Visa is extending its alliance with Amazon Web Services (AWS) to provide a set of composable open blueprints that organizations can use to enable artificial intelligence (AI) agents to autonomously conduct transactions in real time.
Based on the Amazon Bedrock AgentCore platform for building and deploying AI agents, the blueprints will enable organizations to define multi-step agentic workflows for transactions involving, for example, retail shopping, travel booking and payment reconciliation.
Each blueprint is designed to work with a Model Context Protocol (MCP) server for the Visa Intelligent Commerce platform and associated application programming interfaces (APIs) to enable secure, tokenized payments with additional support for multi-network commerce flows coming soon. At the core of the blueprints is the Amazon Bedrock AgentCore framework that provides access to AgentCore Identity, Runtime, Gateway, Memory, and Observability services.
Organizations that have already committed to using the Visa blueprints include Expedia Group, Intuit, lastminute.com and Eurostars Hotel Company. The overall goal is to enable organizations to automate workflows in a way that enables their internal teams to focus more on improving customer experience, says Scott Mullins, managing director of worldwide financial services at AWS. “There will be more time for higher-level thinking,” he says.
As these blueprints mature, the path for builders to design secure, multi-network commerce systems to coordinate bookings, approvals, and payments will reshape how financial interactions are built and delivered, says Mitch Ashley, vice president and practice lead for software lifecycle engineering at the Futurum Group. “It gives developers blueprints that show how real transactions can be reasoned using agents,” he adds.
Instead of narrow payment automations, these patterns combine Visa’s Intelligent Commerce services with AgentCore’s identity, runtime, memory, and observability, enabling agents to act with context and accountability across industries, notes Ashley.
It’s not clear how quickly AI agents will be conducting transitions on behalf of end users, but now it’s more a question of to what degree than if. Organizations that enable AI agents to conduct transactions still need to work through a range of governance policies and safeguards to ensure AI agents are both secure and acting as expected.
Long-term, organizations will also need to determine how AI agents might negotiate with each other. In theory, the end customer is going to task agents to find the best deal available, which may require it to engage multiple AI agents created by providers of services that have been trained to ensure the highest profit margin possible.
In the meantime, organizations should expect that end customers will increasingly be tasking AI agents to purchase products and services on their behalf using, for example, extensions to the agent-to-agent (A2A) protocol now being advanced under the auspices of the Linux Foundation.
The challenge, of course, will be securing those transactions in an era where online fraud is already rampant. In addition to stealing the credentials of an AI agent, it’s probable cybercriminals syndicates will create their own armies of AI agents that are tasked with trying to make fraudulent purchases.
Regardless of the level of potential risk, the agentic AI genie is not going back in the proverbial bottle. Organizations of all sizes will soon need to enable transactions being conducted by AI agents that customers will be rapidly deploying in the months ahead. If those AI agents are not able to engage with a vendor, they will simply move on to another rival provider of similar products and services.
The one certain thing is that AI agents are about to transform the e-commerce landscape in ways that right now are going to be difficult to predict.
