Mike Kane/Bloomberg
It seems like just a couple of years ago, and it was, that banks were figuring out what to do with generative AI. Fresh on its heels is an even more cutting-edge use of artificial intelligence, one that can operate with little or even no human intervention.
Companies that are developing agentic AI technology are hot: the agentic AI market size is projected to total $7.28 billion in 2025 and is on pace to reach $41 billion by 2030, according to
One area of growth is software that supports businesses. Mordor reports that in 2024, fewer than 1% of enterprise software applications harnessed the potential of agentic AI. Most AI-integrated applications still lean heavily on user prompts. However, by 2028, Morder expects this landscape will shift dramatically, with projections suggesting that nearly one-third of all enterprise software will embrace agentic AI.
Is ChatGPT agentic AI?
Agentic AI is distinct from ChatGPT and other forms of generative AI, which use large language models to produce original content with a human to flag and change errors and execute actions, such as a purchase.
In contrast, AI agents have more autonomy, enhancing both the potential and risk.
“It’s super transformational for consumers and for businesses,” Pablo Fourez, chief digital officer of Mastercard, told American Banker. “There are all kinds of use cases.”
It’s early days for agentic AI for any industry, said Alenka Grealish, co-lead of generative AI research at Celent.
A linchpin for agentic AI, the Model Context Protocol (MCP), is only 5 months old, Grealish told American Banker. Developed by Anthopic, MCP standardizes the way AI agents access data sources or third-party services to power agentic AI’s ability to turn that data into action without relying on human direction.
“The banking industry won’t be in the vanguard, but the payments industry already is,” said Grealish, adding that agentic AI could fuel next generation app engagement and e-commerce, and subsequently payments.
Here are examples of some companies that are using agentic AI in an effort to improve payments.
Amazon’s Buy for Me
The e-commerce giant began testing
Amazon will then buy that item for the consumer if they want to complete the purchase. Agentic AI automatically provides the buyer’s name, address and payment details for checkout.
Amazon did not comment. But it appears to lean into the security and concerns that accompany a “no human” payment.
Amazon says on its website that the buyer’s personal account details are encrypted and it cannot view the consumer’s personal information.
Grealish said Amazon’s Buy for Me is a “glimmer of the future” but also stressed the importance of security for all agentic AI deployments.
“Agentic AI in commerce and payments, however, will get nowhere without robust authentication, authorization and auditability,” Grealish said. A variety of entities are working on frameworks for how humans can confidently delegate to AI agents and control permissions; and how mobile apps and e-com sites can authenticate AI agents, he said.
Amazon is also selling the technology that powers checkout-free retail, providing another venue to apply agentic AI.
“There is potential for integrating AI agents into a checkout-free shopping experience, for example, through AI-powered shopping assistants, or for customer recommendations,” Gartner analyst Sandeep Unni told American Banker, noting that retailers are just experimenting in this area, and retailer adoption thus far has lagged.
Lionel Ng/Bloomberg
Mastercard builds Agent Pay
The card network in April released
The platform expands Mastercard’s existing generative AI technology, which enhances customer service, security and onboarding to generate automated responses to customers.
Agent Pay digs deeper into the card network’s data and AI tools to help shoppers curate a mix of purchases for an event, aid merchants on supply-chain management or help a retailer build a marketing or sales program.
The card network is partnering with Microsoft to scale the platform, with other partners including IBM, which is contributing B2B technology, and payment firms like PayPal’s Braintree and Checkout.com for security.
Security and authentication use tokenization, a process that replaces existing account numbers with one-off numbers that make the card useless if stolen. Mastercard is also using Databricks software to train the card network’s AI engine to produce responses to users with less human interaction.
“If you think about your own personal experience, it’s a way to relate to agentic AI,” Fourez said.
For example, a family trip, including tickets, hotel reservations, activities and added stops for a road trip, could be planned through an AI agent.
Details such as degrees of difficulty for a hike, or information about universities for a trip to check out campuses, could be fed to an AI agent, which would produce a detailed itinerary and access to booking, payment and security protections.
“Planning a trip like this can take a lot of time,” Fourez said. “Agentic AI can do this in a few minutes.”
Nathan Laine/Bloomberg
Visa adds agentic AI credentials
Getting consumers to use agentic AI to make payments will require assurances for consumers that provide guardrails to protect against unwanted transactions.
The card network’s agentic AI product also includes personalized shopping recommendations based on a consumer’s spending history.
Visa’s agentic AI partners include Anthropic, IBM, Microsoft, Mistral AI, OpenAI, Perplexity, Stripe and Samsung.
David Paul Morris/Bloomberg
PayPal’s ‘future of commerce’
The payment company in March launched an AI agent framework that integrates with PayPal’s application programming interface.
That means PayPal’s business clients can produce what the company calls an “agentic experience” that enables consumers to pay and access other consumer experiences and services faster, with less navigation and in a more personal and interactive manner.
“Our developers are gathering in our San Jose offices,” PayPal CEO Alex Chriss said during
Bloomberg
Coinbase builds an AI toolkit
The cryptocurrency exchange company in November launched
AgentKit supports multiple large language model developers, such as OpenAI, Anthropic’s Claude, and Llama.
Hollie Adams/Bloomberg
Circle’s agentic guide
The USDC stablecoin issuer recently
Stablecoins and other distributed ledger-supported payments provide a potential for agentic AI to improve access and product development. In addition to Coinbase and Circle, Visa and Mastercard are also pursuing stablecoin payments, and PayPal has issued its own stablecoin.
Writing for
Andrew Harrer/Bloomberg
Block releases a ‘goose’
The payment company in January launched
The first use is software engineering, but the company is developing other adaptations outside of technology development, opening the possibility for payments and other consumer and merchant-facing products.
Goose was released through Block’s open source office, part of a goal to enable the construction of AI agents that can work with multiple parties — improving data, analysis and turning that information into improved transaction processing.
Stripe: A client and a seller
Similar to its
The move is part of a wider
Stripe recently issued Payments Foundation Model, an AI engine trained on tens of billions of transactions that captures hundreds of subtle signals about each payment that specialized models can’t. It also launched Stablecoin Financial Accounts to businesses in 101 countries.
The products come three months after Stripe’s acquisition of stablecoin platform Bridge.