rules should require human approval,” says Romanov.
His restaurant example is useful precisely because it does not claim to demonstrate autonomous purchasing. It shows the earlier task of deciding what software may change inside a business. Adding payments would extend that authority and require the business to define the financial actions it is prepared to delegate.
Evgenii Arsentev, chief executive of AskDocDoc and a medical doctor by education, draws an equally clear distinction. His agents do not buy anything. Their role is software development, and he manages their access and cost through a restricted arrangement.
Arsentev reports using three Claude subscriptions at $200 each, giving the business a monthly subscription ceiling of $600. That is a way of budgeting the tools agents use, rather than a wallet from which they independently select and pay suppliers.
“An agent gets access for a task and it ends when the task ends. Logs are kept where the agent can’t edit them. Anything irreversible needs a human decision,” says Arsentev.
Time Limits
This introduces a time limit as well as a limit on the type of work. Access belongs to an assignment and ends with it. Keeping records beyond the agent’s ability to change them preserves an account of its actions, while irreversible decisions remain with a person.
Arsentev’s concern is an agent doing exactly what it was allowed to do in the wrong context. He says he sets permissions with that problem in mind. A system can follow its instructions and still produce an inappropriate result if the circumstances do not match the assumptions behind those instructions.
For businesses considering agent payments, this suggests that a spending cap alone would be insufficient. A purchase could fall within the permitted amount and still be unsuitable for the task. Romanov’s emphasis on approved actions and Arsentev’s emphasis on context both point towards defining what a payment is intended to achieve.
Their approaches also show that financial exposure can be managed before an agent receives a wallet. AskDocDoc’s reported arrangement establishes a subscription budget. ReNewator’s restaurant deployment establishes which business actions an agent may perform. Neither example provides evidence of an agent choosing a paid resource and settling the bill itself, but both offer experience in managing delegated activity.
Agents Paying
Chandler Fang, co-founder of t54 Labs, describes the next stage: agents paying for APIs, compute, premium data and other digital services needed to complete a task. He says more than 20 million agent-initiated transactions have been protected through t54’s infrastructure, including micropayments for paid services. The figure is the company’s own account of activity; Fang says transaction volume alone is insufficient evidence of customer demand. Repeat usage, retention and payments that help complete real workflows are stronger signals.
“The basic rule should be that the agent spending the money shouldn’t also control the rules around that spending,” says Fang.
His approach connects the existing business controls described by Romanov and Arsentev with payment authority. Fang says t54 has worked with session-level spending limits and trust checks before payments settle. Businesses should set budgets, approved services, transaction limits and the circumstances requiring human approval, allowing small routine payments to proceed within those boundaries.
Romanov argues that responsibility remains with the business deploying the agent. When something goes wrong, it needs to establish what the agent was asked to do, what it actually did and why.
“Giving an AI access to a system shouldn’t mean giving up accountability,” says Romanov.
His position makes records useful for more than reviewing a successful task. They allow a business to investigate a mistake and compare the instruction with the action. Arsentev’s decision to keep logs where an agent cannot edit them addresses the integrity of that record.
Fang also argues that agent payments need the protections developed in consumer finance. Responsibility should be traceable across the agent, payment provider and merchant, with a record of what was authorised and why. Checks before settlement may help identify an untrustworthy recipient, but customers also need practical ways to dispute a payment and recover money when a purchase goes wrong.
“You can’t tell someone their AI made a mistake and therefore their money is gone,” says Fang.
Together, these accounts suggest a practical sequence for businesses preparing to grant payment authority. First define the task and the access it requires, then establish which actions can proceed automatically and which require a person. Financial permissions can be added to that structure, with a clear purpose, a record of what happened and a route to resolve mistakes.
Fang sees stablecoins as useful for tiny, frequent payments to digital services across borders, while cards and bank payments remain important where refunds, disputes and merchant acceptance matter. He expects agents to use both. For smaller publishers, the opportunity could be payment for individual articles, datasets or specialist research. He says that depends on payments being easy to accept, interoperable and inexpensive, without requiring creators to become crypto experts.
Polina’s bookings and AskDocDoc’s development work demonstrate controls businesses can establish before granting spending authority. Fang’s account extends those principles into payments: permission tied to a task, rules controlled by people, records preserved for review and a practical response when something goes wrong. As agents gain the ability to pay, those details will determine how much authority businesses can safely delegate.
News source
Cloudflare, Monetization Gateway beta announcement, 30 September 2026. https://blog.cloudflare.com/monetization-gateway-beta/
Additional commentary: Chandler Fang, co-founder of t54 Labs. https://www.t54.ai/

