Agentic AI in Payments: What's Real in 2026 as Mastercard, Visa Run Pilots Amid Regulatory Wait

Agentic AI in Payments in 2026: What's Real, What's Pilot and What's Still Hype: By Nikita Zelezkins

Agentic AI suddenly feels everywhere. In the past year, it has gone from an abstract research concept to one of the loudest buzzwords in payments. Mastercard announced Agent Pay, followed by Visa's launch of Intelligent Commerce. Stripe, PayPal, and Coinbase launched toolkits for AI-initiated transactions. Amazon and Google rolled out AI shopping experiences that promise to "buy on your behalf." For many, the idea of autonomous agents that shop, decide and pay for you still sounds like something from the future - compelling, but also slightly unsettling. All these announcements make it seem like adoption is happening at scale, and agentic AI is already well embedded in payment processes. But is that really the case, and what lies beneath the marketing layer? In this article, I decided to look at how payment companies are actually using agentic AI today, and what AI agents can realistically do - and not do - in early 2026. What "agentic AI" actually means In a payments context, the term "agentic AI" is often used loosely - and not always accurately. At its core, the distinction between agentic and generative AI comes down to agency and execution. * Generative AI produces outputs in response to prompts: write an email, summarise a document, generate an image, etc. * Agentic AI can independently plan and execute actions to achieve a goal: for example, monitor incoming customer emails, identify urgent issues and draft appropriate responses. Why payments are a special case In payments companies, that distinction matters because acting often means moving money. This introduces legal, regulatory and financial risk. As a result, a truly "agentic" AI system in payments is the one allowed to act within strict permissions, limits and guardrails set by humans. That's why many systems in payments feel agentic, while only a few actually are. To see where the technology really stands, it helps to separate what's live in production from what still requires human oversight and what remains experimental. Where agentic AI is used in payments today Spending on agentic AI is projected to grow rapidly - some estimates suggest it could reach $155 billion by 2030, which is a near exponential adoption curve. But in payments, where that spending goes matters more than how much. Even if companies are spending billions on agentic AI, most of that money is not going into systems that can actually move money autonomously, yet. Below is a practical snapshot of agentic AI adoption in payments today. The most mature form of "agentic" behaviour in payments has existed for years and is embedded directly into transaction flows. These are "agentic-ish" in the sense that they automatically act (approve/decline, step-up, route), but do so within strict guardrails. They are usually used for: Fraud, Risk & Account Takeover Detection According to a recent industry data summary, 87% of global financial institutions now deploy some form of AI or ML-driven fraud detection. Card networks and PSPs continuously score transactions in real time, automatically approving, declining or triggering step-up authentication. Visa has launched genAI-powered fraud capabilities aimed at detecting specific attack patterns such as account takeover. Stripe Radar is explicitly an AI system that evaluates transactions in real time and autonomously intervenes based on risk. Authorisation & Routing Optimisation For large payment service providers like Adyen, Stripe, Worldpay, Checkout.com and others dynamic transaction routing and authorisation optimisation are now standard capabilities, with systems automatically selecting routes and retries to maximise approval rates while controlling cost. These providers don't all market it as "AI", but the function exists across the board. This is a deeply economic use case: small improvements in approval rates translate directly into revenue gains for large merchants. Scam Detection and Consumer Protection Nudges An emerging category is AI-driven scam detection nudges, where users are warned before sending money rather than having transactions blocked outright. Today, this is limited to a small number of providers - most notably PayPal and Venmo - while broader adoption across banks and PSPs remains at an early stage. The second major category of adoption in payments sits firmly in human-in-the-loop systems, and this is where the steepest growth is expected. These tools are often described as agentic, but in practice they function as decision support and acceleration layers, not autonomous actors. They automate analysis, summarisation and drafting, while a human remains responsible for the final decision and execution. Disputes/chargebacks & merchant support GenAI helps merchants and PSPs cope with the rising volume and cost of chargebacks. These tools summarise evidence, draft responses, classify cases and guide operations teams through dispute workflows. Providers such as Stripe and PayPal offer AI-assisted dispute management as part of their platforms, while a growing ecosystem of specialised chargeback vendors like Chargepay, Chargeflow, Justt and others provide AI-supported tools that merchants can integrate, often as a plugin. Compliance and risk operations (KYC/KYB, AML casework, policy Q&A) In compliance and financial crime operations, GenAI is increasingly used as an internal productivity layer for banks, PSPs and regulated fintechs. It supports analysts by reviewing onboarding documentation, summarising alerts, drafting investigation notes and answering internal policy questions. This capability is often delivered as part of platforms from providers such as ComplyAdvantage, Fourthline, Quantexa and others. McKinsey estimates that, in modeled scenarios, moving from human-led to agentic-led KYC workflows can result in 200% to 2,000% productivity gains, even when final decisions remain firmly human-owned. This is where the hype is loudest - and where production rollout is cautious. Why? Because payments involve financial, legal and regulatory risk and are tightly regulated. For example, Payment Services Directive 2 (PSD2) and Strong Customer Authentication (SCA) regimes that apply within the EU and UK require clear human authorisation for payment orders, and there's no current mechanism for AI agents to be treated as equivalent to a human payer. As a result, the main barrier to adoption here is not technological capability, which is largely in place, but legal and regulatory constraints. The regulatory reality: where things stand today Right now, we are at a transition regulatory step. In the EU AI Act is evolving with major obligations taking force in 2026, and related regulatory discussions are happening in the UK. The AI Act rules will set the guardrails for how AI systems are governed - requiring risk classification, human oversight, accountability, transparency, data controls and auditability. It creates the governance framework regulators would need before allowing greater AI autonomy under existing payments law. However, the AI Act and current frameworks do not yet permit fully autonomous payments and leave several important risks unresolved. As highlighted by UK Finance, the trade association representing more than 500 financial services providers, there is still uncertainty over who is ultimately liable when an AI system makes autonomous decisions, how meaningful human oversight can be maintained in multi-agent systems, how responsibility is shared across complex third-party AI supply chains, and whether existing consumer protection and explainability requirements are sufficient when AI acts independently on a customer's behalf. What agentic commerce actually looks like today Today, payment companies are not letting AI spend money freely. Instead, they are using two practical approaches to introduce AI into payments while keeping user control. The first, and the most widely used, is where AI prepares the purchase, but the user approves the payment. This is already live. For example, in Stripe's ChatGPT checkout that rolled out in the US, AI can help find products and fill in checkout details, but the user must confirm the payment before it goes through. The second approach is pre-approved spending mandates. Here, users set rules for spending limits and allowed use cases in advance, and AI can act within those boundaries. This is the model behind Mastercard's Agent Pay and Visa's Intelligent Commerce. Both networks are building the underlying infrastructure, such as identity, tokenisation and control layers but these mandate-based flows are still in early pilots and not yet widely available to consumers. Conclusion: hype or real? Agentic AI is already deeply embedded in payments, but not in the way the headlines suggest. It is widely used for risk, compliance and operational decisioning but is still tightly constrained when it comes to moving money autonomously. Most "agentic" payment experiences today rely on explicit approval rather than free-running AI agents. The shift toward more autonomous payments is happening, but gradually. For now, agentic AI in payments is evolving within clear technical and regulatory boundaries, with meaningful autonomy still some distance away.

