In short
Agentic payments are when an AI agent pays on your behalf, within limits you set beforehand.
The model is being designed primarily for retail.
But paying a bill is not the same as buying a product. The question stops being what to buy and becomes whether, when, and how to pay, and what to do when the payment fails.
An AI agent gets the request for a pair of running shoes below a certain price. It searches, compares, chooses, and pays. That is how almost every piece of content on agentic payments explains the subject.
But paying for a product is not the same as paying an electricity bill, a water bill, a health plan, or school tuition. In recurring services, the bill arrives every month: set amount, single creditor, fixed due date. There is nothing to compare or choose.
On one side, the payer deals with due dates, late fees, interest, installments, and the risk of losing an essential service. On the other, the biller deals with delinquency, collection costs, support costs, and a service that is only released once the payment is confirmed.
For the biller, the question is not how to make an agent buy. It is how to make an agent pay the bill on time.
What are agentic payments?
An agentic payment is a transaction an AI agent authorizes and completes on the user's behalf, within limits defined beforehand: how much it can spend, in which situations, for how long.
What separates this from existing automation is not the absence of a click. It is who decides. An automated rule executes what was programmed. An agent reads the context and chooses within a margin.
Are agentic payments and agentic commerce the same thing?
No. Agentic commerce is the entire purchase journey: discover, compare, choose, and pay. Agentic payment is only the last step.
The full journey of an agentic commerce transaction ends in an agentic payment. But not every agentic commerce journey gets that far: the payment step can still be a link, outside the agent.
In a recurring service journey, the stages are different: knowing the bill has arrived, whether the amount changed, when it is due, which bills are still open, and whether a negotiation is under way.
How does a payment made by an AI agent work?
In three stages:
Intent. The user records the objective and the conditions under which the agent may act. The record is signed and serves as the reference for everything that follows.
Authorization. Before moving money, the agent proves two things: legitimate access to the payment method and permission for that specific transaction.
Settlement. The money runs on the usual rails: card, Pix (Brazil's instant payment system), digital wallets, bank account debit.
Who is building the rails for agentic payments
The landscape is dynamic and constantly evolving. Agentic payments range from an autonomous agent that settles the bill end to end, to models where the final decision stays with the person. The initiatives behind those models come from different directions and shift month to month. Some examples of what is being built:
AP2 (Agent Payments Protocol). Open protocol from Google, released in September 2025, with more than 60 partners. It keeps a signed record of what was authorized and what the agent did. Pix support is on the roadmap.
Mastercard Agent Pay. Has already executed transactions initiated by AI agents in Brazil and Latin America.
Visa Intelligent Commerce. Working on verified agent identity and consent records.
Know Your Agent emerged alongside it: KYC applied to software. If an agent moves money, you need to know who develops it and how it is monitored.
What is the difference between an automatic recurring payment and an agentic payment?
Automatic recurring payment (direct debit, Pix Automático, card-on-file recurrence) | Agentic payment | |
|---|---|---|
Who decides | The user, once, at enrollment | The agent, within the limits it received |
When they decide | Upfront, for all future charges | In each situation, reading the context |
What can vary | Nothing. Fixed rule | Timing, rail, negotiated amount, priority |
Example | Electricity bill debited on the 10th of every month | An agent that defers, splits into installments, or switches methods when the balance falls short |
An automatic recurring payment executes a rule. In the agentic model, the decision is delegated, which does not replace the earlier model: it builds on it. An agent that decides to pay still needs a rail, and Pix Automático is exactly that.
From recommendation to payment: the four stages of the journey
In recurring services, the journey moves through four stages:
Physical. Paper bill, in-person service, payment at the counter.
Digital. App, portal, digital bill. The user does everything, on screen.
Assisted. Conversational agents handle service, negotiation, and payment in the same channel, mostly on WhatsApp. The final decision belongs to the person. Already live in Brazil.
Agentic. Agents resolve bills, deadlines, and outstanding items without human intervention at every step.
Seen in sequence, the agentic stage stops being fiction: it is the continuation of what already runs today.
What changes when the payment is a bill, not a product
The logic is the same in both cases: discover, read the context, decide, pay. What changes are the signals the agent has to read.
In retail, the signal is the product: price, availability, return policy. With a bill, the signals are different. Has the bill arrived? Did the amount change from last month? Pay early or wait for the due date? If the balance falls short, split into installments, switch methods, or defer? If the account is already overdue, which negotiation is open? Each answer changes the final amount and the risk of the service being blocked.
What happens when the payment fails?
Recurring payments fail for many reasons: expired card, exceeded limit, insufficient balance, fraud screening blocking the transaction. It is routine for anyone billing millions of customers every month.
Delegating the decision to an agent does not eliminate the failure. It changes who has to resolve it.
A useful agent knows what to do next: retry at a different time, switch rails, offer installments, send a reminder before the due date. That is not agent intelligence, it is orchestration. Without it, the agent knows how to request payment, not how to secure collection.
It is not enough for an agent to know how to pay. It has to understand bills.
The subject was on the agenda at FEBRABAN TECH 2026, Brazil's largest financial technology event, in Mastercard's workshop "Pagamentos mais inteligentes: a ascensão do comércio baseado em Agentes de IA" (Smarter payments: the rise of AI agent-based commerce). Felipe Goldin, CTO at Bemobi, shared the panel with Guida Sousa, from Mastercard, and André Fialho Tsutsui, from Itaú.
AI agents already search, compare, and recommend. When they start to pay, the conversation moves to another level. And in essential services, understanding bills becomes a competitive differentiator.
When the agent moves beyond recommendation and starts to pay, the bar for what has to work rises sharply. Translating agentic commerce into that context is where Bemobi stands apart, because our sector expertise was built inside the largest recurring services players in the country. With Mastercard, we gain the layer of security, identity, and trust to bring this to our clients faster
Felipe Goldin, CTO at Bemobi
Frequently asked questions about agentic payments
Can an AI agent pay my electricity bill today?
Transactions already executed in Brazil are concentrated in commerce; the application to recurring services is still being built. With Bemobi, bill payment via AI agent will be a reality soon.
Do agentic payments work for overdue bills?
This is an open question. An overdue bill involves charges and conditions that vary by creditor, decisions the current protocols, designed for purchases, do not account for.
Are agentic payments secure?
It depends on verifiable agent identity, signed authorization, and credential protection.
Does Pix work with AI agents?
Pix support is on the AP2 roadmap. There is no Central Bank of Brazil regulation yet.
