Enterprise organizations are increasingly relying on web-based applications for critical business processes, yet many workflows remain manually intensive, creating operational inefficiencies and compliance risks. Despite significant technology investments, knowledge workers routinely navigate between eight to twelve different web applications during standard workflows, constantly changing context and manually transferring information between systems. Data entry and validation tasks take up approximately 25–30% of a worker’s time, while manual processes create compliance barriers and cross-system data consistency challenges that require constant human verification. Traditional automation approaches have significant limitations. While robotic process automation (RPA) works for structured, rules-based processes, it becomes brittle when applications are updated and requires constant maintenance. API-based integration remains optimal, but many legacy systems lack modern capabilities. Business process management platforms provide orchestration but struggle with complex decision points and direct web interactions. As a result, most enterprises operate with a mixed approach where only 30% of workflow tasks are fully automated, 50% require human oversight, and 20% remain completely manual.
These challenges manifest in normal enterprise workflows. For example, purchase order verification requires intelligent navigation through multiple systems to perform three-way matching between purchase orders (POs), receipts, and invoices while maintaining audit trails. Employee on-boarding demands coordinated access provision across identity management, customer relationship management (CRM), enterprise resource planning (ERP), and role-based decision-making collaboration platforms. Finally, e-commerce order processing must intelligently process orders across multiple retailer websites that lack native API access. Artificial intelligence (AI) agents represent a significant advancement beyond these traditional solutions, providing capabilities that can intelligently navigate complexity, adapt to dynamic environments, and dramatically reduce manual intervention in enterprise workflows.
In this post, we demonstrate how an e-commerce order management platform can automate the order processing workflow across multiple retail websites through AI agents Amazon Nova Act And strands agent Using the Amazon Bedrock AgentCore browser at scale.
Ecommerce Order Automation Workflow
This workflow demonstrates how AI agents can intelligently automate complex, multi-step order processing on various retailer websites that lack native API integration, combining adaptive browser navigation with human oversight for exception management.
The following components work together to enable scalable, AI-powered order processing:
- ECS Fargate tasks run a containerized Python FastAPI backend with a React frontend, providing a WebSocket connection for real-time order automation. Tasks automatically scale up based on demand.
- The application integrates with Amazon Bedrock and Amazon Nova Act for AI-powered order automation. The AgentCore browser tool provides a secure, isolated browser environment for web automation. The Main Agent orchestrates the Nova Act Agent and the Strands + Playwright Agent for intelligent browser control.
E-commerce order automation workflow represents a common enterprise challenge where businesses need to process orders across multiple retailer websites without native API access. This workflow demonstrates the full capabilities of AI-powered browser automation, from initial navigation to complex decision making to human-in-the-loop intervention. We have a sample agentic ecommerce automation that we’ve open sourced aws-samples repository on GitHub,
workflow process
Users of the e-commerce order management system submit customer orders via web interface or batch CSV upload, including product details (URL, size, color), customer information, and shipping address. The system specifies the priority level and queuing order for processing. When an order is initiated, the Amazon Bedrock AgentCore browser creates a separate browser session with Chrome DevTools Protocol (CDP) connectivity. The Amazon Bedrock AgentCore browser provides a secure, cloud-based browser that enables AI agents (in this case the Amazon Nova Act and Strands agents) to interact with websites. It includes security features such as session isolation, built-in observability through live viewing, AWS CloudTrail logging, and session replay capabilities. The system retrieves retailer credentials from AWS Secrets Manager and generates a live view URL using Amazon DCV Streaming for real-time monitoring. The following diagram shows the sequence of the entire workflow process.
Browser automation with form filling and order submission
Form filling represents a critical capability where the agent intelligently detects and populates different field types across different retailer checkout layouts. The AI ​​agent visits the product page, handles authentication if necessary, and analyzes the page to identify size selectors, color options, and cart buttons. It selects the specified options, adds items to cart, and proceeds to checkout, filling in shipping information with intelligent field detection in different retailer layouts. If products are out of stock or unavailable, the agent moves on to human review in terms of alternatives.
The sample application takes two different approaches depending on the automation method. Amazon Nova Act Uses visual understanding of the webpage and DOM structure, allowing the Nova Act agent to receive natural language instructions such as “fill in the shipping address” and automatically identify form fields from screenshots, adapting to a variety of layouts without predefined selectors. On the contrary, Varieties , Playwright Model Context Protocol (MCP) composition uses the Bedrock model to analyze the page’s Document Object Model (DOM) structure, determine appropriate form field selectors, and then perform low-level browser interactions to populate the fields with Playwright MCP customer data. Both approaches automatically adapt to diverse retailer checkout interfaces, eliminating the brittleness of traditional selector-based automation.
human-in-the-loop
When encountering CAPTCHA or complex challenges, the agent stops the automation and notifies operators via WebSocket. Operators access Live View to see the exact browser status, manually resolve the issue, and restart. The AgentCore browser allows the human browser to take over and give control back to the agent. The agent continues from the current state without restarting the entire process.
Observability and scale
Throughout the execution, the system captures detailed execution logs with session recordings, screenshots and timestamps at critical steps, stored in S3. Operators monitor progress through real-time dashboards showing order status, current stage and progress percentage. For high-volume scenarios, batch processing supports parallel execution of multiple orders with configurable workers (1-10), priority-based queuing, and automatic retry logic for transient failures.
conclusion
AI agent-powered browser automation represents a fundamental shift in the way enterprises approach workflow management. By combining intelligent decision making, adaptive navigation, and human-in-the-loop capabilities, organizations can move beyond the 30-50-20 split of traditional automation toward significantly higher automation rates in complex, multi-system workflows. The e-commerce order automation example demonstrates that AI agents do not replace traditional RPA – they enable the automation of workflows previously considered too dynamic or complex for automation, handle diverse user interfaces, make relevant decisions, and maintain full compliance and auditability.
As enterprises face increasing pressure to improve operational efficiency while managing legacy systems and complex integrations, AI agents offer a practical path forward. Instead of investing in costly system overhauls or accepting the inefficiencies of manual processes, organizations can deploy intelligent browser automation that fits their existing technology landscape. This results in reduced operating costs, faster processing times, improved compliance and most importantly, freedom for knowledge workers from repetitive data entry and system navigation tasks – allowing them to focus on high-value activities that impact the business.
About the authors
Kosti Vasilkakis is a Principal PM at AWS in the Agentic AI team, where he led the design and development of several Bedrock AgentCore services, including Runtime, Browser, Code Interpreter, and Identity. He previously worked on Amazon SageMaker from its early days, launching AI/ML capabilities that are now used by thousands of companies worldwide. Early in his career, Costi was a data scientist. Outside of work, he builds personal productivity automation, plays tennis, and enjoys life with his wife and children.
Ved Raman Senior Solutions Architect for Generative AI at Amazon Nova and Agentic AI at AWS. She helps customers design and build agentic AI solutions using the Amazon Nova Model and Bedrock AgentCore. He previously worked with customers building ML solutions using Amazon SageMaker and also worked as a Serverless Solutions Architect at AWS.
Sanghwa na is a Generative AI Specialist Solutions Architect at Amazon Web Services. Based in San Francisco, he works with customers to design and build generic AI solutions using large language models and Foundation Models on AWS. He focuses on helping organizations adopt AI technologies that drive real business value.
