CFOtech Ireland - Technology news for CFOs & financial decision-makers
Ireland
Cumbuca sees AI agents driving new payment dispute rules

Cumbuca sees AI agents driving new payment dispute rules

Fri, 20th Nov 2026 (Today)
Karen Joy Bacudo
KAREN JOY BACUDO Finance Editor

Cumbuca has outlined how it expects AI agents and Open Finance payments to reshape banking and payments. Co-Founder and Chief Executive Officer Daniel Ruhman said new dispute rules for AI-driven payments could emerge in 2027.

He said banks, card networks and payment providers are already grappling with who verifies an AI agent, who authorises it and who bears the cost when something goes wrong. Those questions, he argued, will become more urgent as software agents begin acting on behalf of consumers within banking and payments systems.

Cumbuca is a Brazilian fintech focused on data consent through Open Finance. Ruhman is also a founding member of Brazil's National Open Finance Association and took part in early discussions around Pix, the country's instant payments system.

Dispute rules

Cumbuca expects dedicated dispute frameworks will be needed once AI agents can initiate or complete transactions in ways banks cannot easily distinguish from direct customer activity. Ruhman identified control, authorisation and liability as the central issues likely to shape the next phase of rule-making.

"We expect 2027 to bring the first dispute rules written specifically for payments made by AI agents. At Money20/20 USA in October, the debate kept returning to who verifies the agent, who authorised it and who pays when something goes wrong. The concern is well founded, because AI agents under evaluation at OpenAI escaped their controls this summer and compromised parts of Hugging Face's infrastructure. Once an agent can use the banking apps on a customer's phone, it can act as the customer and the bank cannot easily tell the two apart. Visa, Mastercard and Ant International have begun work on making their separate Know Your Agent checks interoperable, but identity checks are only the first layer of protection. The new rules will need to flag payments made by software, compare an agent's behaviour with its owner's history and settle cases where the agent went beyond its instructions. Pix starts with an advantage: every payment started outside the bank's app is approved with a credential tied to the customer's phone, usually unlocked with biometrics. But a biometric check only proves the customer was there, not that they understood what they approved, and malware that operates banking apps already hijacks Pix in Brazil. Rules for agents will have to tie what the customer sees to what they sign, flag payments made by software and cap long-lived permissions. Pix's fraud refund mechanism also excludes commercial disputes, so an agent that buys the wrong thing from a legitimate seller has no remedy today. The markets that fix both first will set the rules for everyone else," Ruhman said.

That view places Brazil's Pix system within a broader international debate over authentication, redress and consumer protection. Biometric approval and device-linked credentials can help validate a transaction, but they do not answer whether a user understood or intended a purchase made by software acting in their name.

Closed systems

Ruhman also argued that the rise of AI agents threatens a core retail banking model: keeping customers inside bank-owned apps where they can be sold loans, insurance and other services. In his view, larger banks are more likely to respond by building agents inside systems they control than by opening broad interfaces to third-party software.

He said the economics of agent-based finance will favour businesses that can process high transaction volumes with permissions and consents already in place. Requiring customers to authorise every action individually would remove much of the convenience that makes agents useful in the first place, he added.

"The largest banks will answer AI agents in 2027 by building their own, inside closed systems they control. Retail banking adopted the attention model of large consumer apps, because the banking app is where customers are sold loans and insurance. Agents remove that time on screen in the same way Open Finance does. If every bank offered an open agent interface, an agent could move a customer's money to the best return every second. Banking would start to look like quantitative trading, which is one potential future incumbents are keen to avoid. That explains why so few banks have published interfaces that agents can use, despite all the public enthusiasm. It also explains the significance of Itaú's agreement in late October to buy the AI company NeoSpace for R$2.9 billion. We see that deal as the template for the largest banks, which will buy model and data capability and run their own agents. Fintechs and agent platforms will build on open rails instead and earn their living from transaction volume. To win that volume, a business must handle large numbers of agent transactions at once, with permissions and consents already in place. Asking a customer for consent on every transaction would recreate the friction that agents exist to remove. The economics favour that second group, because compute is the binding constraint on any agent product and data alone will not pay for it. Instinct, an AI agent that books flights, hotels and purchases through WhatsApp, charges users nothing and intends to take a small fee on each transaction. It says it is approaching a billion dollars in transaction volume, which is an early sign of the model that will win. The same technology will do for financial services what the smartphone and Nubank did for access to a bank account. Wealth managers, insurers and financial advisers will use AI to understand what each customer needs and suggest specific products. Human advisers will then step in to close the deal, so customers get solutions tailored to them rather than a standard account," Ruhman said.

The comments point to a split market between incumbent banks building tightly controlled in-house systems and fintechs relying on open payment rails and transaction fees. That could shape how AI-led financial services are distributed, particularly in markets where Open Finance rules already allow data sharing and payment initiation.

Back office first

On payment adoption, Ruhman said account-to-account transactions are more likely to gain traction first in back-office functions such as collections, lending, payroll and treasury than in consumer checkout flows. He cited Brazil's Open Finance payment volumes as evidence that regulation and common technical standards do not automatically change consumer behaviour.

He also pointed to growth in recurring Pix payments and automated Open Finance transfers as a sign that the strongest early use cases are often those that run in the background, with less direct consumer decision-making at the point of payment.

"Account-to-account payments will grow fastest in 2027 in collections, lending, payroll and treasury, and only later at the online checkout. Open Finance payments grew far more slowly in Brazil in 2026 than we, and most of the market, expected. They still add up to only around half of one per cent of the seven billion or so Pix transactions processed each month. That is despite a regulatory mandate, a national technical standard and a central bank committed to their success. Britain and Europe are now promoting pay by bank at the online checkout, and the Brazilian figure is a warning worth heeding. It shows that rails and mandates do not change consumer behaviour on their own. A consumer will only connect an account if somebody gives them a strong reason to do it the first time. After that, paying by bank has to feel as quick and simple as using a card, or customers will abandon the purchase. Where adoption is happening, it's through the kind of payments that customers barely notice. Recurring Pix payments grew from fewer than one million in January to almost nine million in September. Itaú reports a 281 per cent rise in settled recurring volume over six months. In April, automated Open Finance flows such as smart transfers completed 19 million transactions, against 12 million for one-off payment initiation. Smart transfers move money automatically between accounts held by the same person when a pre-agreed event occurs. The trigger can be a better savings rate elsewhere, an overdraft charge about to fall due, or the end of an introductory offer. That said, most of the growth will come from credit. Lenders will use smart transfers to collect instalments, moving money from the borrower's main account into a repayments account in the borrower's name. Since February, every institution with Pix accounts has had to support biometric payments without redirecting to the banking app. That removes the single step that has done the most damage to conversion rates so far. Britain is on a similar path, since the recurring payments reported by Open Banking Limited are still almost all sweeping. Its commercial scheme, launched in June, starts with utilities, government and financial services rather than retail. The same lesson applies to FedNow, whose volume grew by more than 80 per cent between the first and second quarters of this year. The payments that scale first will be the unglamorous ones that save a Finance Director money every month," Ruhman said.