Fintechs want to bank on agentic commerce

Fintech startups are gearing up for AI-powered commerce, linking digital payments with agentic platforms like ChatGPT and Claude. Payment firms including Razorpay, Cashfree, Visa, Mastercard and Pine Labs are building pipelines ahead of wide-scale adoption. Currently, end-to-end AI transactions aren't allowed by the RBI, but companies are preparing for the future. Fintech startups are hopping onto the artificial intelligence bandwagon as digital payments emerge as a critical linking point between search operations on AI-powered platforms and actual agentic commerce. Payment aggregators like Razorpay and Cashfree, card networks such as Visa and Mastercard, and merchant payment processors including PayU and Pine Labs are all working towards integrating with OpenAI's ChatGPT, Anthropic's Claude and other such platforms. While end-to-end AI-based commercial transactions are not allowed by the Reserve Bank of India, digital payment companies are building the pipelines to ensure that they are ready once agentic commerce sees wide-scale adoption. "As of now adoption is very low, but people are searching for merchandise on AI-powered apps. So, the next step is they will complete the transaction within those apps only; the payments industry is preparing for that," said the chief executive officer of a large digital payment firm. Payment processor Pine Labs announced that the company is integrating with Open AI's ChatGPT to enable agentic commerce and also will help merchants enable such transactions on AI-powered applications. "We will power merchants who want to be ready for commerce in an AI-powered world. Eventually we will also look to enable AI agents undertaking complete payment services," Pine Labs CEO Amrish Rau said. Pine Labs is looking to run experiments with AI agents undertaking complete transactions in Southeast Asia and the Gulf countries. Once the RBI opens up for agentic commerce, players will enable these payments in India too, Rau said. Currently, UPI's Reserve Pay is the only payment mechanism through which an AI agent can undertake a transaction, Razorpay CEO Harshil Mathur pointed out. UPI Reserve Pay is a mechanism where consumers are allowed to set a transaction limit for any specific merchant on a UPI app. Subsequent transactions with that merchant within the limit can flow without an OTP. "We are already live on Reserve Pay, this is a mode which is allowed by the RBI and where an agent can actually undertake the transaction on behalf of the customer without a second factor of authentication. For the other payment flows, human intervention is still needed," Mathur told ET. Card networks are pushing tokenised payments via AI agents, where the real card data remain masked and tokens can be stored with these AI platforms, thereby ensuring safety of transactions. Gautam Aggarwal, president, India and South Asia at Mastercard, told ET that these payments were awaiting RBI approvals and it could take some more months to go live. AI-powered commerce is still in a very nascent stage in India, industry insiders pointed out. Additionally, for card transactions to be enabled on AI systems, banks will need to come on board. So, the technology needs wider adoption and large industry players to come together. "First merchants need to enable such payments, then consumers need to come in as well, only then will the flywheel be in place. As of now, the plumbing is being put in place," Mathur of Razorpay said.

