From transactions to intelligence: how generative and agentic AI is reshaping payments

In the past, the act of paying has been mostly transactional — a process of moving value from one entity to another. What we’re now witnessing across the financial ecosystem is a fundamental shift. Payments are becoming intelligent, context‑aware, and adaptive.

As artificial intelligence (AI) continues to evolve, generative and agentic AI is set to transform customer and employee experiences. These technologies are redefining how work gets done by shifting digital experiences from manual, user-driven interactions to seamless and contextual outcome-driven workflows with minimal human intervention.

Unlike traditional AI tools that classify, forecast, segment data, and merely answer questions or provide suggestions, agentic AI acts with intent. It introduces intelligent agents capable of reasoning, acting, collaborating with other agents, and completing multistep tasks on the user’s behalf. It operates seamlessly towards goals defined by policy and context — not by static rules. In practice, that means a payments agent can decide which network to clear through, whether a transaction needs manual review, or when to execute a cross‑border settlement for the most efficient foreign exchange rate.

Some of the key use cases that our customers demonstrate how they are using a combination of data and AI/ML (including generative and agentic AI) to transform customer engagement and deliver new experiences include:

• Agentic commerce: agentic commerce is shifting how consumers and businesses initiate, authorise, and complete transactions. Instead of manual steps across multiple apps and websites, agents can coordinate discovery, decision-making, and secure payments in the background, making payments more seamless, contextual, and intelligent.

For example, Visa launched Visa Intelligent Commerce leveraging Amazon Bedrock AgentCore to build multi-agent workflows such as travel booking, retail shopping, and business-to-business (B2B) payments. It enables businesses to develop end-to-end agentic payment experiences using Visa’s suite of agentic commerce tools available on AWS Marketplace and connect the businesses’ agentic payment applications directly to Visa’s payment network and trigger transactions with straightforward, natural language commands to enable seamless, secure, and contextual payment flows.

Another example is Stripe, who built their agent to consumer (A2C) onsite shopping advisor with Amazon Bedrock and built-in guardrails that allow its customers to build their own onsite branded experience, which lets a shopper ask all kinds of shopping questions to get hyper-personalized responses.

• Fraud and financial crime prevention: in the fight against financial crime, payments organisations are implementing enhanced data capabilities to generate deeper insight into fraudulent activities. The combination of AI and generative AI is being used to accelerate investigations and prevent bad actors from exploiting the system. This technology-driven approach enables real-time monitoring and faster response to potential threats while reducing the burden of manual compliance processes.

For example, Visa built an AI powered fraud prevention solution Visa Protect A2A, specifically for non-card account-to-account transactions to enhance the security of these payment flows. The solution leverages AWS Nitro Enclaves and Amazon EKS, and compute-level isolation that allowed Visa to achieve sub-250 millisecond fraud detection latency and provides real-time transaction scoring through real-time APIs allowing financial institutions to better assess and prevent fraud before it happens. 

• Customer experience and support: payments organizations are taking a comprehensive end-to-end approach to customer experience, spanning from acquisition through to retention and growth. Organizations we work with are leveraging vast amounts of data from multiple sources, including payments transactions, call centre interactions, and digital engagement metrics, to better anticipate and serve their customers’ needs. This data-driven approach enables them to provide highly personalised services across all touchpoints, creating a seamless experience whether customers interact through chatbots, call centers, or digital platforms. 

For example, Remitly implemented a generative AI chatbot solution leveraging Amazon Bedrock's foundation models for intent classification and response generation, combining traditional chatbot functionality with natural language capabilities and custom security controls. Not only were they able to rapidly set up the generative AI solution in just 2 months, but they also achieved high customer satisfaction, with only 3% of the users escalating to a human agent. 

• Payment processing and operations: traditional payments operations functions have historically been plagued with friction points and inefficiencies that have frustrated customers and staff. Payments organizations are now actively implementing generative and agentic AI solutions to streamline these processes. These technologies are transforming manual workflows into automated processes, significantly reducing processing times and improving decision-making capabilities. The focus is on creating seamless, efficient payment operations that enhance both customer satisfaction and operational efficiency.

For example, Stripe revamped their compliance operations by building a large language model (LLM) powered AI research agent on Amazon Bedrock that automates complex merchant risk assessments, resulting in a 26% reduction in review handling time through agentic automation. They built reliability and observability into agent operations, handling thousands of investigations daily while ensuring accuracy of AI-generated insights.

When implementing agentic AI capabilities, some of the key considerations that our payments customers are focused on include scale, security, and system evaluation. Along with the complexities related to fraud and monitoring agents deployed at scale, agentic AI systems introduce persistent memory, tool orchestration, identity and agency challenges, and external system integration, expanding the risks that organizations must address. Agents initiate actions based on goals and environmental triggers that might, or might not, require human prompts or approval. This creates risks of unauthorised actions, runaway processes, and decisions that exceed intended boundaries if agents misinterpret objectives or operate on compromised instructions. 

Working with customers deploying these systems, we’ve observed that traditional AI security frameworks don’t always extend into the agentic space. The seamless and contextual nature of agentic systems requires fundamentally different and enhanced security approaches. 

To address this, we have developed a structured model and framework for understanding and addressing the security challenges of agentic AI systems based on connectivity and self-directing levels, which help organizations to confidently build these capabilities and to deploy agentic AI while managing the landscape of associated risks. Organisations can also leverage integrated security guardrails through Amazon Bedrock and minimize hallucinations and data ambiguity using automated reasoning checks.

Additionally, traditional LLM evaluation methods treat agent systems as black boxes and evaluate only the outcome, failing to provide sufficient insights to determine why AI agents fail or pinpoint the root causes. The robust self-reflection and error handling in AI agents requires a systematic assessment of how agents detect, classify, and recover from failures across the execution lifecycle in reasoning, tool-use, memory handling, and action taking. To address this, we have developed a holistic agentic AI evaluation framework across quality, performance, responsibility, and cost dimensions, in addition to continuous production monitoring and human-in-the-loop validation. Amazon Bedrock AgentCore evaluations provide automated assessment tools to measure how well agents perform specific tasks, handle edge cases, and maintain consistency across different inputs and contexts.

As we look ahead, some payments are most likely to fade into the background — executed intelligently by systems that know when, why, and how to act. Generative and agentic AI will create networks that learn continuously from real‑time behaviour, while human oversight will help ensure every action reflects intent and compliance.

About Vishal Arora

Vishal is a Financial Services industry specialist and leads AWS’ worldwide business and market development efforts for Agentic and Generative AI for Payments, where he is responsible for defining the global strategy, developing sales motions and go-to-market strategies, and helping customers and partners to achieve their business goals leveraging AWS capabilities across the end-to-end value chain. Over the past 7 years, he has held leadership roles in AWS’ Financial Services and Partnerships businesses. Before joining AWS, Vishal spent over 15 years with PwC Consulting in New York, Tokyo, and Hong Kong, where he held various senior leadership roles and was responsible for leading global strategic initiatives for Fortune 100 financial institutions on finance and digital transformation, mergers & acquisitions, and strategic risk and regulatory management.