Mastercard and Visa enlist banks for agentic payment pilots

This content has been selected, created and edited by the Finextra editorial team based upon its relevance and interest to our community. Singapore-based DBS has become the first bank in Asia Pacific to pilot Visa Intelligent Commerce, the payment giant's suite of APIs and partner programme for consent-driven payments by AI agents on behalf of consumers. More than three quarters of Singapore residents are already using generative AI tools such as chatbots in their daily lives, according to a Visa-commissioned study. This momentum is also reflected in online shopping behaviours -- with eight in 10 Singapore consumers now relying on AI assistance for this. DBS and Visa have carried out a series of real-world food and beverage transactions to demonstrate that AI-powered agents can complete everyday tasks on behalf of customers using credit and debit cards via secure, issuer-controlled flows. Now, the partners plan to explore a wider range of agentic commerce transactions, such as online shopping, travel bookings and more. Ananya Sen, group head, regional consumer products, DBS Bank, says: "Our collaboration with Visa shows how agent-led payments can be deployed securely and safely at scale, giving customers confidence in how transactions are made in an AI environment." Separately, Visa rival Mastercard has been working with Westpac to carry out the first transaction in New Zealand to use its Agent Pay framework. Mastercard used a Westpac-issued debit card to purchase cinema tickets. The transaction was fully authorised with cardholder consent, and every participant in the payment flow - issuer, acquirer and merchant - could see and recognise that an agent conducted the transaction. Mastercard and Westpac have already worked together on agentic payment transactions in Australia. Says Westpac NZ MD, product, sustainability and marketing, Sarah Hearn: "Agentic AI has huge potential to improve payment experiences, and we're pleased to be working with Mastercard to bring the technology to our customers in the future." Mastercard has also been showing off the technology in India, using cards issued by Axis Bank and RBL Bank to complete transactions at the AI Impact Summit in New Delhi. "By completing a fully authenticated agentic commerce transaction on our network within a Large Language Model, we have shown what the future looks like: seamless, secure, end-to-end commerce powered by trusted AI," says Gautam Aggarwal, president, India and South Asia, Mastercard.

Mastercard's AI agent-led pay awaits approval: Top executive

Mastercard demonstrated India's first AI-driven agentic commerce transaction in a controlled sandbox, with real payments but simulated merchant environments. A full commercial rollout depends on regulatory clarity, expected in the coming months. The technology aims to automate purchases without human input while using existing secure payment systems like cards or UPI. Mastercard's 'first fully authenticated agentic commerce transaction' in India, which refers to an AI-driven purchase requiring no human input, was a controlled sandbox demonstration rather than a live commercial rollout, a top company executive told ET. A commercial rollout of such a technology hinges on regulatory clarity, which Mastercard's India and South Asia President Gautam Aggarwal said was expected over the next few months. He said the payments during the showcase were live on real systems and real cards, but merchants were connected through sandbox environments as no Indian platform is yet fully embedded into an AI agent ecosystem. "The payment was real, but the commerce side was still controlled," he said. The timeline for commercial rollout will depend on whether regulators treat agent-led payments as extensions of already approved technologies such as tokenisation and passkeys or require fresh approvals. "If no fresh approvals are needed, it could be a matter of a few months. If new clearances are required, it could take longer," Aggarwal said. He stressed that agentic commerce is rail-agnostic and can operate across cards or UPI, making it particularly relevant for India where both payment methods have wide adoption. "This is not about the payment method. It is about automating the entire commerce journey," he said. On security, Aggarwal said agent-led transactions rely on the same safeguards already used in digital payments. "We are not introducing any new technology. The transactions are as secure as what consumers use today. The real change is building trust in the agent," he explained. India's ability to scale digital innovations rapidly makes it well placed to adopt agentic commerce, with e-commerce, payments and banking likely to lead early usage, he noted. He said Mastercard is comfortable becoming the invisible infrastructure behind AI-driven transactions. "We already operate as plumbing in the system, and plumbing is where the scale will be in the future," Aggarwal said, signalling the company's long-term bet on powering commerce in the background rather than owning the consumer interface.